Anda di halaman 1dari 15

Is B&B an efficient algorithm?

• NO
• For larger problems the tree size grows exponentially.
• Need to solve an LP at each node of the tree.
• Possible remedy: Early termination.
• For maximization problem
– ZINC is an lower bound.
– Max(ZLP) over all dangling nodes is an upper bound.
– Optimality Gap : % difference between the two.
– Stop when the gap is less than 1-2%.
– Heuristic Solution.
• Quite often, the optimal solution is reached quite early in the search, but it takes a long
time to prove optimality.
Problem 5 (20)

This question should be answered directly on this sheet.

When solving an integer programming problem with Maximization objective function, the branch and
bound tree at present has 7 dangling nodes. The LP objective function values at these nodes are 378,
418, 288, 313, 407, 383 and 356.

a) Indicate True/False/Maybe for each of the following statements. (12)

Statement True/False/Maybe
1. Optimal solution value for this
problem is 431
2. Optimal solution value for this
problem is more than 418
3. Optimal solution value for this
problem is less than 376
4. Current Incumbent value more than
295
5. Current Incumbent value is less than
288
6. Current Incumbent value is less than
270
a) If the node with best upper bound value is selected for further branching, and
two new nodes are created, the LP objective function values at these nodes are
376 and 349, respectively. Moreover, the solution with value 376 is integral.

What will be the number of dangling nodes in the tree after updating the incumbent
and removing the fathomed nodes? List the values of LP solutions at these nodes.

Calculate the maximum percentage gap between the optimum solution and the
current incumbent.

If the search is to continue, which node should be selected for further branching.
Indicate the LP solution value of that node.
Travelling Salesman Problem
• Given a set of cities and distances among them, find a minimum distance tour
that visits each city exactly once.
• Applications
– Routing of trucks
– Circuit board drilling
– Machine Scheduling (when setup time depends on previous as well as next job)
• Decision Variables?
• Constraints?
• Objective Function?
• xij = 1 if salesman travels from city i to city j, =0 otherwise
• n(n-1) variables
Minimize i j cijxij
subject to: j xij = 1 i (I)
j xji = 1 i (II)
iSjS xij  1  SN (III) sub-tour breaking cons.
xij = 0,1
• Relax (III), yields an assignment problem.
• Easy to solve.
• Provides a lower bound.
• Solution is integral.
• What is violated? Sub-tour breaking constraints.
• How to fix that? Can it be fixed via branching?
• Consider a sub-tour. 1-4-5-1
• Sub-tour 1-4-5-1 has three arcs: 1-4, 4-5 and 5-1
• Can all three arcs be present in a valid tour? X14=0 X51=0
• NO! at least one must be absent.
X45=0
• Use this for branching
– subdivide the set of all tours into three subsets.
– One of three arcs is excluded in each subset.
Solve 5-node TSP with following distance matrix

999 31 90 48 24
31 999 66 26 35
90 59 999 32 67
48 26 32 999 25
24 35 67 25 999
Root Node Total Cost = 154
999 31 90 48 24
31 999 66 26 35
90 59 999 32 67
48 26 32 999 25
24 35 67 25 999
X(3,4) = 0 Total Cost = 165
999 31 90 48 24
31 999 66 26 35
90 59 999 999 67
48 26 32 999 25
24 35 67 25 999
X(4,3) = 0 Total Cost = 180
999 31 90 48 24
31 999 66 26 35
90 59 999 32 67
48 26 999 999 25
24 35 67 25 999
X(3,4) = 0, X(1,5)=0 Total Cost = 180
999 31 90 48 999
31 999 66 26 35
90 59 999 999 67
48 26 32 999 25
24 35 67 25 999
X(3,4) = 0, X(5,1)=0 Total Cost = 171
999 31 90 48 24
31 999 66 26 35
90 59 999 999 67
48 26 32 999 25
999 35 67 25 999
Role of Heuristics
• Heuristic Algorithm: does not guarantee optimality
• Important for solving real world problems because B&B will take
too long.
• Why is a heuristic algorithm important in the context of branch
and bound?
• Provides a good initial incumbent solution
• Quicker fathoming, fewer nodes, faster convergence.
100
Zinc = Infinity

Zinc = 111 105 107

112 113

113 115

110
114 116
• Q.: A branch and bound tree, at some stage of computation, has 5 dangling nodes. The LP
solution values at these nodes are: 1) 98.5, 2) 77.3, 3) 87.6, 4) 86.3, 5) 103.4. The current
incumbent solution value is 104.5.

• Is this a minimization problem or a maximization problem? Explain why?

• Which node will you select for further branching, and why?

• Suppose that the LP solution values for the two nodes created by branching on the selected
node are 88.2 and 91.5. Furthermore, both these solutions are integral solutions. How many
dangling nodes will the tree have in the next iteration. Explain your answer.

• The user has decided that branch and bound search will be terminated if the current
incumbent solution is guaranteed to be within 2.5% of the optimal solution. In light of this
criterion, should the search be continued or stopped? Explain.

Anda mungkin juga menyukai