Anda di halaman 1dari 130


FOREWORD People in every walk of life are required to make decisions, irrespective of their level or function and an officer in the defence services, is no exception. In planning of an operation which includes selection of objectives, allocation of resources to tasks and scheduling of activities against a time frame, he is to select the most suitable option amongst the various alternatives. Further, during the implementation fresh decisions in respect of readjustments, reallocations and use of reserves are to be made. An important and essential ingredient to effective decision making is accurate and timely information. Statistical techniques offer powerful tools for sifting raw data and extracting precise and relevant information. When quantified, the information is precise and less vulnerable to distortions and misinterpretations. Armed with options generated through quantitative techniques, a decision-maker taking into account such qualitative information that has a bearing on the decision would make an effective and right decision. The increasing application of quantitative methods today makes it imperative for anyone in the military or otherwise to have a basic understanding of the available modern decision-making tools. A comprehension of the fundamentals

enables him to identify, situations in which quantitative methods might be usefully applied, as well as to appreciate the type and quality of informational inputs required for various decision-making models. With increasing availability of analytic software and computer awareness, organisations have ample opportunity to apply quantitative models in their decision making. Operations Research, as a problem solving technique must be viewed as both, a science and an art. The science aspect lies in providing mathematical techniques and algorithms for solving appropriate decision problems. It is an art because success of the solution of a mathematical model in all its phases depends on creativity and personal ability of the decision-making analyst. The handbook, Quantitative Aids to Decision Making, attempts to give the reader an exposure to various OR techniques and their application through few Defence related examples. Deliberate effort has been made not to delve into complex mathematical derivations, and to focus on the concepts and formulation of the problems. The handbook, in 1982, was initially prepared by Col MR Narayanan of the faculty. Gp Capt SK Manocha of R&C Faculty has developed the additional chapter on network flows. Credit is also due to him for the illustrative presentation and graphics included in this edition.

College of Defence Management Bolaram Post Secunderabad - 500 010 6th October 1997

(NP Sambhi) Brigadier Officiating Commandant

PREFACE The handbook on Quantitative Aids to Decision Making is designed to provide the reader with an introduction to some of the important quantitative methods that are available to aid the decision making. The illustrative problems used to explain the techniques have been deliberately kept simple with sole purpose of putting them in a straightforward and understandable manner. Many of the problems and examples in the handbook are adopted for military application, but are equally relevant to management of resources anywhere. Some of the examples, to the doubting Thomas, regarding the potential of the quantitative techniques may appear to be too simplistic and not real-life solutions. But these should be taken as the springboard to tackle much larger and more complex problems encountered in real life. However, for solving real life problems an integral and systems approach is required. First section of the book has six chapters on statistics. Starting with basic concepts of the discipline, this section through examples, explains different facets of data presentation and analysis field of descriptive statistics. In the area of inferential statistics fundamentals of probability theory and some important distributions have been dealt. The section finally gives a brief exposure on sampling and hypothesis testing. Diversity of the situations where Operations Research can be applied is unfathomable. Specific approach adopted and technique selected to solve any complex problem depends on the creativity and ingenuity of the analyst. Therefore, it is important that the nuances and limitations of the mathematical model are well understood by both the users and OR practitioners. The viewpoint that an OR user need not learn about the mathematics of OR because computer can take care of solving the problems, is dangerous. Computer solves the model as presented by the user. Therefore, the user must be

aware of the process of formulation and conceptualisation of the problem. Second section of the handbook attempts to provide this basic knowledge. The section on Operations Research opens with a brief overview of the subject. Some of the commonly applicable techniques have been covered without going Into any complex mathematical treatment. That is why Simplex method of linear Programming has not been covered, but the problem formulation has been taken up. A new chapter on Network Flows has been included in this edition to present the potential of some simple algorithms which can be applied to reallife problems and affect appreciable saving in resources. It is hoped that this handbook will be of value to those getting first exposure to OR as well as to others as a ready reference for the basics on the subject.

CONTENTS SECTION - I Chapter 1. Statistics for Management - Basic Concepts 2. Data Presentation and Analysis: Frequency Distribution 3. Data Presentation and Analysis: Measures of Central Tendency and Dispersion 4. Probability Concepts 5. Probability Distributions 6. Inferences Based on Sample Information SECTION - 2 7. Operations Research: Introduction 8. Linear Programming 9. Transportation Model 10. Decision Theory 11. Queueing (Waiting Lines) 12. Monte Carlo Simulation 13. Network Flows Bibliography

SECTION 1 CHAPTER 1 STATISTICS FOR MANAGEMENT BASIC CONCEPTS Introduction Decision making, at commanders level, is becoming more and more complex, involving comprehension of a large number of interrelated/interdependent factors impinging on the decision process. Simultaneously, the commanders are also faced with the problem of ever increasing number of alternatives to choose from. Under the circumstances, decision making, purely based on past experience, judgement and intuition, has become rather difficult. Moreover, the researchers are of the view that the human mind, generally, is not capable of perceiving in all details more than seven parameters, on an average, at a time. Statistical techniques offer a useful and powerful tool in the decision making process, by their ability to analyse and interpret quantitative information in a scientific and objective manner as also in providing certain conceptual framework to the decision maker, to enable him to comprehend qualitative information in a more objective way. Scope of Statistics Statistics is concerned with scientific methods of collecting, organising, summarising, presenting and analysing data, and what

is even more important, is drawing valid conclusions and making reasonable decisions based on such analysis. Importance of Statistics Whatever be the field of human endeavour, complete information can seldom be obtained due to high costs and time factor. In real life, partial information forms the basis of most of our decisions. Statistical techniques enable us to:

Identify what information or data is worth seeking. Decide when and how judgements may be made on the basis of partial information. Measure the extent of doubt and uncertainty.

The key distinction between normative techniques and statistical techniques lies in the fact that statistical methods are explicit in nature and provide a clearly defined measure of error. For example, when we say, With 90% confidence it can be said that the mission will succeed, thereby implying that nine out of ten times the mission will succeed, we are quantifying the error that one out of ten times the statement may turn out to be false. CLASSIFICATION OF STATISTICAL METHODS Statistical methods may be classified as Descriptive or Inductive, the latter pertaining to statistical inference. Descriptive Statistics When statistical methods are used, a problem is always formulated in terms of population or universe- defined as all the elements about which conclusions or decisions are made. For example, suppose we want to know the satisfaction level of our naval officers in the present appraisal system, all the naval officers of our navy would represent the population about whom conclusion

is to be drawn. Similarly, if general opinion of the officers on the usefulness of a particular course is required, then all the officers on that course would constitute the universe or the population. Census. If information or data is taken from. each and every element of the universe to reach a conclusion about the population we are dealing with Descriptive Statistics. In survey language, this process of eliciting data or information from every element of the population or universe is called census. Statistical Inference Another way conclusions or decisions are made, is through the use of a portion or sample of elements from the universe. The sample data is analysed, and based on the sample evidence, conclusions are generalised about the target population. For example, to assess the satisfaction level of the naval officers, we may take a representative sample of officers at random and based on the limited observations (opinions) taken from this sample, we would draw conclusions about the population of all the naval officers of the Indian Navy. This method is referred to as Statistical Inference ie estimation of the characteristics of a large group (or population), based on limited observations of a small portion of it. Sampling. Since only a portion of the population is studied, this method is called sampling as against census where each and every element of the population is studied. Merits and Demerits of Sampling. Some of the important merits and demerits of sampling are given below:

Saving in Cost and Time. Cost of a census is often prohibitive, both in terms of rupee cost and time For example, if the Personnel Branch has to collect data from each and every officer of the service concerned, regarding their confidence level in the existing appraisal system, it is obvious that the

process will be tedious, time-consuming and costly. By using sampling techniques, however, the same conclusion may be obtained by interviewing a relatively small representative sample of officers at much lower cost and time.

Limited Information. During war, time may not be available to obtain full information, and decisions are to be taken based on partial or sample information. Destructive Testing. In a number of problems involving quality control, destructive testing of a representative sample is resorted to for ascertaining the quality of a batch. For example, while ascertaining the quality of a batch of ammunition manufactured in an ordnance factory, destructive testing of only a random sample of cartridges is carried out. Obviously one cannot carry out destructive testing of each and every cartridge. Accuracy. When we draw conclusions about the population based on sample evidence, there will be an element of uncertainty or error in our conclusion. This is called sampling error. Statistical methods have been generated for deciding:

What would be the sample Size consistent with the population pattern /dispersion and predetermined acceptable error in the estimate of the parameter related to the problem. What is the degree of uncertainty due to sampling.

Many of the methods of Statistical Inference are based upon results that have been taken from the field of probability theory, for coping with the sampling error. Statistical Investigation Statistical Investigations may be broken down into following stages:
q q

Formulation or definition of the problem. Collection and validation of data.

Classification, tabulation and description of results. Analysis and interpretation.

Formulation or Definition of the Problem. This is the most important stage in the statistical investigation of a problem as perceived, with a view to arrive at the real problem (based on cause-effect chain) and then to decide on the information and/or data required. Collection of Data. The next step in statistical investigation is the collection and validation of data. It is a time consuming and costly component of the statistical process. Validation of data is a pre-requisite for problem definition and solution. Classification, Tabulation and Description of Results. Collected data to be of any use must be classified in a systematic manner and presented in the form of tables or as graphs and charts to enable the decision maker to comprehend the information quickly. Analysis and interpretation. If the universe or population has been studied and data collected, decisions can be arrived at using techniques of descriptive statistics. If generalisations or inferences are to be made for the population based on the study of a sample or samples, theories and techniques of inductive statistics apply.

CHAPTER 2 DATA PRESENTATION AND ANALYSIS: FREQUENCY DISTRIBUTION Data and Information Data or quantitative information is normally obtained as under:

Data. Continuous Data Obtained as a result of measurement eg, distance, timings, weights of men and material, dimension of obstacles - both artificial and natural. The data obtained as thus is referred to as continuous as it can assume all possible values, subject to the limitations of the accuracy of the measurement ordered or the measuring device used. Discrete Data. Scoring- eg range classification results, appraisal data of officers based on an accepted scoring norm.

Counting - eg number of accidents, missile kills, failures of weapons sub-systems, number of mines in a strip, number of prisoners of war.

Ranking - eg listing in order of preference the various courses of actions based on qualitative criteria.

Counted, or ranked data is called discrete as it can assume certain whole number values only and not the fractional values. Frequency Distribution If we want to obtain any meaningful conclusions from data, discrete or continuous, we must first organise the data properly

and then be in a position to describe the data, its pattern or distribution, so as to render the large number of observaions readily understandable. Since this is the first step with any set of data, we will consider in detail this aspect of organising data with an example. Monthly Flying Hours. Let us say that it is required to find the monthly flying hours achieved by a squadron. For this purpose, the monthly flying hours achieved by five squadrons during 20 months from November 1995 to June 1997 was obtained. Table 2-1 gives the figures as recorded in the sequence in which they were obtained.

Table 2-1: Number of Monthly Flying Hours Achieved by Five Squadrons During 20 Months

A scrutiny of the table at first reveals very little. We may perhaps be able to locate the maximum and minimum values with a little difficulty. To get more information from these observations, it must be organised in some systematic fashion. The simplest device is to form an array which is an orderly arrangement of values of items regarding the magnitude. These values can either be arranged in an ascending or descending order. The observations shown in Table 2-1 has been re-arranged in ascending order in Table 2-2

Table 2-2: Array of Flying Hours Achieved - Ascending Order

Such an array gives a slightly better picture of the data than the original set of observations. For example, a quick look at the array shows that the minimum flying hours achieved is 124 and the maximum 268. Thus we draw the useful conclusion that the range of the set of observations is 144 hours. When crucial decisions have to be made on statistical studies, the number of observations is usually large and arranging them in an array is time consuming and unwieldy. To grasp the significance or understand the underlying pattern inherent in such massive data, the usual technique is to condense such data by suitably dividing the range of the observations. Such a tabulation of observed data into classes

is referred to as a Frequency Distribution Each group of observations is, called a class, which is bound by an upper and a lower limit. By judicious selection of the number of classes, the inherent pattern of the data is revealed. The pattern is generally of great interest to the decision maker. An example of a tabulated frequency distribution relating to the number of monthly flying hours achieved in 100 squadron months is given in Table, 2-3.

Table 2-3: Frequency Distribution of Monthly Flying Hours Achieved by Squadrons

Graphical Representation of the frequency Table Frequency distributions may be presented graphically so as to give an effective presentation of the pattern of the data. Histogram and frequency polygon are the two common methods used for this purpose. For the data above, histogram and frequency polygon are shown in the Figures 2-1 and 2-2, respectively. Histogram, which is like a bar chart shows the frequency of each class clearly, whereas the frequency polygon presents a more effective picture of the variation in the distribution of data. The latter can also be used for a more meaningful comparison of related sets of data, eg: range

Figure 2-1: Histogram

Figure 2-2: Frequency Polygon classification performance of two units, rocket firing results of various air force squadrons etc. If instead of the absolute value of the frequency, the relative frequencies are used, the shape of the histogram will be identical except that the scale on the Y axis will be reduced by n times, ie to 1/n, where n is the total number of observations.

Frequency Curve. We may also draw a smooth curve. instead of the frequency polygon. This is known as a frequency curve (Figure 2-3) and is extensively used to analyse and interpret data mathematically. When the curve is drawn for relative frequency instead of absolute values, the area under the curve is treated as unity and we may obtain the relative frequency between any two values on the X axis by the area bounded by the two ordinates corresponding to these points.

Figure 2-3: Frequency Curve Skewness and Kurtosis The frequency curve is characterised by skewness and kurtosis. These are explained below: Skewness. Frequency distribution that is not symmetrical ie extreme values at one end and not balanced at the other end. The distribution may be right skewed or left skewed.

Right Skewed(+ve).

Extreme values in the direction of higher measurement tail to right, as shown in Figure 2-4.

Left Skewed (-ve)

Extreme values in the direction of lower measurement Tail to left as also shown in Figure 2-4.

Figure 2-4: Skewness - Left and Right Kurtosis. Refers to degree of peakedness. There are three types as under and are depicted in Figure 2-5:
q q q

Platy-Kurtic Meso-Kurtic Lepto-Kurtic

Figure 2-5: Kurtosis Ogives (Cumulative Frequency Diagrams) Very frequently, it is desired to show in diagrammatic form, not the frequency or relative frequency of occurrence in the various intervals, but the cumulative frequency above or below a given value. For example, we may wish to be able to read off from the chart the number of squadron months where number of flying hours

achieved was not less than a given number of flying hours. Charts of this type are called Cumulative Frequency Diagrams or Ogives, pronounced as Ohjives. These are of two types, viz Less Than curves and Or More curves. The table and chart for the two curves are shown at Table 24 and Figure 2-6.

Table 2-4: Cumulative Relative Frequency

Figure 2-6: Cumulative Frequency Diagram(Ogives)

CHAPTER 3 DATA PRESENTATION AND ANALYSIS: MEASURES OF CENTRAL TENDENCY AND DISPERSION In Chapter 2, we have seen how a mass of unconnected data can be transformed into a more readily comprehensible form by means of frequency distribution. But even this may not be adequate. We should be able to present a gist of the data in a nutshell. A set of observations can be described satisfactorily by

Measures of central tendency. Measures of dispersion.

Measures of Central Tendency Generally, all data tend to cluster around a point in the middle region of its range. The measures by which this central tendency is described are:

Mean. Mode. Median.

Mean (m). Is also referred to as Arithmetic Mean or Average. Given a number of values Mean is easily obtained by adding the individual values and dividing the total by the number of values.

Similarly, mean for the squadron monthly flying hours works out to 205.

Merits and Demerits of Mean. It is the most familiar concept and is easily understood. One major advantage of mean is that it is the most reliable measure when we deal with samples for estimating the average value of the population. However, one has to be careful in using and interpreting average (mean) as representative value for drawing conclusions. The pitfalls of using mean value alone for decision making purposes can be grasped from the following examples:

For a beach landing operation, a site was selected carefully where average depth of water was 110 cm. The troops selected for this operation had an average height of 150 cm. When the troops actually landed at this site, more than 40% had to be rescued from drowning. There are number of examples where the staff have worked out the logistics provisioning based on average requirements, resulting in stockouts on 50% of the occasions.

Reasons for the above can easily be understood. Average values very often can completely distort reality if extreme low or high values occur in the data. While arriving at average depth of 110 cm

with 15 or 20 readings in the proposed landing site, even three or four values less than 20 cm would have pulled the average to 110 cm although a large number of readings could have been 150 cm or more. Median. Median is the value of the middle item of the series of observations when arranged in any array. Examples:

Number of values is odd Median value = 6

3, 4, 4, 5, 6, 8, 8, 8, 10

Number of values is even Median value =

5, 5, 7, 9. 11, 12, 15, 16 = 10

9+ 11 2

Merits and Demerits of Median.

q q q q

It is not affected by extreme values. It is easy to determine because it is the mid value. It may be graphically arrived at. Even for open ended distributions, median can be determined. It is a positional measure.

Mode. Mode Mode is the value that occurs most often and around which the values of other items tend to cluster. When data is grouped and represented pictorially, mode will stand-out having the maximum height.

Merits and Demerits of Mode.


It is also not affected by extreme values and is easily determined. It is the least precise of the three measures of central tendency It is difficult to calculate unless data is grouped. It is the least understood and has limited statistical use.

q q

Measures of Dispersion Measures of central tendency like mean, are only representative single value of the data or observations. They do not indicate how the data is dispersed. When we say that average height of a jawan in a Gurkha unit is 140 cm, it is only a typical figure. But it does not tell us how the heights of all the jawans are distributed relative to this figure. It could be between 120 cm and 170 cm or perhaps 135 cm and 155 cm. Let us consider another example. Consider two units, each with say 100 firers. Mean score of both may be 20, out of a maximum possible 30. While in one unit there may be a wide difference in scores ranging from 10 to 30, in the other, the dispersion may be less - say from 15 to 25. Describing the performance of the unit on mean value alone will not be realistic, as it conceals the difference in the individuals standards. If the desired minimum score is 18, there should be more number of failures in the unit where dispersion in scores is from 10 to 30. We, therefore, require a measure of dispersion. Smaller its value would imply greater uniformity in the data. Two measures of dispersion will be considered here:

Range Standard Deviation.

Range. It is the difference between the highest and lowest values in a set of data.

Merits and Demerits of Range.

q q

It is well understood and easily computable. Its main disadvantage is the emphasis on the extreme values only, not being, based on each and every item. It cannot tell about the character of the distribution. It cannot be computed for open ended distributions.

q q

Uses. Being simple, it is extensively used in statistical quality control in industry for Process control. Another application area is for computing share values in the stock market.

Standard Deviation. This is the most important and mathematically precise measure of variability or dispersion. It is also called Root Mean Square Deviation because it is the square root of the squared deviations from the mean. Let us try to understand this definition with an example. Consider the data shown at Table 3-1.

Table 3-1: Calculation of Standard Deviation

In the above example if the values were, say 30, 35, 40, 45 and 50, mean (m) would be 40. But We would intuitively realise that this data has less dispersion or variation with respect to mean. Standard deviation (s) in this case works out to 7.07 or 7 approximately.

Advantages of Standard Deviation. The standard deviation is a comprehensive measure of the variability, since it takes into account every item in the series. Besides, it has mathematical properties and is used extensively in advanced statistical theory.

Disadvantages of Standard Deviation. The major disadvantage is that it is an abstract measure and a little difficult to understand. However, one needs to realise that it represents a kind of average of the individual amounts of variability and greater the average, greater is the variability in the series. As in the case of the mean, the standard deviation cannot be calculated for an openended distribution because of the lack of a mid-point in the openended interval.

The exact significance of the standard deviation will become clear after a discussion of the Normal Distribution considered in Chapter 5. Convention of Symbols The statistical symbols used for different parameters in respect of population and samples are at Table 3-2.

Table 3-2: Conventional Statistical Symbols

Data Analysis - Flying Hours Example Let us take the flying hours data given earlier at Table 2-1 as an example and based on various descriptive measures of that data set, examine various conclusions that could be made. Treating all 100 observations as a data set, we can attempt to describe the overall data behaviour through the following numerical measures:
q q q q q

Mean Median Mode Range Standard Deviation

= = = = =

204.00 hrs. 205.13 hrs. 206.56 hrs. 144.00 hrs. 31.00 hrs

The mean, median and mode being very close to each other, the distribution could be treated as symmetrical (infact, however it is slightly left-skewed). As will be seen later in explanation of the normal distribution, one can construct intervals about the mean in terms of standard deviations and conclude about the proportion of flying hours lying within such intervals. For example, within an interval 204 31 hrs ie, between 173 hrs and 235 hrs, approximately 68% of the observations would lie. And an interval of 204 62 hrs would contain about 95.4% of the observations. If one has to evaluate the overall performance of all the squadrons, one would like to question whether the mean monthly flying hours achieved under the existing conditions is acceptable? If so, the next question would be whether the dispersion as dictated by a standard deviation of 31 hrs is acceptable? The first question pertains to the mean performance standards while the second to the variability or consistency. If the performance is deemed to be unacceptable, measures could be thought of to enhance the mean and decrease the standard deviation. After implementation of such

measures, again similar data could be obtained and analysed. The new mean and standard deviation could then be compared with the original values to see if the differences are significant or not ie, whether the differences are due to random variability or due to the effect of the measures taken to improve the performance. The above analysis, however, is carried out using techniques of inferential statistics. If one has to compare the performance of the five squadrons with one another, the approach will be different. The mean flying hours and the standard deviations for each squadron (ie for 20 data items each) have to be worked out. The results are shown in Table 3-3, here:

Table 3-3: Flying Effort Comparison of Five Squadron One should not judge the squadrons performance by looking at the mean alone. The standard deviation also has to be examined simultaneously. In this example, Squadron 3 appears to have the best performance with the highest mean and the lowest standard deviation. This combination may not always happen. Therefore, both these measures are combined into one measure known as Coefficient of Variation.

Coefficient of Variation =

Standard Deviation Mean

X 100

and wherever a higher mean value dictates better performance (this is not so if the variable under consideration is time for performing a specific task), the lower the value of this coefficient, the better is the performance. The conclusion, based on the coefficient of variation, that Squadron 3 has performed the best may not really be valid because of random variability present in the measurement and/or observation. Secondly, a total of 20 observations in case of each squadron may not be adequate to provide us with reliable inference. Strictly speaking, the variations amongst the individual means and standard deviations may not be statistically significant. Or they may be. To test for this, inferential statistics provides suitable techniques and the statistician carries out suitable tests before concluding about the superiority of the squadron in question.


Introduction Uncertainty is an important element in most decision making situations. There is uncertainty when we deal with happenings in the future, for it is difficult to forecast the future with complete certainty. Again, there is uncertainty when information is not complete, and decisions have to be taken with limited information. The effect of uncertainty is all the more pronounced at higher levels because decision making at higher levels involves looking deeper into the future, usually with less information. Our need to cope with uncertainty leads us to the study and use of probability theory - In fact, usages of the concepts of the probability is part of our every day life. In personal as well as managerial decisions, we face uncertainty and use probability concepts, whether we admit the use of something so sophisticated or not. Perhaps we do so intuitively, largely sub-consciously. When we hear of a weather forecast indicating high likelihood of rain, when looking at the sky, our subjective assessment indicates such a likelihood, we drop the idea of a party out in the lawns and plan to organise the same indoors. While playing bridge, we do make some probability estimates say, before. attempting a finesse. Again, before putting in an attack, we do make a deliberate attempt to consider the chance of success of own forces. In fact, in tactical appreciation, consideration of factors such as ground and relative strength and that of courses of action, is largely devoted to estimating the likelihood of own success and evolving a plan that attempts tomaximise the same.

In so doing, what we are really delving in is the domain of probability. Probability may be defined as likelihood of occurrence of an event. It is represented on a zero-to-one interval scale. Probability of one indicates that an event is sure to take place and zero, that its occurrence is impossible. Probabilities associated with some events are shown in Figure 4.1. Among these, probabilty of 0.5 for a coin coming up heads on a toss is easily recognisable, for, there is a fifty percent chance of experiments leading to heads or tails. Probability Event

Figure 4- 1: Probabilities of Some Events Probability concepts help us in:

q Expressing ourselves more precisely, through a universally accepted language, thus enabling better communicaion.

Finding answers to certain decision situations by use of simple probability laws. Inferring the characteristics of the whole population from the data obtained from a sample.

Approaches to Probability There are two approaches to probability estimation and interpretation - objective and subjective. The objective approach may be further sub-divided into the classical approach and the empirical approach. Classical Theory of Probability. The classical theory defines probability (p) as:


Number of favourable outcomes Total number of equally likely outcomes

Let us consider the probability of a coin coming up heads on a toss. There are just two possible outcomes, either heads or tails. Assuming that the coin is unbiased, there is no sufficient reason to infer that either of these outcomes is more likely to take place than the other. We thus have a situation wherein there is a total of two equally likely outcomes and the probability of occurrence of one among these, ie heads, (or alternatively, tails) is thus given by p = 1/2 or 0.5. Similarly, the probability of a die showing a onespot on a roll is 1/ 6 as there is just one favourable outcome among a total of six equally likely outcomes which are thrown up by the geometry of symmetry or construction. Although the classical theory helps in defining probability precisely, it cannot be used in most practical situations because it is not possible to talk of total number of equally likely outcomes for want of symmetry. The probability of an RCL gun shot hitting a tank at a specific range is one such case. Empirical Methods. This approach is based on experimentation and helps in overcoming the limitation of the classical theory. Empirical probability may be defined as:


Number of favourable outcomes Total number of trials

To determine probability by this method, the experiment has to be repeated until a stabilised value is reached. A large number of trials are thus required to be carried out. This aspect is illustrated by the results of an experiment shown at Table 4-1 wherein data obtained through repeated trials has been tabulated and the cumulative probability of an ENTAC missile hitting a Patton tank at a range of 1500 mtrs has been worked out as 0.7.

Table 4-1: Cumulative Probability of an ENTAC Missile Hitting a Patton Tank at 1500 mtrs

Figure 4-2 depicts the cumulative probability obtained from the experiment in which trials have been repeated until a stabilized value of probability is reached. In classical as well as empirical approach, probability may be interpreted as the long run relative frequency. Thus probability of 0.7 would indicate that if the experiment is repeated sufficiently large number of times, 70 percent of favourable outcomes or successes may be expected on an average.

Figure 4-2: Cumulative Probability of Hits Subjective Probability. There are situations where experimentation too is not feasible; take for example, the probability of success in an attack that own forces are planning to launch. In such cases, estimation of probability is subjective, based on informed judgement of the commander. In such an assessment, knowledge of whatever information is available, experience of similar happenings in the past and, of course, good judgement on the part of the commander form important constituents. It thus represents the strength of belief of the decision maker that a given event would occur. Subjective probability estimates, although approximate, are useful in employing analytical methods and can certainly be refined through practice and experience and through use of the Standard Gamble Technique (a formal method designed for this purpose). An excellent example of subjective probability is provided by the bookmakers odds on a race course. Probability Laws These help us in determining the overall probability of a set of events, once that of the constituent events are known or established. There are two probability laws - multiplicative and additive. The meaning of these terms and the usefulness of these laws can be best illustrated by an example.

Say there are two firers, A and B, and their respective single shot probability of hitting the target is p (A) = 0.8 and p (B). = 0.7. Since a given firer can either hit or miss the target and the sum of these probabilities must equal unity, the probability of the two firers missing the target would be p (A) = 0.2 and p (B) = 0.3. If both A and B fire one shot each, what is the probability of

Both the shots hitting the target, ie p(A and B)?

The target being hit, or in other words, at least one shot hitting the target, ie p (A or B)?

p(A) = 0.8; p(A) = 0.2; p(B) = 0.7, p(B) = 0.3 Figure 4-3: Probability Laws Both these situations are illustrated on the probability space shown in Figure 4-3. The probability of both A and B hitting the target is given by the Multiplicative Law, that is: p(A and B) = p(A) x p(B) = 0.8 x 0.7 = 0.56 Similarly, the probability of at least one firer hitting the target is given by the Additive Law, that is: p(A or B) = p(A) + p(B) - p(A and B) = 0.8 + 0.7 - 0.56 = 0.94

An alternative approach to the Additive Law is provided by the fact that the requirement corresponding to p (A or B) is met unless neither firer hits the targets. Thus p (A or B) = = = = 1 - P(A and B) 1 - p(A) x p(B) 1 - 0.2 x 1-0.06 0.94

A useful method of pictorially representing the various events in a sequential manner, so as to assist comprehension, is provided by the probability tree. The situation representing the above example in the form of a probability tree is depicted at Figure 4-4 here.

Figure 4-4: Probability Tree The method indicated in alternative approach to the Additive Law given earlier, is particularly useful when an experiment is repeated a large number of times. We shall illustrate this with an example. Let us say an anti-aircraft gun fires 1000 rounds per minute and has a single shot hit probability (SSHP) of 0.005. Two such guns, sited together to protect a vulnerable point, engage an aircraft for 3 seconds. What is the probability of the aircraft being hit?

As long as even one shot hits the aircraft, the purpose is achieved.Thus, the required probability may be determined as follows: Number of rounds fired per gun in 3 seconds = 1000 x 3 60 = 50 Total number of rounds fired (by 2 guns) = 2 x 50 = 100, SSHP = 0.005, Probability of a shot not hitting the aircraft = 1 - 0.005 = 0.995 Probability of all 100 shots missing the aircraft = 0.995 x 0.995 x 0.995 = (0.995)100 = 0.59 Hence, probability of the aircraft being hit = 1 - 0.59 = 0.41 In actual practice SSHP is much less than the figure indicated here. It is little wonder then that rarely is an aircraft downed by anti-aircraft guns. Of course, they do act as deterrents. A General Application It is seen from the previous examples that in order to calculate at least one hit one has to understand the event as a complementary event to the event no hitor all misses and compute the probability by subtracting the probability of all misses from 1. This concept helps us to calculate the effectiveness of number of weapons delivered and conversely to calculate the number of weapons to be delivered to achieve a desired effectiveness. The example given below illustrates this. Let us consider a situation where n missiles are fired, each having an SSKP (single shot kill probability) of 0.7. Thus the miss ..... (100 times)

probability is (1 - 0.7) ie 0.3. In general. the probability of success can be obtained by subtracting the total miss probability (ie probability of all misses) from 1. We can tabulate the results progressively as in Table 4-2.

Table 4-2: Calculation of Success Probability The results can be graphed as shown in Figure 4-5. The important conclusion from the above analysis is that with every additional launch, the increase in overall kill probability gradually diminishes. After a certain point an additional missile does not increase the overall effectiveness substantially. In this example, while second missiles brings about an increase of 0.21 kill probability over a single missile, fourth missile results in an increase of only 0.019 over three missiles. A military commander, therefore, has to view the problem in this manner and decide whether the additional effectiveness is worth the expenditure of an

Figure 4-5: Number of Missiles vs Success Probability additional weapon. In this case, expenditure on more than 3 missiles may seem to be infructuous. Similar arguments help us to calculate the number of weapons necessary to achieve a desired overall kill probability. Suppose the SSKP is given as 0.5 and we desire to achieve an overall kill probability of 0.95 ie 95% satisfaction level, then we could proceed as follows: Given, SSKP = 0.5; Thus, Miss probability = (1 - 0.5) = 0.5 If n is the number of missiles necessary, then we may write, 1 - (0.5) = 0.95, and solve for n. (0.5) = 1 - 0.95 = 0.05 Taking logarithms on both sides. n log (0.5) = log (0.05) log (0.05) n= log (0.5) = 5. = 4.32,
n n

Thus a total of 5 missiles would give us a total kill probability of more than 0.95, ie 96.875%. (1 -0.5n = 0.96875). Dependent Events So far, the discussion has been restricted to independent events, that is, such events where occurrence (or non-occurrence) of one does not affect the probability of occurrence of the other. Thus, the probability of a firer hitting the target is independent of the result of the others fire. In case of dependent events, however, occurrence of one event affects the probability of occurrence of the other. For example, the probability of an aircraft destroying a ground target would be dependent on whether the said target has been provided air defence cover based on missiles. Let us assume the probability of destruction, when there are no missiles, is 0.6 and that the same reduce to half that value 0.3 when the target is protected by missiles. Also, say, our subjective assessment indicates a probability of 0.7 that the air defence is missile based. What then is the probability of destroying such a target? Such a situation may be viewed as that composed of two events. Event A representing that the defence is missile based (A that it is not based on missiles) and Event B indicating destruction of the target (B that the target is not destroyed). The probability space corresponding to this situation is shown at Figure 4-6. The two areas shown shaded represent the probability of destruction depending on whether the target is protected by missiles or not. Each of these probabilities is obtained through the use of the Multiplicative Law. The overall probability of destruction, whether missiles are there are not, is thus given by adding the two shaded areas and so equals 0.39. This may also be called Marginal Probability of destruction. An interesting off-shoot with regard to dependent events is provided by Bayes Theorem which deals with updating the

Total p(Destruction) = 0.21 + 0.18 = 0.39 Fligure 4-6: Dependent Events - Probability of Destruction of Ground Target by an Aircraft

probability of occurrence of one event based on information obtained regarding another related event. Although this concept is of particular interest in defence applications, an advanced treatment of this nature is beyond the scope of this handbook. Conclusion In this chapter, we have dealt with some fundamental probability concepts including approaches to probability and probability laws. The concept of dependent events has also been briefly brought out. Although this knowledge would be helpful in tackling relatively simple problems, the reader is advised to refer to any standard book on statistics for a more complete treatment. A broad understanding of these concepts is also a pre-requisite for dealing with probability distributions that follow in the next chapter.

CHAPTER 5 PROBABILITY DISTRIBUTIONS Introduction The probability of getting a head on toss of a coin is 0.5. If the coin is tossed ten times, we should generally expect five heads. But we know that the outcome of such an experiment could well be anything from zero at an extreme to ten at the other. Intuitively we could also say that the probability of an outcome in the extreme ends of the range would be rather small, its value increasing as the number of heads approaches five. We also know that the sum total of the probabilities over the entire range of outcomes must equal unit. A plot of each of these probabilities would tell us how the Probabilities are distributed with respect to outcomes. Such a plot is called a probability distribution. Most of our planning is based on averages or means. The time to lay a bridge, useful life of an equipment and the percentage of casualties to be expected in an operation of a given intensity are just some examples. However, by definition, the mean implies a value that would be exceeded about half the time. In all such planning, therefore, our assurance level is of the order of 50 per cent only. Surely, in critical areas we need a much higher assurance level than that. This, and many other useful aspects of planning can be effectively dealt with through a knowledge of probability distribution. Two distributions of particular interest to us are the Normal and Poisson.

Normal Distribution Perhaps the most important probability distribution in statistics is the normal distribution. This is because it is applicable to a great many situations including human characteristics (weights, heights and lQs), outputs from physical processes (dimensions and yields) and other measures of interest to management. Its physical appearance is that of a symmetrical bellshaped curve, extending infinitely far into positive and negative directions. Of course, most real-life populations do not extend for ever in both directions, but for such populations the normal distribution is a convenient approximation. Parameters: Mean and Standard Deviation. In fact, there is a family of normal curves. To define a particular normal curve we need two parameters; the mean (m) and the standard deviation (s). Because of the symmetry of the normal distribution, its mean (as also the median and mode) lies at the centre of the normal curve. In other words, the mean places the centre of the horizontal axis, and the standard deviation governs, the dispersion of values. Figure 5-1 shows three normal distributions, for the time taken for bridge laying by three; A, B and C engineer regiments. Each of them has the same mean ie 30 mins but a different standard deviation of 1, 5 and 10 mins.

Figure 5-1: Time for Bridge Laying - A, B and C Engineer Regiments

If we were to take a large enough population of soldiers and draw the frequency curve of their heights, we should expect a bellshaped curve. In a comparison of height between Sikh and Gurkha troops, the following should be intuitively visualised:
q q

Sikhs being generally tall, would have a higher mean. The curve for Gurkhas would show smaller dispersion as variation in height of Gurkhas is relatively less.

Area Under the Curve. No matter what the values of m and s are, for a normal probability distribution the total area under the curve is taken as unity; thus we may think of areas under the curve as probabilities. Mathematically this means that 68.3 per cent of all values in a normally distributed population be within one standard deviation (plus and minus) from the mean, 95.4 percent within two standard deviations and 99.7 per cent within three standard deviations. This is shown graphically in Figure 5-2. Practically, 100 per cent of the area may be considered to lie within three standard deviations (plus and minus) from the mean. The graph clearly shows the relationship between the area under the curve and distance from the mean measured in Standard Deviations, for a normal proability distribution.


5-2: Relationship Between the Area Under the Curve and the Distance from the mean measured in

Standard Deviation, for a Normal Probibility Distribution

It is neither possible nor necessary to have different tables of areas for every possible normal curve. Instead, we can use a standard table of areas, under the normal curve. With this table, we can determine the area; or probability, that the values of a normally distributed variable will he within certain distances from the mean. These distances are defined in terms of Standard Deviations. Z Table. The table at Appendix B at the end of this handbook shows the area under the normal curve between the mean (m) and any other value (X) of the variable. The value X of the variable with which we are concerned is required to be converted into another variable Z for using the table. The value of Z representsthe distance from the mean (m) in terms of standard deviations, and so is derived from the formula: X- m Z= s Thus, if m = 50 and s = 10, the probability of the variable lying between 50 (ie the mean) and 60 (ie Z = 1) will be 0.3413, as indicated in the table for a value of Z equal to 1.00. Assurance Level Let us say past data tells us that Engineer Regiment A takes a mean time of 210 min to lay a bridge and that

Figure 5-3: Time Required for 95% Assurance Level

the standard deviation is 10 min. We wish to lay down a suitable time for bridge laying in the operation order so as to be 95 per cent confident that the task will be completed in time. The normal curve for Regiment As timings is shown as Figure 5-3. As the curve is symmetrical, only half the area under the curve would correspond to timings upto 210 min. In other words, a time span of 210 min provides only 0.5 probability, or 50 per cent assurance of the task being completed in time. To increase the assurance level we must cater for more time. The required assurance level being 95 per cent, an increase of 45 per cent, or 0.45 in terms of area, is desired. Referring to the table we find that this (0.4495) = 0.45 corresponds to a value of Z = 1.64. Thus time is required to be increased by 1.64 (s) or approximately 16 min. Therefore 226 min should be laid down for completion of this task. Comparison of an Assurance Level. With a similar treatment we can also compare the performance of two regiments. Let us assume that similar data for B Engineer Regiment indicates X = 200 min and s = 20 min. Although the mean time taken by B Regiment is less, the time this unit requires to provide 95 per cent assurance is 200 + 1.64 x 20 ie 233 min. The case for both the regiments is shown in Figure 5-4. on next page. As can be seen, the choice is in favour of A Regiment - an alternative which is not obvious without this analysis.

Dispersion in Mean Expected Life : An Application. Let us consider yet another application. Say, a new vehicle is being introduced, its mean expected life being 6 years and Standard Deviation one year. We wish to determine the year wise wastages, so that their replacement can be planned ahead. We know that the entire useful life may be assumed to spread over 3 Standard Deviations either side of the mean. Thus, wastages will take place

Figure 5-4: Comparison of Two Normal Distributions at Given Assurance Level between the fourth and the ninth year. Taking approximations for the values indicated earlier to be 68, 95 and 100 per cent respectively for 1,2 and 3 Standard Deviations (plus and minus) away from the mean, the wastages can be easily worked out as percentages of the number introduced. This is illustrated in Figure 5-5.

Figure 5-5: Wastage of Equipment Dispersion Application Areas. Typical application areas of normal distribution are:

Performance of man and machine, eg: Time taken to complete an activity or mission (bridge laying / crossing, movement).

- Distribution of rounds and/or bombs on target.

q q

Equipment life cycle, wastages and replacements.

Equipment Mean Time Between Failure (MTBF) and reliability.


Wastage pattern of general stores.

Poisson Distribution Poisson Distribution deals with events that take place in a continuum of time or space. An accident or a flat wheel of vehicle is a typical example. It can take place at any moment or point in a continuum of time or space, given a certain environment - say, a given highway. The probability of its occurrence in a short interval (say, 1 min or 100m) is small and remains constant, and although we can easily count the number of occurrences over an interval, it makes little sense to talk of non-occurrence of such an event. Basically such events are also rare events. Parameter : The Mean. Unlike the normal distribution, Poisson distribution is fully defined by just one parameter, the mean (m). To take an example, let us assume that vehicles plying between Chennai and Delhi have, on an average, two flat tyres. Then, although the mean equals two, there would surely be vehicles that have no flat tyres at all, those that have just one and others that have two, three or more flat tyres. There would also be cases where vehicles have as many as, say six flat tyres. Of course, we can say intuitively that the proportion of vehicles with six or more flat tyres would be extremely small. Poisson distribution tables (not attached here, but available in any standard book on Statistics) give us the probability associated with each of these occurrences. The case for m = 2 is graphically represented at Figure 5-6. As can be seen there is a sharp fall in probabilities beyond two occurrences. An Application. We shall now illustrate a typical application of Poisson distribution. Let us say that at an aircraft repair depot, the

Figure 5-6: A Plot of Poisson Distribution for m - 2 requirement of aircraft engines over a period of ten months is 32, ie 3.2 is monthly mean requirement. Stocks are replenished monthly. We wish to determine the minimum stock holding to provide an assurance level of 95 per cent. The values of probability as tabulated in Table 5-1 can be obtained from Poisson tables for m = 3.2:

Table 5-1: Poisson Probabilities

As can be seen, the minimum stock holding for 95% assurance level is six engines. If the stock held is five, the assurance provided is 89.46% ie approximately 90%. Application Areas. Typical application areas for Poisson distribution are in the estimation of :

Accidents, battle casualties. Failure of components and consequently assessment of demands for spares and the stocking policy. Arrivals at servicing facilities such as workshops and airfields. Applications pertaining to the field of arrivals have been dealt with separately under Queueing (Chapter 11: Waiting Lines).

CHAPTER 6 INFERENCES BASED ON SAMPLE INFORMATION Introduction Decision making could be based on qualitative aspects such as intuition, hunch, knowledge, experience etc or quantitative inputs in the form of data. As statistical techniques basically handle quantitative data, qualitative data is required to go through a certain process of quantification for any worthwhile analysis and interpretation. When qualitative information is in the form of claims, guesses, bunches, opinions or,unauthenticated authoritative statements, acceptance or rejection of such information is vital for a decision situation. How does one go about the process of quantification and subsequent analysis under these circumstances in order to determine the validity of these claims? Let us consider a few examples of such claims in defence situations: A manufacturer supplying vehicles to the Army claims that his vehicles can meet the cross-country capability as per the General Staff Qualitative Requirements (GSQR).
q q A foreign manufacturer claims that Single Shot Kill Probability (SSKP) of Anti-Tank Guided Missile (ATGM) developed by them is better than 0.7. q There is no difference in the firing performance of different types of RCL Guns. q A type of troops have better physical endurance in mountainous terrain than B type.

It is claimed that difference in firing performance of guns is due to capabilities of firers and not due to characteristics of the guns.

For normal conversational purposes, such statements may be acceptable but when decisions are to be made in operational situations, the decision process would pivot on the true values of their capabilities or characteristics. It would be apparent that it is not possible to verify the accuracy or the truthfulness of the claims which are under scrutiny. Conclusion or inference, perforce has to be made based on testing a portion or sample selected randomly from the entire lot. In statistical terminology, making an inference about the population based on the sample evidence is called Inductive Reasoning and a statistical technique called Hypothesis Testing is used to test the accuracy of such judgements which are based on sample information. HYPOTHESIS TESTING Hypothesis means a law which is not universally accepted. A claim may, therefore, be considered as a hypothesis. In Hypothesis Testing, the mean value of randomly selected sample elements is compared with hypothesised mean value. If the difference is statistically significant, the claim is rejected, otherwise the claim is accepted to be true. This is the process of inductive reasoning which implies that, based on samples information, the accuracy of the hypothesised value of a lot or population is inferred. In this process, there is bound to be certain degree of error which may be quantified to assist the decision maker. Let us consider a few examples to understand the technique. Vehicle Manufacturers Claim Say, the Service HQ have evolved a QR for 2.5 Ton Vehicle for use in the services and have laid down 400 km cross-country performance with dispersion 370-430 km. Vehicle manufacturers claim is that his vehicle has a mean cross - country radius of action

of 400 km. Army is planning to procure, say, 500 vehicles. The manufacturer also has indicated that the radius of action would meet the QR and has offered a sample of 36 vehicles for trial. The performance of the vehicle may be expected to follow normal distribution as shown in Figure 6-1 with the mean value of 400 km and Standard Deviation (or dispersion factor) of 10 km assuming that the manufacturers claim is correct.

Figure 6-1: Cross Country Radius of Action (Normal Distribution) Individual vehicles cross country performance would be between 370 and 430 km The implication of performance of vehicles following normal distribution is that if there are 100 vehicles, 68 would have mean value of cross-country radius of action between 390 and 410 km, 95 vehicles between 380 and 420 km and 99 vehicles between 370 and 430 km. The mean value of radius of action is 400 km and its standard deviation is 10 km which is within the specifications as per the QR But how do we validate the claim of the manufacturer? Obviously, we cannot test the performance of each and every vehicle in the total population of 500 vehicles supplied by the manufacturer. He has offered a sample of 36 vehicles for trials. Let us assume that the mean obtained with this sample was 380 km. With this information should the manufacturers claim be accepted? On the face of it,

it appears that 380 km is close to 400 km claimed by the manufacturer and hence the claim could be accepted. Based on theory of sampling, individual values of the sample should also vary between 370 and 430 km and their mean should be close to 400 km but in the present case it is 380 km which means individual values could vary between 350 and 410 km assuming a dispersion factor of 10 km. Therefore the manufacturers claim of mean value of 400 km could turn out to be wrong as may be seen in Figure 6-2.

Figure 6-2: Manufacturers Claim and Sample Results Let us statistically analyse further. Based on sampling theory and discussion earlier, individual values of performance of the 36 vehicles which we had taken as sample, could vary between 370 and 430 km and the sample mean close to 400 km. It is quite possible that mean value of 380 km obtained could be the actual mean of the population with individual values varying between 350 km and 410 km or perhaps, in reality the mean value was 390 km with values varying between 360 and 420 km. See Figure 6-2. Let us view the problem a little differently, with the following information:

As per manufacturers claim, the population has a mean crosscountry radius of action of 400 km.

The sample of 36 vehicles after trials, yields a mean of 380 Km.

The question is whether this sample under consideration could have come from the population whose mean radius of action is 400 km, or not. To answer this question, we must study the statistical behaviour of all possible samples which could be drawn from the population in question, and check whether our sample conforms or does not conform to such behaviour. If we examine all possible samples of size n from a population of size N and study the distribution of all the sample means, we find that such distribution known as Sampling Distribution is characterised by a mean which equals to the population mean and a Standard Deviation which is the population standard deviation of the sampling distribution is known as Standard Error, (SE). Let us understand above by using our example:

Population Parameters.
q q

Range of cross country radius of action = 370 to 430 km Population Mean = 400 km

q q

Population Standard Deviation = 10 km Sampling Distribution Sample Size (n) = 36

Mean of the sampling distribution = Population Mean = 400 km

Population SD 10 10

Standard Error =

= 36 6

= 1.66

The Sampling Distribution of Means of all samples of size 36 drawn from a Population with its mean equal to 400 kms is shown in Figure 6-3.

Figure 6-3: Sampling Distribution of Sampling Means If we construct an interval of 3 SE about the mean of 400 km, we know that almost all sample means (ie 99.7%) would be within this interval. Thus, we may consider 395.02 km as the lowest acceptable limit below which we will not accept any sample mean. In other words, if any sample mean is below 395.02 km, it is highly unlikely that this sample could have come from a population whose mean is 400 km. In the present case, the sample mean is 380 km and thus the manufacturers claim is to be rejected. Sometimes, depending on the tolerance level of the decision maker, the critical limit could even be placed at 396.68 ie at - 2 SE from the mean. In this case in 95.5% instances, we will not go wrong if we reject the claim. However, we will commit an error of 2.25% in we demarcate our rejection region as shaded in Figure 6-3. In other words, there will be 2.25% instances when due to randomness, typical sample from the claimed population may yield a mean which could be very close to 396.68 but fall in the rejection region. In such cases because of our decision rule, we shall reject the claim even if the

claim is justified and true. This error is known as the level of significance and is always decided in advance before conducting the hypothesis test. Manufacturers Claim that SSKP of ATGM is 0.7 There are many situations where due to reasons of cost and complexity, verification of claim has to be carried out with small samples. Consider the case where a manufacturer has offered only 10 ATGMs for testing and we shall assume that there were 5 hits. Do we accept the claim that the SSKP is 0.7. If we do not accept, what is the degree of risk we are committing in rejecting the claim (or hypothesis) which in actual fact was true? Such cases involve small sample sizes. Depending on the nature of investigation, a binominal or a t-distribution would need to be used for the hypothesis testing. These cases, however, are not discussed as they are beyond the scope of this elementary handbook. Checking the Firing Performance of Guns The claims so far discussed could be verified with quantified data available. However, certain claims could be in the form of qualitative statements and their accuracy would require to be validated. For example, verifying the claim that there is no difference in the performance of four types of recoilless (RCL) guns. The guns are taken on the range and made to fire same or different number of rounds on a standard target at a given effective range. The results obtained are shown in Table 6-1.

Table 6-1: Gun Performance Data

Looking at the performance of these guns, one finds that the results are quite close to each other but can one say with confidence that there is no difference in their performance? A statistical technique called the Chi Square (X ) test is used to verify such cases. This technique is applicable to test whether all the four guns belong to a population of similar characteristics or not and whether the difference between their firing performance is significant or not. This implies that variation in the firing could be due to characteristics of guns or chance element Depending on the cumulative value of the difference, it can be verified whether these samples belong to the same population or not. The values given in the tables are the observed values F0 of hits and misses obtained out of total round fired. These observed values are compared with expected values. The concept of expected Fe is that assuming that guns kill probability is 50% and if all guns are absolutely identical, then proportion of hits to misses will be same in all cases. If kill probability is 60% then for all guns the proportion of hits to misses will be in the proportion of 60:40 in all cases. Any difference in the firing characteristics will create more deviation from the expected values. In the given data, the expected values are calculated based on this principle, except that the value of kill probability is not known at this point in time because its value at the induction stage may have changed over a period of time. Kill probability and expected values in such cases may be calculated based on the collected data. Say, there were 102 hits out of 192 fired. Probability of hit p(Hit) =102 192 In case of Gun A, 40 rounds are fired Expected value of hit = 40 x 102 = 21.3 192

X Tables are available for various levels of significance. A level of significance of 5% gives the extent of risk of committing an error of rejecting a true hypothesis - called Type 1 error. Degrees of freedom is obtained by the formula (m - 1) x (n - 1) where m is number of rows and n is number of columns in the table. In the above case degrees of freedom is (2 - 1) x (4 - 1) = 3. At 5% level of significance and 3 degrees of freedom, the val of X 2 from the table is 7.81. The computed value which in this case is 1.76 is compared with the table value. If it is less, then there is no difference in the firing performance of these guns. If the computed value is more, then there is a difference. Conclusion in this case is that there is no significant difference in the firing performance of these guns. Conclusion An attempt has been made in this chapter to bring out relevance of some of the useful statistical techniques such as Hypothesis Testing and Chi-Square Test in validating and interpreting qualitative information when they are in the form of claims, opinions or authoritative statements. These are rather important in the defence context as there are circumstances where accuracy of the statements need to be verified at a given measure of error. Awareness of relevant statistical techniques is essential for commanders at all levels when faced with such decision situations.

SECTION 2 CHAPTER 7 OPERATIONS RESEARCH INTRODUCTION Introduction Operations Research (OR) may be defined as the application of the scientific method to provide the decision maker with a quantitative basis for decisions regarding complex operations under his control. OR quantitatively, evaluates alternative courses of action which the decision maker can take, and then provides this information to him so that he can choose that alternative which will be in the best interest of the organisation as a whole. There are basically four distinguishing characteristics of OR. First, it is broad in scope. It is concerned with the organisation as a whole rather than with sub-organisation units. Furthermore, it is applicable to any organisation having definable objectives, where the attainment of these objectives is subject to limiting constraints and there are alternative ways of pursuing them. Second, it emphasises the development of an analytical technique for solving individual problems rather than the application of general models. No two problems are exactly alike. Therefore, it attempts to develop a model to fit the problem rather than modify the problem to fit a model. In the development of models, OR leans heavily on mathematics and other basic sciences.

Third, it uses a team effort in dealing with problems. While it may often be possible for one person to do OR work, many problems are of such dimension that the team approach frequently results in faster and more efficient method of obtaining a solution. The team should bring together persons from various disciplines and backgrounds and with different experiences. Fourth, it is concerned with the practical management of an organisation. When a scientist tackles a real world problem, he seeks a solution in the light of his own background and experience. But when the solution is obtained it must be presented in the language of the real world. In addition to the use of the scientific method,OR requires that definite conclusions be drawn, and that these conclusions be communicated to the decision maker. Operations Research, neither is, nor does it claim to be a magic formula which can be manipulated to yield the ultimate unique solution, it is merely a logical systematic approach to , provide a rational basis for the decision makers decision. Specifically, OR applies the scientific method which is simply a logical systematic approach for conducting research or solving problems. The scientific method for solving OR problems consists of the following steps:
q q q q q q

Formulate the problem. Construct a model. Obtain a solution by the model. Validate the solution given by the model. Establish controls over the model for an implementable solution. Implement the solution.

Some of the techniques of OR are discussed in the succeeding chapters.

CHAPTER 8 LINEAR PROGRAMMING Introduction Linear Programming (LP) is concerned with the study of interrelated components of a complex system so that limited resources can be used as efficiently as possible in attaining a desired objective. It was during the World War II that the theory of LP was developed. In order to maintain the huge war effort, it was crucial that troops be recruited, trained, and deployed to battle on schedule and the vital equipments be produced as needed. All of this, of course, required extensive planning. An USAF research group, later known as Project SCOOP (Scientific Computation of Optimum Programs), developed elaborate planning techniques, which shortly after the war resulted in the theory of LP. George Dantzig, one of the leaders in this research group, is credited with being the father of LP In peacetime economy, an effort was made to apply LP to nonmilitary uses. It soon became clear that the sophisticated organisational procedures developed by Project SCOOP had wide application in industry also. Significant uses of LP were made in operating petroleum refineries, exploring and producing oil, planning efficient transportation routes of manufactured goods, producing steel, awarding contracts etc. In fact, LP is so widely used today in military, government, commerce and industry that it would be impossible to list all its many applications. If the value of a theory is measured by its ability

to solve problems, then surely LP must be considered one of the most important areas of OR. Requirements of LP Problems The following requirements must be met before the LP technique can be applied: There must be an objective to achieve and which could be in terms of optimum resources allocation, maximise profit, minimise cost, maximise operational effectiveness or minimise casualties.

There must be alternative course of action - one of which will achieve the objective. Resources must be in limited supply.

The objective and constraints must be expressed in mathematical linear equations or inequalities. An Example - LP Graphical Two types of aircraft are available for a bombing mission, against dug-in and evenly dispersed enemy troops. Bombs are plentiful but aircraft, fuel and maintenance time are limited. The data is available in Table 8-1.

Table 8-1: Mission Parameters Only 400 KL of fuel is available for this mission. No more than 30 planes can receive the required maintenance prior to departure time. How many planes of each type should be sent on the mission so as to maximise bomb tonnage delivered?

The above problem can now be formulated as follows:

These inequalities can now be plotted as shown at Figure 8-1.

Figure 8-1: Linear Programming - Graphical Solution

Using the Extreme Point Theorem (proof not being provided here), it is known that in a convex polygon OPQRTS, the optimum solution lies at one of the vertices. Therefore the bomb loads delivered at points, P, Q,R, S, T by substituting the values of variables A and B (two types of aircraft) at these points, will be: Pt Pt Pt Pt Pt T(26, 0) T(26, 4) R(20, 10) Q(10, 15) Q(0, 15) Z Z Z Z Z = 2 x 26 = 2 x 26 = 2x 2O = 2 x 10 =2x0 + 3 x 0 = 52 + 3 x 4 = 64 + 3x 10 = 70 Optimum + 3 x 15 = 65 + 3 x 15 = 45

The optimum solution is at R ie we must send 20 of Type A and 10 of Type B aircraft for this mission to deliver a maximum of 70 tons bomb load. This solution as you would have noticed does not violate any of our constraints. Summary of the Solution Procedure The solution by graphical method involves three steps:

Step 1 - Transform the given technical specifications of the problem into inequalities and make a precise statement of the objective function. Step 2 - Graph each inequality in its limiting case and obtain a region of possible solution ie the solution space.

Step 3 - Note the values of variables at all vertices and select that vertex which gives the highest value of Z (objective function).

Comments on the Graphical Method The graphical method of solving LP problem is limited to cases in which the variables are two or three. Even for three variables, we move into a three dimensional space and the visualisation of the problem becomes more difficult than in case of two variables.

But what about problems involving variables greater than three? Most problems in actual practice do fall under this category. These problems can be solved by the Simplex Method which is the most general and more powerful. It is not being presented in this handbook, as it is fairly involved technique to be solved manually, but can be quite easily solved with the help of commercially available OR computer software packages. Some examples of military problems which can be solved using the LP technique are given below along with the formulation of the LP equations, which constitute the crux of the problem solution. Once these have been correctly formulated, the final solution can be easily arrived at using any computer software for LP like QS, CMMS, TORA etc. AIRCRAFT ALLOCATION A decision has been taken, as a strategic move to strike a fatal blow to the enemys capacity for aircraft production. The enemy has four major plants located at different cities, and the destruction of any one plant will effectively halt the production of aircraft by the enemy. The Bomber Command is faced with shortage of fuel and the fuel supply for this particular mission is limited to 240 KL. Any bomber sent to any particular city, must atleast have enough fuel for the round trip plus a reserve of 500 litres. The number of bombers available to the command their description and fuel consumption are given at Table 8-2 here.

Table 8-2: Data on Aircraft Information about the location of enemy production plants and their vulnerability to attack by bombers is as in Table 8-3.

Table 8-3: Enemys Plants and Vulnerability

The requirement is to plan the number of bombers required to maximise probability of destruction of enemy targets. ARMY TRANSPORTATION Army is planning to carry out the trials of the newly developed main battle tank at five areas (operational bases) spread over west, north and east India. These tanks can be supplied from three depots. The requirements of tanks for trials at five bases, the availability of tanks at the three depots and the distances in km from the bases to the depots are given in Table 8-4 as follows.

Table 8-4: Depot-Base Distance, Availability and Requirement of Tanks

Find a schedule to transport the tanks from depots to bases that will minimise the total distance involved in meeting the requirements of tanks. NAVY WEAPON MIX Indian Navy in considering three types of attack aircraft - a supersonic type, a transonic type, and V/STOL type for its aircraft carrier. The effectiveness of the aircraft to the carrier is determined by the relative expected military value of targets that it can destroy during a military engagement of a certain length. These have been estimated for the three types and mentioned in first row of Table 8-5.

Table 8-5: Expected Values and Requirements of the Three Aircrafts If the entire deck space available for aircraft on the carrier is allocated to one specific type, the numbers of aircraft that could be accommodated are given in Table 8-5, second row. Personnel requirements and costs are also included in the table.

Considering that a carrier has facilities for 1500 attack aircraft personnel and the Navys planned expenditure for attack aircraft for the carrier is 325 crores, the problem is to determine how many of each of the three types of aircraft should be purchased for the carrier in order to maximise the value of its attack capability. SOLUTIONS : FORMULATION OF LP EQUATIONS Aircraft Allocation


Introduction A typical Transportation Model deals with a special class of linear programming problems in which the objective is to transport a single commodity from various origins to different destinations at a minimum total cost. The total supply available and the total requirement are given. Further each allocation has a known origin and a known destination. Also all relationships are assumed to be linear. Transportation problems being a sub-set of the general LP problems can be solved by the simplex method. However, the transportation algorithm provides a much more efficient method of handling such a problem. To state explicitly, before transportation model can be applied, following assumptions/conditions must be made:

All relationships are assumed to be linear. It deals with only one (single) commodity which is required to be moved from different origins/depots, where it is available, to various destinations/units where it is required.

A limiting factor is that all demands at destinations must equal total availability at various origins. This, however, does not pose any problem as the two aspects can be made equal by a simple method which is described in paragraph titled Balancing Demand and Supply in this chapter.

like the general LP model, we can solve the problem for a single objective at a time eg, minimise cost, distance or time etc.

Basic transportation model deals with a minimisation case. In order to handle a maximisation case (say the objective may be to maximise profits) the approach has, therefore, to be slightly modified which is described in paragraphs titled Maximisation Case in this chapter. Making Initial Transportation Tableau A pictorial representation of the mathematical description is available through a transportation tableau. The transportation tableau is a matrix with rows (representing sources of supply) and columns (representing destinations to satisfy demands). Therefore, each cell in the matrix represents a unique routing for shipments from sources to destinations. The figures on the right and bottom sides of the matrix show, respectively, the amount of supplies available at each source and the amount required at each destination. The objective of the problem is to identify the amount to be shipped from each source to satisfy the demands at each destination with the minimum transportation costs. Let us see this with the help of an example. Total of 21 tanks are available at three origins and are required at five destinations and the cost of transportation from each origin to various destinations is known and the data is arranged in Table 9-1.

Table 9-1: Transportation Cost-Per Tank (Rs thousands) with Availability and Demand

Balancing Demands and Availability In the above example, the demand and availability are both equal so the tableau prepared above is ready for further operation. However, in real life such situations would rarely occur. An example of this is at Table 9-2. The procedure adopted for balancing the demand and availability is as under:

Table 9--2: Unbalanced Matrix Demand Exceeds Supply. For purposes of solution of the problem, we create a fictitious source or resource where the supply available is just sufficient to meet the excess of the demand over the actual supply. This source of supply is called a dummy supply. Transportation cost from a dummy supply to any destination is naturally zero because no actual transfers to any destination can take place. Also if in the solution a given destination receives a supply of d units from the dummy and s units from an actual supply, it has been in reality short-supplied to the extent of d units. As is evident from balanced matrix at Table 9-3, there would be a short fall of 50.
q q

Supply Exceeds Demand. Similarly as in the above case a fictitious consumption point is created for the sake of balancing and solving the problem. The requirement at the dummy destination is exactly equal to the difference between actual supply and demand. Once again the cost of this factitious transfer is zero. If a source is deemed to have transferred d units to the dummy, it is interpreted that source has a left over stock of d units.

Table 9-3: Balanced Matrix VOGELS APPROXIMATION METHOD (VAM) Vogels Approximation Method gives near optimal or optimal solutions. VAM is based on the use of the difference associated with each row and column in a matrix giving the unit cost. A row or column difference is calculated as the arithmetic difference between the smallest and the next higher cost element in that row or column. This value provides a measure of the priorities for making allocations to the respective rows and columns, since it indicates the minimum unit penalty incurred by failing to make an allocation to the smallest cost cell in that row or column. Thus, this procedure repeatedly makes the maximum feasible allocation in the smallest cost cell of the row or column with the largest difference. It never changes an allocation previously made, so the procedure stops as soon as the total demand and supply are exhausted. The detailed steps involved in Vogels approximation method are summarised here: Step 1 - Construct the cost-data table like Table 9-2 for the problem. Following is the detailed structure of this table:
q q

One column listing the sources (at the far left of the table), one column for each of the (n) destinations, and one column for the supply entries (at the far right of the table).

One row identifying the destination points (at the top of the table), one row for each of the (m) sources, and one row summarising the demand values (at the bottom of the table).

Small inset boxes for entering the costs associated with each source-destination pairing or cell.

q The table is complete if the sum of the supply column equals the sum of the demand row. If the sum of the supply column does not equal the sum of the demand row, adjust the table as follows. For total demand greater than the supply; add a dummy row with cost entries equal to zero and the row requirement set equal to the excess demand. For total demand less than total supply, add a column with cost entries equal to zero and the column requirement set equal to the excess supply.

Step 2 - For each row, calculate the difference between the lowest and next higher cost entries. (If the two lowest cost elements are equal, the difference is zero). Step 3 - Similarly, for each column calculate the difference between the lowest and next higher entries (If the two lowest cost elements are equal the difference is zero). Step 4 - Select the row or column that has the greatest difference. In the event of a tie, the selection is arbitrary. Step 5 - In the row or column identified in Step 4 select the cell that has the lowest cost entry. Step 6 - Assign the maximum possible number of units to the cell selected in Step 5 (ie, the smaller of the two between the availability row and demand column). This will completely exhaust a row or a column (perhaps both). Step 7 - Reapply Steps 2 to 6 using the remaining rows and columns. Repeat the process until total supply has been

exhausted. If the calculations are correct, this will also exhaust total demand. Illustrative Problem-Solution The above method is applied to the tank transportation problem whose data has been given in Table 9-1 earlier. The iterations are explained below in reference to Table 9-4. Note: For ease of understanding, the iteration number is indicated on left hand top comer of the cell in which allotment is made.

Table 9-4: Transportation Table - VAM Solution Iteration 1. The highest difference is 10 in column two. Hence allot the maximum feasible amount to the lowest cost cell (02, D2) in this column. The availability in row two being 4, only 4 can be allotted to cell (02, D2). The balance requirement in column two is also amended to 1, quantity 4 having been allotted. Row two is now eliminated from further consideration and the remaining matrix is operated upon as per above procedure in the next step.

Iteration 2. Against maximum column difference 9, maximum feasible quantity 3 is allotted to cell (03 , D1,) and zero put in requirement of column one. Column one is now eliminated from further consideration. The above procedure is continued till all allocations have been made. The total cost of transportation works out to 150 units of money. It might appear that the procedure is tedious but after a little practice, one could probably execute this entire procedure on the original cost and requirements table. One would need merely to cross out rows and columns as they are completed and revise the supplies and demands. Salient Features
q Follows principle of opportunity costs ie, it examines the cheapest route from each origin to different destinations and to reach the destinations from different origins.

Compares opportunity cost or penalty for not following the cheapest route.

Makes maximum allocation to that route which has high penalty.


Gives a solution which is optimal or very close to it. Procedure is tedious but not difficult. Being repetitive in nature lends itself to easy handling by computers.

Maximisation Case Though solution methods for transportation basically cater to minimisation of costs, occasionally, a problem is most naturally stated in terms of expected profit instead of cost of transportation. This situation is accommodated in the transportation method by treating profits as negative losses and then applying minimisation techniques.

A convenient method for converting a profit statement to cost relationships is to subtract all stated profits from the highest profit; the resulting numbers are relative costs and are treated as regular costs in solving the problem. For example, if the largest profit in any resource transfer is Rs 100 (or 100 units), all transfer profits would be subtracted from 100. This policy gives 100 - 100 = 0 as the relative cost for the more profitable route, which of course, makes it the most attractive route, for cost minimisation. Similarly the lowest profit route would have the highest relative cost. After all source-to-destination transfer charges are calculated, the methods described for a minimisation case can be used to develop a solution. From the preferred routes indicated in the solution, the total profit expectation can be calculated by using original profit figures. Modified Distribution Method (MODI) Transportation problems are a special class of linear programming problems. VAM gives a solution which is very close (or even in some cases equal) to the optimal solution. For an optimal answer one may make use of MODI. It iteratively improves on an initial basic feasible solution obtained by any one of the approximate methods, including VAM. It is based on a cell evaluation method of an empty cell, keeping the rim requirements ie supply and demand intact. Readjustments to previous allocations are made iteratively till finally optimal solution is obtained. The commercially available OR computer software packages give optimal solutions.

CHAPTER 10 DECISION THEORY Introduction In every walk of life, be it at individual level or extremely complex modern organisational level, we are concerned with taking effective decisions. Administrators in the government, commanders in the services, managers in the public or private sectors, irrespective of their functions or level, are all involved in the decision making process. Managerial Decision Making It is a process whereby a commander or a manager when confronted with a problem, chooses a course of action from a set of possible alternative courses available. Decision making is a response to a problem situation which generally arises as a result of a discrepancy between prevailing conditions and organisation goals. Uncertainty will always be there about the future since one cannot be absolutely sure of the ultimate consequences of the decision made. Elements of a Decision In a problem situation where decisions are to be taken, the following elements can be identified:

Decision maker - who may be an individual or a group. A set of possible and viable alternative courses of action

q A set of states of nature or occurances which may or may not be predictable.

A set of consequences or effects related to each possible course of action depending upon the state of nature (causeeffect relationship).

The complexity of a decision problem would depend upon the nature of decision element It is extremely difficult for a decision maker to intuitively take into consideration, the implications of consequences of each possible course of action vis-a-vis states of nature. Decision analysis provides the necessary base for defining, analysing and solving a problem in a rational, logical and scientific manner based on data, facts, information and logic when dealing with complex decision problems. The focus of such an approach is to study as to how decisions should be made rather than how people actually make decisions. Pay Off Matrix Pay Off. Decision making is choosing a course of action from among the alternatives. Effect or consequence of a course of action would depend upon the state of nature. Consequences are; that could be measured like profit number of targets destroyed, number of casualties or losses. Such outcomes could be viewed as pay off of a decision. But before taking a decision, pay off values could be assigned for the different courses of action to facilitate choice among the alternatives. Decision Situations. All decision situations fall under one or more of the following categories:

Certainty. For a set of given initial conditions, future events can be predicted with certainty. Risk. Although future cannot be predicted, it is possible to assess or predict the likelihood (or probability) of future events.

Uncertainty. Where probability of future events cannot be assessed.

Assignment of Pay Offs. Assignment of pay off values for certainity situation is comparatively simple. For example, in an investment decision of whether to invest in a bank, reputed company or in shares, the return on investment could be viewed as pay offs. The choice as to where to invest would, however, depend upon the risk profile of the decision maker. Similarly, measures of effectiveness such as target destruction potential (number of targets that can be destroyed) of a weapon or fire unit can be assigned with reasonable level of assurance which could be viewed as direct measures of the pay off values. But in many of the defence situations, where future events cannot be predicted, we will be dealing with situations of risk and uncertainty. In such situation pay offs may have to be assigned subjectively. In order to reduce individual subjectivity, the concept of participative decision making may be resorted to, wherein personnel involved in the decision making process and in its subsequent implem entation are associated in the assignment of pay off values. Decision Matrix Decision matrix is an extension of pay off matrix and there are two main aspects involved:

Assigning of probabilities to the occurrence of different states of nature. Assigning of pay offs to various cells in the matrix.

For assigning probabilities, methods available are:


Determining how frequently similar events have occurred in the past, in case such data is available and thus calculate the probability. Available intelligence about the states of nature. Informed judgement of the decision maker based on his knowledge, professional skill and experience.

Basic Steps for Decision Matrix


Listing of all viable alternatives. These are also called strategies as these are under the direct control of the decision maker. Listing of events affecting these alternatives. There are called states of nature and are beyond the control of decision maker. Construction of pay off matrix by assigning pay off values. Assigning probabilities to the states of nature. Calculation of expected value of each alternative. Presenting the expected value of each alternative to the decision maker for taking the decision.

q q q q

In normal military appreciation, although there may be a large number of own and enemy courses of action, by applying value judgements, the choices are narrowed down. However, decision matrix poses no such restrictions and can simultaneously analyse a number of own and enemy courses of action. In fact, more complex the situation greater could be the advantage of using a decision matrix. Methodology of analysing a risk decision situation using Decision Matrix will now be considered with the help of an example. Sketch at Figure 10-1 shows three different approaches for attack and capture of an enemy post.

Figure 10-1: Alternate Approaches for Attack

Approaches for Attack (Alternatives)


A3 - Longer than Al. Enemy position tackled from flank. No minefield encountered. Cover available. Takes less time than A2. Enemy may withdraw. Al - Shortest approach. It is across a minefield. Maximum casualties expected. Enemy can withdraw.

A2 - Longest approach. Enemy tackled from the flank. No minefield encountered. Cuts off enemy route of withdrawal. Limited cover available.

Courses Open to the Enemy (States of Nature)


S1 - Enemy upsticks and withdraws before own assault is mounted. S2 - Enemy gives a limited fight and withdraws under pressure. S3 - Enemy fights to last man, last round.

The next step is to construct pay off matrix by assigning pay off values that may accrue by adopting a particular strategy and matching it against a probable state of nature. This necessitates finding a measure for the pay offs, or a measure of effectiveness for the goal to be achieved. In any military operation results may be measured in terms of attrition (ie own and enemy casualties), time taken and spatial gain. In this particular case, the factors considered are:

Enemy casualties. Own casualties. Speed of operations.

Consideration of these factors would lead to a descriptive analysis as shown in Table 10-1 and these can be quantified as given in Table 10-2.

Table 10-1: Evaluation of Strategies The factors can be weighted, if required. Qualification of the descriptive standards has been done on a scale of 0 to 10 to give the values as given in Table 10-2 on next page. Probabilities of occurrence are next assigned to the states of nature and a pay off matrix constructed as shown at Table 10-3. Expected value of pay off is worked out for each alternative, by adding product of probability associated to state of nature and pay off for each strategy for respective alternative. For example, expected value for alternative Al is worked out as; (28 x 0.5) + (22 x 0.3) + (18 x 0.2) = 24.2. Decision Matrix in Tactical Appreciation In the normal tactical appreciation, courses of action available to the enemy are considered and the course most likely to be adopted

Table 10-2: Pay Offs

(Alternative Al, is selected since it has the highest expected value) Table 10-3: Calculation of Expected Value by him arrived at. Thereafter own plans are evaluated only in the light of the enemys most likely course. Thus a situation of risk is

considered as one of certainity by reducing the problem to a single state of nature. The decision matrix permits evaluation of own courses against all states of nature that may possibly occur. In case the states of nature are likely to occur with different probabilities, according to the strategy adopted, the situation can be easily handled as shown in Table 10-4 below.

Note: Figures in the inset boxes indicate probabilities of occurance of states of nature Table 10-4: Decision Matrix for Tactical Appreciation The strategy yielding the highest expected value should be adopted which is Al in this example. Decision Tree In decision matrix, one adopts the best choice based on information and criteria, information being probability associated with states of nature while criteria being expected value. In analysing decision situations, the decision maker, apart from considering all the choices open to him as well as the uncertainties associated with each, has to identify the time sequence in which various actions and consequent events would occur. Decision making may involve several stages and choices at each stage; each of

the choices open will result in a different outcome and pay-off. It Is possible to depict the situation graphically by branches of events. Each branch represents a possible event resulting from a potential choice. The combinations of all the possible actions and potential events are depicted in the form of a decision tree consisting of nodes and branches where each branch represents an alternative course of action and/or an event. Decision trees are basically a combination of decision matrix and probability tree as it represent both the strategies as well as states of nature. The technique of decision tree enables commanders to take more direct account of:

Impact of possible future decisions. Impact of uncertainity. Relative importance of present and future pay-offs.

This technique has the following deficiencies:


Unable to display the logical inter-relationship between the different steps of a complicated situation. Difficult to introduce time relationship especially in multidecisional problem. Difficult to depict sequential decisions as Decision Tree represents a single decision problem at a Given point in time.

Characteristics. Decision tree is a pictograph of the whole situation as it chronologically depicts the sequence of actions and the outcomes as they unfold. It has a tremendous potential as a decision making tool for long range planning and multi-state decision as it can clarify for the commander the choice, the objectives, the monetary gains and the information needs. it mainly consists of nodes and branches. A Square box represents decision/ alternative node or all alternative courses of actions. The branches end either in chance

node or to the final pay-off-value. Chance/Event node representing states of nature or alternative outcomes of the chances event, are represented by a circle as shown here. Decision Node Chance Node

Selection of Best Alternative. All possible decision paths are analysed, and depending on the criteria optimal strategy is selected. Having identified the possible courses of action, the potential states of nature and the outcome for each course of action in a sequence, the decision maker evaluates the expected value of the pay-offs of each branch and its associated probability. The path with highest expected value is identified as the best alternative. In a single stage problem, there will be one decision node and number of branches indicating alternatives and then that many chance nodes from which all possible events will emanate. In a multistage problem, there may be two, three or four sets of decision nodes, the branches of which ending in that many number of chances. Laying the Tree. It implies spreading the problem into decision nodes, chance nodes, alternatives, events and the pay-off values and the probability indicated. Folding Back the Tree. It means calculating expected value at the chance node and annotating the highest value at the decision node. In case there is another chance node, new expected value with the highest expected value of each decision node is calculated. The branch to which the highest expected value refers is left unmarked, while other branches are slashed. The methodology for laying and folding back a tree is explained with help of an example of air attack. Example : Air Attack It is desired to carry out an air attack on an enemy position. For this purpose three alternative are available to the station commander.

Use bombers by day. Use bombers by night Use fighter-bombers by day.

There is a 0.6 probability that the target will be well defended and 0.4 probability that main target will not be well defended. The aircraft on arrival at the target area may decide to attack the main target or take alternative target. The decision tree of the problem is drawn at Figure 10-2 on the foldout sheet. Probability of total/partial destruction and expected pay-offs in each contingency are as shown on the decision tree. These probabilities and pay-offs have been assigned by the station commander. Methodology of Constructing a Decision Tree It will be seen that there are two sequential decisions involved. The decisions required and their associated alternatives and state of nature are as under: Decision 1

q q q

Use bombers by day. Use bombers by night. Use fighter-bomber by day. Probability ... 0.6 ... 0.4

States of Nature Target well defended

Target not well defended

Decision 2 q Alternatives q Attack main target. q Attack alternative target.

States of Nature q Target may be fully destroyed. q Target may be partially destroyed.

Folding back of the tree will be done from right to left Starting from last chance node, calculate expected value of each path and add. Taking the topmost branch, for example; the expected value for last chance node 2 on this branch works out to be (0.4 x 10 + 0.6 x 2) = 5.2. The added value is put on top of each chance node. Then move left to the decision node and place the value highest of the branches on it. Slash the branches of the node other than with the highest value. Similarly, the highest value for each of the decision nodes is annotated. Move left again to the next chance node. Compute the expected value for each branch and annotate the highest one for the node. Slash the other branches. Repeat the process till the first decision node is reached. Mark the highest expected value. The path left inslashed would give the option for maximum expected value. Continuing with the example for the topmost branch; moving left to last decision node 2 the value annotated is 5.2, higher of the two (5.2 and 3.1). Following the unfolding procedure, finally value 7.06 for first decision node l is annotated. The analysis of decision tree tells us to use fighterbomber by day and attack the main target irrespective of the fact whether it is well defended or not. Generally optimal decisions are shown by dark branches/ lines emanating from decision nodes. Conclusion Decision tree is a pictorial presentation of a decision situation where each possible action choice, related risks and possible outcomes are identified easily. It helps to sharpen the thinking process of the decision maker and assist him in focusing on key issues. This structured decision process can help in making better, orderly and effective decisions.

CHAPTER 11 QUEUEING (WAITING LINES) General A queue forms when existing demand exceeds existing capacity of the service facility and customers do not receive service immediately on arrival but have to wait, as service facility is temporarily busy providing service to those arrived earlier. In other words, when the number of units arriving is more than the capacity of the service facility for a particular period, a queue will start forming whose length will increase or decrease depending on the rate of subsequent arrivals. The obvious answer to the reduction of the queue length is to increase the service facilities but then the facilities may also he idle for longer durations. The concept of queueing theory helps us to determine optimum trade-off between the minimum waiting time and maximum utilisation of service facility. The service facility could be a bridge, workshop, machine, service station, dockyard, airport etc. while the respective customers could be vehicles, defective radio sets, vehicles, ships, aircraft and so on. Concept The basic cause for long queue lengths or long waiting times of the customers is because neither the arrivals nor time taken for servicing follow a fixed pattern. A fixed pattern is practical only in fully automated systems and not in most real life systems. In fact, long waiting occurs because of randomness or variability in the arrival pattern of customers as well as in the time required in the

service to customers. In real life, it is extremely difficult to exactly predict the arrival pattern and the pattern of time taken for service. Because of the randomness in the arrival and service patterns, probabilistic models have been developed based on which mathematical theory of queues has been formulated. These models enable us to calculate for a given situation, what kind of a queue would result and how long the customers may have to wait for the service. In order to make use of the queueing concepts and the associated formulae, values of two basic parameters must be available. These are:

Average rate of arrival : l (lamda) Average rate of servicing: m (Mu)

If the average rate of arrival (l) is less than the average rate of service and both are stabilized values, the system will eventually settle down to a steady state. In such a situation, probability of a particular length of queue forming will be the same at any time. If the rates are not constant, the system will not reach a steady state. If the rate of arrival is equal or less than rate of servicing, the system is unstable and has very high probability of a long queue whose length is steadily increasing. Let us explain this concept with some situations:

Situation One Arrivals are at a constant rate of 10 per hour which means that in every six minutes one unit arrives. Servicing is at a constant rate of 12 per hour which means that every arrival takes exactly five minutes to service.

A queue will not form but after one hour it will be realised that service facility was lying idle for 10 minutes.

Situation Two

Arrivals are at a constant rate of 10 per hour.

Servicing at a constant rate is also 10 per hour.

In this case there is a perfect match between the requirements of customers and pattern of service. In other words demands equal service facility, therefore, neither a queue will form nor service facility will be idle.

Situation Three

Arrivals has a constant rate of 10 per hour. Servicing is at a constant rate of 8 per hour which implies that it takes 7.5 minutes to service the customer.

In one hour there will be 10 arrivals and only 8 can be serviced, therefore, there will be two units waiting for servicing after the end of the hour. If the situation continues, after every one hour the length of the queue will go on increasing by 2. In all the above situations, a very basic assumption has been made of constant arrivals and constant servicing and calculation of whether a queue will form or not seems very easy. However, we all know that it is difficult to maintain arrivals at a constant rate and it is also not possible to ensure that time taken for servicing each arrival is exactly the same. Because of this variability in these two parameters, queuing theory tells us that a queue of infinite length will occur in Situations Two and Three as any arrival later than the constant rate will make the service facility idle and this loss of servicing time will never be recouped. Further, if servicing time is more than the constant rate, it will further aggravate the queue situation. Even in Situation One, the queue occurs due to this variability and queueing theory helps us to determine the queue lengths for the situation. Therefore, to apply queuing concepts, it must be ensured that average rate of arrivals is less than average rate of servicing.

Mathematical Formula The formulae associated with a single service station are given:

Utilisation Factor For a single service station having queue discipline of first come first service, one can completely predict the behaviour of the queue as it is totally dependent on the ratio of average arrival rate and average service rate. This ratio is called utilisation factor and is denoted by p (Rho), and mathematically;

A graph indicating the relationship-utilisation factor and mean queue length is shown at Figure 11-1. It will be seen that queues start forming for a utilisation factor of 0.5 and both the queue length and waiting time tend to become infinite for utilisation factor of 1.

Figure 11-1: Exponential Relationship between Utilisation Factor and Mean Queue Length

It will be further seen that for even a slight increase beyond the utilisation factor of 0.75, there is disproportionate increase in the queue length and waiting time. Therefore, if any system is planned to function with p values of more than 0.75, the system will become inefficient thereby adversely affecting the operational efficiency. Let us consider an example to illustrate this. In an opposed river/canal crossing operation, one vehicle takes one minute to cross the bridge, and if it is planned to release one vehicle every minute for arrival at the bridge entry point, it may cause a lot of congestion of vehicles at the bridge during execution stage. As the average bridge crossing time is one minute, it is essential to have a planning figure of approximately 1.2 minutes for arrival of vehicles at the bridge entry point so that value of p is 0.75. This will, of course, reduce number of vehicles which could be crossed in one hour thus affecting the total number of vehicles which could be planned for a given time of availability of bridge before first light. ILLUSTRATIVE EXAMPLE Field Repair Workshop Normal Load. A field repair workshops receives varying number of vehicles each month. The monthly average works out to 80 vehicles. Similarly, the average service capability is 100 vehicles per month. We observe that :

Input Increased. A couple of months before a formations collective training, the number of vehicles reporting to workshop increases, and this trend continues till after a month of the exercise. The average input thus rises to 95 vehicles per month. Let us see its effect on waiting length of vehicles and waiting time.

Observation. It may be noted that an average increase of 15 vehicles a month has shot up the queue length from 3 vehicles to 18 vehicles and the waiting time from 1 day to 6 days.

Increased Output. It is only natural that the user units and the OC field repair workshop would like to see all vehicles on road or under active repairs with no waiting time particularly during the collective training period. OC workshop declares a minor emergency, stops all sports and social activities and orders overtime in the evening. The double shift increases his average output from 100 to 140 vehicles a. month.

Let us see the effect on waiting length and waiting time for vehicles: It means that on an average, there is one to two vehicle in the queue waiting for less than a day. Note: The increase of average output to 140 vehicles per month achieved by overtime could also be achieved by providing additional resources as an alternative to overtime and/or double shift.

Conclusion In this chapter the basic concepts of queueing have been explained in very simple terms but today queue models and formulae for almost all contingencies are available. These models tell us, for example, if there are two bridges, what is an efficient system having two parallel lanes or one feeding lane which further distributes into two lanes, before the entry to bridges. Basic formulae are available for calculating mean queue length for say a twin canal system, number of facilities required for a particular operating system, number of runways required for an airport, number of servers required at ammunition and/or supply points etc. The option selected using queuing concepts will provide cost effective and efficient service.

CHAPTER 12 MONTE CARLO SIMULATION Introduction During World War 11, the work of John von Neumann and Ulam at the Los Almos Project encountered a major problem emanating from the behaviour of neutrons as they passed through various materials. It was dangerous and hazardous to experiment on a trial and error basis, yet it was absolutely necessary to determine the thickness of the appropriate shielding materials. Finding that pure analytical procedures alone would not solve their problem, Neumann and Ulam simulated the environment by using fictitious neutrons and sending them through representations of various materials about which basic data relating to speed, distance travelled, probable absorption rates, etc, were known. The key to the whole process was the use of a device to generate the random and independent selection of the particles, their paths, speeds, absorption rates, and so forth. The device used was a high speed digital computer and millions of iterations of the experiment were conducted. This experiment was given the code name Monte-Carlo, the gambling casino connotation coming naturally, and applications of the process have used the name ever since. Today, when Monte-Carlo is associated with an experiment, some device is used that provides a subset of values from a population in such a way that each population element has the same chance of being selected. Typical devices used to generate Monte-Carlo samples are random number tables and tapes, and digital computer programs written to generate random numbers.

random number tables and tapes, and digital computer programs written to generate random numbers. Directly related to Monte-Carlo is the concept of simulation, and more specifically Monte-Carlo simulation. And directly related to simulation is the process of modelling. Placing these processes and techniques in proper perspective, requires a look at the basic abstraction process known as model building. Models A model is a representation of an object a system, or a process. A descriptive synonym for the word representation in the above definition is abstraction. Whether the model is physical (a lookalike representation) or analogue (where one set of properties is used to represent another set) or symbolic (where the elements of what is being represented and their interrelationships are represented by mathematical equations), abstraction is the key activity involved. In Monte-Carlo Simulation we primarily use mathematical models for exploratory, predictive, or control purposes. The mere formulation of mathematical models does not necessarily ensure that valid predictions about the underlying system can be made using either analytical methods or experimentation. Figure 12-1 illustrates the typical modelling process.

Figure 12-1: A Typical Modelling Process

Simulation enters the scene when any of the implied steps of modelling depicted in the diagram are either impossible or too costly to perform. For example, Block A may have insufficient accumulated data or observations of the system to adequately specify the hypothesis which finally becomes the symbolic model. Block B may be inadequately specified because of the complexity of the observed system. The state of analytic technology, for symbolic manipulation (Block C) may be inadequate to provide solutions to the model, and therefore to provide predictions (Block D) about the future behaviour of the system, even if sufficient information were available from Block E. And finally, Block F, the validation and evaluation of the model, might be too costly to perform, or have too little test data available coming from Block G. The paramount difficulty experienced about in Blocks A, E, F and G is the lack of sufficient data or information. It is because of these problems that Monte-Carlo Simulation offers assistance to the modeler. Simulation can be defined simply as a numerical technique for experimenting with the elements and relationships of mathematical models for predicting most likely outcome of a situation. Steps or Phases in a Simulation Study Proper problem formulation is the first step in simulation (Figure 12-2). This involves translation of the vague idea that the organisation has, into an explicit problem statement by the analyst through continual interactions, questionnaires, and estimations of results. Second step - the most time consuming - is of data collection. Quantitative data are required to describe the real system and for generation of various stochastic variables in the system. Also, it is required for testing and validation of the model. Next step of data analysis leads to deriving the proper generating functions, after determining the distribution that describes the raw data. Model formulation is a difficult step in the process as simulation model is an abstraction of some real phenomenon or system. It should adequately describe the real system at a minimum cost of human and

Figure 12-2: Phases in Simulation Study computer resources. Computer is an absolute necessity of program generation in simulations, therefore, computer language used is a matter of some consequence. Special purpose simulation languages, such as SIMSCRIPT, GPSS, DYNAMO and GASP simplify programming with features like the simulation clock and next-event logic being pre-programmed. Model validation provides confidence in the model by making certain verifications. It involves testing for congruence between simulator design and program; outputs and real-world variables; subjective human validation and finally comparison of output with historical data. Once a simulation model has been implemented and validated, it can be used for its original purpose, experimentation, to gather the information necessary for decision making. Analysis of simulation results, is fairly straightforward drawing of inferences for rational decision-making often using certain statistical techniques.

When to Simulate On formulation of the problem, the analyst must decide whether or not to attempt to solve the problem using discrete digital simulation. Figure 12-3 depicts this decision process. A solution technique other than simulation is often judged to be more appropriate. The important point is, selection of solution methodology must follow problem formulation and not visa-versa. If a classical analytical technique is not available then decision between simulation and making an intuitive, seat-of-the-pant decision be made.

Figure 12-3: When to Simulate Purpose of Simulation The advantages and purpose of simulation could be understood from the following:

Conducting of exercise is a very costly method of testing in terms money, time and resources. Even physical observation of a system becomes too expensive, at times. A number of aspects may be missed out in a large scale exercise and if it so to be investigated, repetition may not be possible.

Mathematical models give us greater insight into the system, are easier to handle and lend themselves to computerisation. However, it is not possible to develop mathematical solutions for all kinds of problems. When dealing with future weapon systems and future combat environment, it is not possible to conduct exercises. At times, systems may be so complex with large number of probabilistic elements which do not lend themselves to easy handling by even mathematical models. Sufficient time is not available to allow the system to operate extensively. Actual operation and observation of the system may be too disruptive.

Limitations of Simulation Simulation is not without limitations and some of them are:

Is not precise, or an optimisation process, or can it yield an exact answer; but merely provides a set of system responses to different operating conditions. It may consume tremendous amount of time and money to develop and run the simulation. Not all situations can be evaluated by simulation. Simulation can generate a way to evaluating solutions but does not generate solutions themselves. SIMULATION PROCESS

q q

Bomber Mission. The process of simulation will now be explained with an example of a bomber mission. Scenario. A bombing mission is sent to bomb a military target, which is rectangular in shape and has the dimensions 75 mtr by

150 mtr. The bombers will drop 10 bombs altogether, all aimed at the geometric centre of the target. It is assumed that the bombing run is made parallel to the longest dimension of the target, that deviation of the impact point from the aiming point is normal with mean zero, and standard deviation 50 mtr in each dimension and that these two deviations are independent random variables. Use Monte Carlo simulation to estimate the expected number of bomb hits, and compare the result with the exact value. Solution. Let, X be the range error (undershoot or overshoot with respect to aiming point), and Y be the line error (left or right of flight line) as shown in Figure 124 here. A hit results considering the target area when: - 75 < X < 75 and - 37.5 < Y < 37.5. If either of these conditions is violated the result would be a miss.

X Range Error ( Overshoot +ve Undershoot -ve) Y Line Error ( Right +ve Left -ve )

Figure 12-4: Bombing Mission - Target Parameters Now if X and Y are measured in terms of their standard deviates say, u and v respectively, and the standard deviation being 50 mtr in both direction, we can say, (X - 0) U= 50 and v = 50 (Y - 0)

By substituting values for X and Y; the corresponding range for u and v for a hit can be found as

- 1.5 < u < 1.5 < and - 0.75 < v < 0.75 With help of a random Normal Number table (like one extracted at Table 12-1), the results of bombing trials can be simulated. A Random Normal Number table, if not readily available, can be constructed with the help of random number table and the

Table 12.1: Random Normal Table in Units of Standard Deviates (Z) Standard Normal Probability Distribution table (similar to Appendix A). The simple procedure is to pick one random number and look for it by putting a decimal before it in the body of normal probability table. Note down the corresponding Z value with its sign. If the table has maximum value upto 0.4990 and the number is 0.5 or greater (random number 5000 or more) then number is reduced by subtracting 0.5 and then looking for the remainder, and a negative sign is assigned to Z value. For example, if random number happens to be 7612, look for 0.2612 (ie 0.7612 - 0.5) in the table. The Z value corresponds to -0.71 and is noted in sequence, to construct the Random Normal Number table. Likewise referring to several random numbers in sequence the table can be constructed.

The simulation trials are carried out by reading standard deviates from Table 12.1 sequentially either in rows or columns in the usual method of referring to random tables. The values, for example, we pick from this table are 0.80, -0.54, 0.42, -0.48, ... and so on. These are copied sequentially under columns u and v. If both u and v lie within their respective ranges of 1.5 and 0.75, then only result is a hit, otherwise a miss. Result column of the table is accordingly completed, and total hits for the trial counted. Results of the first three trials are given at Table 12-2. The trials are carried out by reading the standard deviates from the Table 12-1 sequentially either in rows or columns. In this table, the standard deviates occur randomly within the normal frequency distribution ie values around zero standard deviation occur more frequently and values around 3 standard deviations occur very rarely and normally reducing in between. These values which we get from the table when we read in columns are 0.80, - 0.54, 0.42, - 0.48, ... and so on, in this example. These three trials yield 4.67 as the average number of hits per mission. Many more trials should be conducted before one can have any real confidence in the results. One way of estimating how many trials are necessary is to list the cumulated mean at the end of each trial and to stop the trials when the mean seems to settle down to a stable value. In this example we have: After trial number : 1 2 5 3 4.67 4 5 6 -

Cumulated means as : So, more trials are necessary.

In this problem, unlike most Monte Carlo problems, an exact calculation of the answer is much easier than the Monte Carlo calculation. The probability of a hit with a single bomb is equal to

Table 12-2: Results of Simulation Trials

area under the normal curve between -1.50 to 1.50 multiplied by area between -0.75 to 0.75 = 2 (0.4332) x 2 (0.2734) (from Z Table at Appendix A) = 0.866 x 0.547 = 0.474 The mean number of hits in a mission dropping 10 bombs is just 10 times this, or 4.74. This figure nearly agrees with the result of simulation which is 4.67. Conclusion Monte Carlo Simulation is one of the most powerful and frequently used technique in the discipline of Operations Research. It has been effectively used in many areas such as inventory, allocation, queueing, sequencing, replacement, war gaming, weapon systems analysis and strategic planning. Simulation is most widespread as far as military problems are concerned, and continues to be essential in the field of war gaming in most advanced countries, an area in which we are also now making a headway.

CHAPTER 13 NETWORK FLOWS The network structure of some typical real world problems that deal with physical distribution systems, communication routes, pipelines, airline flight legs and so on, makes them amenable to solution by specific algorithms for Network Flows. Another reason for using network models is their solution efficiency and capability to handle thousands of variables with help of computers in a reasonable time, where at times other approaches are impractical. Before going into a step-by-step procedure or algorithm for solution of network related problems, it would be appropriate to first familiarise with few terms associated with network flows. Figure 13-1 shows a simple network to explain these.

Fig 13-1: An Illustrative Simple Network


Network. A set of nodes linked by branches or arcs associated with flow of some type, is a network. Nodes. In a network, nodes represent beginning or end of an activity (or end points of a line representing distance, time,

capacity etc). These are depicted by circles. A source node is a point of entry in specific types of networks, and sink (or output) node is the final exit point.

Branches. Lines joining nodes and referring to activities, distances, costs, times etc as the case may be are called branches or arcs. Path. A sequence of connected branches such that in the alternation of nodes and branches no node is repeated, is a path in a network. Acyclic and Cyclic Network. A network is said to be acyclic if it contains no loops; otherwise it is cyclic.

Most of the problems related to network flows also called route scheduling or net flows, generally fall under one of the general categories viz shortest path, minimal spanning tree, maximal flows, capacitated network (CNet) and travelling salesman solution. Figure 13-2 depicts these diagrammatically.

Shortest Route on a network depicts the connected branches (representing distance, cost, or time) in a network that

Fig 13-2: Some Typical Network Flow Problems

collectively comprise shortest distance between a source and destination. The path may not necessarily pass through all nodes.

Minimal Spanning Tree is a connected set of all nodes of the network containing no loops and connecting all nodes at minimum total value (distance, cost, time). Maximal Flow Network is a directed network characterised by flow capacity of the branches and determines the maximum flow that can pass from one input node (source) through the network to one output node (sink) during a specified period. Travelling Salesmans Problem is concerned with the minimisation of total travel cost, distance, time, or combination of these. Where all nodes are required to be connected only once in a continuous path and return to the origin. There is no general algorithm available for this, apparently trivial, problem. There are a large number of solving methods but all are enumerative to some degree or another and hence become impractical for large sized problems. This specific type of problem is not being covered here. SHORTEST ROUTE PROBLEM

Ferry Crossing No 543 Engr Regt operates four ferry services daily across Brahamputra river from Dibrugarh to Sikraghat. Due to direction of river flow, strength of currents, presence of intervening islands and minimum water depth required for cruising; the route followed is circuitous and takes well over an hour. The CO of the regiment is concerned with high consumption of diesel and also wants to reduce the ferry crossing time for passengers. Over the period, a number of points on various navigable routes have been identified and travel times, in minutes, for each segment recorded. These are depicted besides the sketch at Figure 13-3.

Figure 13-3: Shortest Route Problem - Ferry Crossing The problem is to locate that route which would take minimum time for the crossing the river each way. The solution for one way crossing from Dibrugarh to Sikragbat is given here. Corollary of the problem with related data could be solved for the return journey in a similar way. Solution. The algorithm to solve shortest route problem is iterative in nature; at each step determining the shortest distance to one node, thus after (n-1) steps the procedure ends for a network with n nodes. The network at Figure 13-4 has been drawn from the information at the sketch. Segment timings which are not directly related to respective distances have been marked near each node

Figure 13-4: Segment Times of Ferry Crossing

corresponding to direction of travel. The iterative algorithm to find a route from Dibrugarh to Sikraghat with shortest time is as under:

Step 1. Construct a master-list, tabulating each node in the first row (refers Table 13-1 Next page). Under each node, in ascending order of their values, list all the branches incident on that node. Thus start node of each branch in a column is same as the one at the top of the column. Omit in listing all those branches which have the source (origin) as their endnodes or the sink (destination) as their start-nodes. Step 2. Mark a star on the source node and assign it a value 0. Locate the branch with the smallest value under the source column and circle it Star the node that matches the end-node of this branch, and assign this node a value equal to that of the branch. Strike off from the master list all other branches that have the newly starred node as their end-node. Step 3. If the newly starred node is the sink, go to step 6, otherwise continue. Step 4. Consider all those starred node columns that have any uncrossed or uncircled branches left in the master list. For each column, add to the assigned value of the node value of the uncircled branch with smallest value in that column. Denote the smallest of these sums as Z, and circle that branch whose value contributed to Z. Star the end-node of this branch and assign it the value Z. Step 5. Delete from the master-list all other branches having this newly starred node as end-node. Go to Step 3. Step 6. Z* is the final value assigned to the sink. Identify the shortest route recursively.

Corresponding steps and iterations in above example are:


Step 1. In the master-list BA, CA, GB, GD, GF and GE are omitted.

Fig 13-1: Master-Sheet at Successive Iterations

q Step 2. Node A (the source) is starred and assigned value 0. AC10 is circled. Node C is starred and assigned value 10. Branches DC and EC are crossed out.

Step 3. Newly starred node is C, an intermediate node, hence go to next step.

q q

Step 4. Columns A and C are considered. Z is lesser of sums from columns A and C, ie between 40 (0 +40) and 20 (10 + 10) respectively. D is starred and assigned value 20.

Step 5. Branch BD is struck off.

q Continuing Step 3 through 4. Z being least of the sums; 40, 30 and 40 for columns A, C and D respectively, equals 30 at this stage. Therefore CE is circled, E is starred and assigned value 30. FE is struck off. Next Z = 35, via EF is assigned to F. Least Z at this stage is obtained via AB and DB, both being equal to 40. AB is arbitrarily selected and circled. Column B is starred and assigned value 40. Next Z = 45 via FG takes us to the sink, G, and the process completes. q Step 6. Z* = 45. Beginning with sink, by including in the path each circled branch whose end node belongs to the path. Starting at G, column F provides connecting branch FG, column E gives next branch EF, C Provides CE and finally column A gives AC. The shortest path for the ferry from Dibrugarh to Sikraghat thus identified is ACEFG = 45 mins, sum of the branches on the path.

Figure 13-5 shows the route marked on the sketch. Reader may work out the return route.

Figure 13-5: Ferry Crossing - Shortest Route Solution

MINIMUM SPANNING TREE A problem under this group involves selection of a set of branches that span (connect) all nodes of a network so that total value of connected branches is to be minimum. The structure resembles a tree graph connecting all nodes and forming no closed loops. A tree with n nodes will contain n-1 branches. Laying of Underground (UG)Cable With ensuing termination of long-term contract between DoT and the Army, 456 Signals Company is to plan laying of underground cable for telephone communication within Rampur Cantonment. On the sketch at Figure 13-6, location of the Signal Centre (Y), civil exchange (Z) and all the required distribution points (DPs) are marked. DPs have been identified taking into consideration requirements of load, location of subscribers, expansion plans, and other factors of feasibility of routings. The routings from DPs are to be in clusters and overground - however not a part of this project. Lines cannot be laid along direct straight lines connecting any two DPs due to existing structures, natural obstructions, and at places for reasons of security, ownership of land and maintenance

Fig 13-6: laying UG Cable - Minimal Spanning Tree Problem

considerations. The costs to interconnect DPs on all possible routes have been worked out taking the facts on ground and all associated costs into account. These costs (in lacs of rupees) have been indicated on the sketch. The problem is to identify all the branches which would interconnect all nodes (DPs) and the exchanges - military and civil - of the network, with the aim of minimising the overall cost. Solution. The graphical algorithm for finding minimal spanning tree is quite straight forward as given here.

Step 1. Start at any node at random and join it to its nearest neighbouring node. Step 2. Choose the unconnected node which is nearest to any of the connected nodes, and join it to that node. Step 3. If there are still any unconnected nodes, repeat Step 2, otherwise stop.

For a small network, solution can be obtained graphically as above. But for a large network the sequence should be followed in a tabular format to avoid any errors of omission or incorrect selection. Tabular Method. A table listing all nodes for columns and rows in the same sequence is made. Body of the table indicates branch values between the nodes. The branches between nodes that cannot be connected, are left blank indicating non-availability. Such a table for UG Cable problem is at Table 13-2 (next page).

Arbitrarily we select any node, say node Y and mark it by # sign. Cross out column Y. Circle the least value in the row ie 8 which is in column F. Next we mark F row with #, cross out column F and find the least value (among 17, 13, 22, 20, 14 and 22) in marked rows (Y and F) and circle it. Thus value 13 in column I is circled.

Table 13-2: Minimal Spanning Tree Tabular It takes us to I row for the third iteration. Crossing out column I, and considering F, Y and I rows the least value 11 from column H is circled. The process is continued.

After tenth iteration we find that no cell is available hence, the process is complete. Each # marked or circled node as well as cross-out lines for columns have been appended with iteration serial numbers within small triangles, to facilitate following the step-by-step procedure by the reader. The resulting tree is traced by highlighting the circled branches on the sketch. Figure 13-7 (next page) shows this for the UG Cable problem. Minimum cost, sum of the highlighted branches, equals Rs 102 lacs. One could start from any other node and confirm the results.

Figure 13-7: Laying UG Cable - Minimal Spanning Tree Solution

MAXIMAL FLOW PROBLEM When the objective is to determine the maximum amount of flow (ie fluid, traffic, information etc) that can be transmitted through a network, Maximal Flow algorithm can be used. It concerns with a single input node (the source) and a single output node (the sink), and is helpful in capacity planning of varied distribution systems. The maximal flow problem is characterised by the branches of the network having finite capacities which limit the amount of flow that can pass through the branch in a given amount of time. The problem is to determine a break-through path, subject to capacity limits, the amount of flow across each branch that will permit the maximal total flow from source to sink. The problem may pertain to networks consisting of directed or undirected branches. Physical interpretation of a directed branch is one-way street in city traffic system with its associated flow capacity. A network of gaspipelines, connected at valves, represents a network of undirected branches - gas may flow in either direction through a pipeline regardless of flow direction.

This special algorithms first guideline is that except for source -and sink, whatever flows into a node must flow out of the node. Second, in a flow across a path from source to sink. the maximum amount that can flow equals the minimum capacity of anybranch on that path. This is similar to the weakest link in the chain analogy. Finally a mechanism should be provided in the search for optimal allocation along various paths for adjusting (or changing) them at a later stage in the search if they are not optimal. A two step mechanism enables this to be achieved. While assigning a flow to a branch, first the capacity in the direction of flow by the amount of the flow is reduced, and second the capacity in the opposite direction of flow increased by the amount of flow. Second step might seem strange, but allows to undo a flow that is not optimal. Maximal Flow Algorithm. Following the guidelines given above the steps in the procedure are:

Step 1. Find a path from source node to sink node with positive flow capacity. If no path with positive flow capacity exists stop. The current flows are optimal. The start can be made from any path from the source node. Step 2. Find the branch in the path that has the minimum flow capacity, Cmin. Increase the flow along the path by Cmin Step 3. Decrease the capacities in the direction of flow by Cmin for all branches along the path. Increase the capacities in the opposite direction by Cmin for all the branches in the path. Go to Step 1.

Different decision makers might make different choices in Step 1. However, properly executed algorithm will provide optimal solution. Transportation of Helicopter Assemblies HQ Eastern Air Command, IAF is faced with a situation where 30 helicopters in completely knocked down (CKD) state, are to be

collected from Mumbai. The consignment is arriving by sea this week-end and should reach Gauhati within a fortnight for assembly there by the foreign suppliers team. The crates can be transported via Calcutta by three different means, viz by-road in special trailer vehicles, by-rail and by-air in service aircraft only. The consignment may be moved via the same mode of transportation for the complete trip, or it may switch modes at Calcutta. There are constraints on all the three modes. Only limited numbers of specialised trailers are available at Mumbai and Calcutta for transportation by road. The transportation by rail is restricted due to limited number of wagons readily available at both places. The air-staff do not wish to launch special sorties for air-lift and would use the routine courier flights from both cities to Gauhati. Complete information on the capacities of three modes available for moving the consignment during next fortnight, is now available with the staff. The information is at Table 13-3. Movements from harbour to airport and railway-yard at Mumbai, and from railway yard to airport at Gauhati can be handled locally.

Table 13-3: Capacities of Various Branches The air-staff wish to determine if the task can be achieved using existing capabilities.

Solution. The data in Table 13-3 have been converted into a network at Figure 13-8. The flow capacities along each branch are indicated next to each node. These may vary according to direction of flow which is also indicated. The branches labelled A, T and R represent air, trailer (road) and rail links respectively. Nodes CA, CT and CR represent the three transhipment points; ie airport, road terminal and railway yard respectively, at Calcutta.

Figure 13-8: Network for Transportation of Helicopter Assembiles

Referring to Figure 13-9 and following the algorithm presented earlier under para titled Maximal Flow Algorithm, we identify path M-CA-G (the air route via Calcutta) as allowing positive flows through each branch on the path. The flow capacity for branch M-CA is 8. The flow capacity for M-CA is compared with C-AG which gives optimum flow = 8. Thus 8 CKD helicopters are allocated along this path. Flow capacities of all nodes are adjusted along the path as indicated in Figure 13-9(i). Arbitrarily, we identify flow potential along path M-CT-G. Comparing M-CT = 20 and CT-G = 10, we allocate only 10 CKD helicopters to this path and re-adjust the flow capacities. Refers Figure 13-9 Q. Further, flow potential is identified along the path M-CT-CA-G (ie by road-trailer to Calcutta and then by air to Gauhati). Optimal

Figure 13-9: Step-by-Step Allocations for Transportation of Helicopter Assemblies

which is 6, is allocated and all flow capacities along the path adjusted as in Figure 13-9 (iii). At this stage it can be seen that the air capacity has been fully used. Continuing further, when all the three branches to Gauhati get fully exhausted, as is the case depicted in Figure 13-9(iv),the solution is obtained. It suggests a total capacity of 30, therefore, the task can be just met with the available capacity of the resources. The schedule according to this solution is; send 8 CKD helicopters by air, 6 by road trailers and 16 by rail to Calcutta. From thereon the distribution could be 14 by air, 10 by road trailers and 6 by rail to Gauhati. Arbitrary selection of paths gives different combinations but finally maximum capacity remains the same. In this case, for example, if air effort from Mumbai is not used, 30 CKD helicopters could still be transported to Calcutta by road-trailers and rail in various possible combinations, and then to Gauhati as suggested in the solution. Problems with large number of nodes solved with help of a computer will bring out various alternatives from which decision maker can select the most suitable one. Conclusion In this chapter three kinds of applications of network modelling were presented. Although special algorithms for each network model have been discussed, most of the special algorithms are subsume of Simplex method. Nevertheless, the specialised algorithms do offer computational advantage. Variations of tabular and graphical network algorithms are also found with their merits. For example, different algorithms can be used in finding shortest-route in acyclic and cyclic networks. The one for the cyclic network being more general includes the acyclic case but is not as efficient because it entails more computation.

Applications of these techniques extend beyond the problems dealing with distances, costs and time of transportation, flow or capacity. For example, shortest-route algorithm may be applied to decide on equipment replacement policy or viewed as a transhipment model - a special case of transportation model. Considering the branches as probabilities, the most reliable route can be determined. A variation of minimum spanning tree emerges when it is needed that all nodes are connected to one central node. If the branches have capacity limitations imposed, the problem becomes a capacitated minimum spanning tree much more difficult to solve. However heuristic and optimisation procedures do exist to solve it. Large and complex problems of networks as mentioned can be handled by various computer software packages.

APPENDIX A Areas under the Standard Normal Probability Distribution between the Mean and Successive Value of z.

Example: To find the area under the curve between the mean and a point 2.24 standard deviations to the right of the mean look up the value opposite 2.2 and below 0.04 in the table. 0.4875 of the area under the curve lies between the mean and a Z value of 2.24. z 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.O 2.1 .00 .0000 .0398 0793 .1179 .1554 .1915 .2257 .2580 .2881 .3159 .3413 .3643 .3849 .4032 .4192 .4332 .4452 .4554 .4641 .4713 .4772 .4821 .01 .0040 .0438 .0832 .1217 .1591 .1950 .2291 .2611 .2910 .3186 .3438 .3665 .3869 .4049 .4207 .4345 .4463 .4564 .4649 .4719 .4778 .4826 .4864 .3896 .4920 .4940 .4955 .4966 .4975 .4982 .4987 .02 .0080 .0478 .0871 .1255 .1628 .1985 .2324 .2642 .2939 .3212 .3461 .3686 .3888 .4066 .4222 .4357 .4474 .4573 .4656 .4726 .4783 .4830 .4868 .3898 .4922 .4941 .4956 .4967 .4976 .4983 .4987 .03 .0120 .0517 .0910 .1293 .1664 .2019 .2357 .2673 .2967 .3238 .3485 .3708 .3907 .4082 .4236 .4370 .4484 .4582 .4664 .4732 .4788 .4834 .4875 .4901 .4925 .4943 .4957 .4968 .4977 .4983 .4988 .04 .0160 .0.577 .0948 .1331 1700 .2054 .2389 .2704 .2995 .3264 .3508 .3729 .3925 .4099 .4251 .4382 .4495 .4591 .4671 .4738 .4793 .4838 .4871 .4904 .4927 .4945 .4959 .4969 .497 7 .4984 .4988 .05 .0199 .0596 .0987 .1368 .1736 .2088 .2454 .2734 .3023 .3289 .3531 .3749 .3944 .4115 .4265 .4394 .4505 .4599 .4678 .4744 .4798 .4842 .06 .0239 .0636 .1026 .1406 .1772 .2123 .2486 .2764 .3051 .3315 .3554 .3770 .3962 .4131 .4279 .4406 .4515 .4608 .4686 .4750 .4803 .4846 .07 .0279 .0675 .1064 .1443 .1808 .2157 .2486 .2794 .3078 .3340 .3577 .3790 .3980 .4147 .4292 .4418 .4525 .4616 .4693 .4756 .4808 .4850 .4884 .4911 .4932 .4949 .4962 .4972 .4979 .4985 .4989 .08 .0319 .0714 .1103 .1480 .1844 .2190 .2517 .2823 .3106 .3365 .3599 .3810 .3997 .4162 .4306 .4429 .4535 .4625 .4699 .4761 .4812 .4854 .4887 .4913 .4934 .4951 .4963 .4973 .4980 .4986 .4990 .09 .0359 .0753 .1141 .1517 .1879 .2224 .2549 .2852 .3133 .3389 .3621 .3830 .4015 .4177 .4319 .4441 .4545 .4633 .4706 .4767 .4817 .4857 .4890 .4916 .4936 .4952 .4946 .4974 .4981 .4986 .4990

2.2 .4861 2.3 .4893 2.4 .4918 2.5 .4938 2.6 .4953 2.7 .4965 2.8 .4974 2.9 .4981 3.0 .4987

.4878 .4881 .4906 .4909 .4929 .4931 .4946 .4948 .4960 .4961 .4970 .4971 .4978 .4979 .4984 .4985 .4989 .4989

From Robert D Mason, Essentials of Statistics, Prentice Hall,Inc.1976.

BIBLIOGRAPHY STATISTICS 1. Richard I. Levin. Statistics for Management, Prentice Hall of India. 2. Simpson and Kafka. Basic Statistics, Oxford & IBH. 3. Kazmier. Statistical Analysis for Business and Economics, Mc Graw Hill. 4. Moroney. Facts from Figure, Pelican Books. OPERATIONS RESEARCH 5. Levin and Kirkpatrick. Quantative Approaches to Management, Mc Graw Hill. 6. Loomba. Managment - A quantitative perspective, Collier MacMillan. 7. Sasiehi, Yaspan, Friedman. Operations Research Methods & Problems, Wiley International. 8. Wanger. Principles of Operations Research, Prentice Hall. 9. Vohra ND. Quantitative Techniques in Management, Tata Mc Graw-Hill, New Delhi. 10. Budnick Fronk S, Mcleavey & Mojena R. Principles of Operations Research for Management; Richard D Irwin Inc, Illinois. 11. Cook Thomas M & Russel RA Introduction to Management Science. Prentice-Hall Inc, NJ.