Branch & Bound Algorithms

Briana B. Morrison With thanks to Dr. Hung

Topics

Define branch & bound 0-1 Knapsack problem

Breadth-First Search Best-First Search

Assignment Problem Traveling Salesperson

Introduction

The branch-and-bound design strategy is very similar to backtracking in that a state space tree is used to solve a problem. The differences are that the branch-and-bound method 1) does not limit us to any particular way of traversing the tree, and 2) is used only for optimization problems. A branch-and-bound algorithm computes a number (bound) at a node to determine whether the node is promising.

Introduction

The number is a bound on the value of the solution that could be obtained by expanding beyond the node. If that bound is no better than the value of the best solution found so far, the node is nonpromising. Otherwise, it is promising. The backtracking algorithm for the 0-1 Knapsack problem is actually a branch-and-bound algorithm. A backtracking algorithm, however, does not exploit the real advantage of using branch-and-bound.
4

Introduction

Besides using the bound to determine whether a node is promising, we can compare the bounds of promising nodes and visit the children of the one with the best bound. This approach is called best-first search with branch-and-bound pruning. The implementation of this approach is a modification of the breadth-first search with branch-and-bound pruning.

Branch and Bound

An enhancement of backtracking Applicable to optimization problems Uses a lower bound for the value of the objective function for each node (partial solution) so as to: guide the search through state-space rule out certain branches as unpromising
6

0-1 Knapsack

To learn about branch-and-bound, first we look at breadth-first search using the knapsack problem Then we will improve it by using best-first search. Remember the default strategy for the 0-1 knapsack problem was to use a depth-first strategy, not expanding nodes that were not an improvement on the previously found solution.
7

BackTracking (depth-first)
8 17
10

17

20

12

13

14

11
8

Breadth-first Search

We can implement this search using a queue. All child nodes are placed in the queue for later processing if they are promising. Calculate an integer value for each node that represents the maximum possible profit if we pick that node. If the maximum possible profit is not greater than the best total so far, dont expand the branch.
9

Breadth-first Search

The breadth-first search strategy has no advantage over a depth-first search (backtracking). However, we can improve our search by using our bound to do more than just determine whether a node is promising.

10

Branch and Bound (breadth-first)

8 17
10

17

20

12

13

14

11
11

0-1 Knapsack

0-1 Knapsack using the branch and bound. Now look at all promising, unexpanded nodes and expand beyond the one with the best bound. We often arrive at an optimal solution more quickly than if we simply proceeded blindly in a predetermined order.

12

Best-first Search

Best-first search expands the node with the best bounds next. How would you implement a best-first search?

Depth-first is a stack Breadth-first is a queue Best-first is a ???

13

Branch and Bound (best first)

8 17
10

17

20

12

13

14

11
14

0-1 Knapsack

Capacity W is 10 Upper bound is $100 (use fractional value)

Item Weight Value

1 4 $40

Value / weight
10

2
3 4

7
5 3

$42
$25 $12

6
5 4
15

Computing Upper Bound

To compute the upper bound, use

ub = v + (W w)(vi+1/wi+1)

So the maximum upper bound is

pick no items, take maximum profit item ub = (10 0)*($10) = $100

all of item 1 (4, $40) + partial of item 2 (7, $42) $40 + (10-4)*6 = $76 ub = (10 0) * ($6) = $60
16

After we pick item 1, we calculate the upper bound as

If we dont pick item 1:

State Space Tree

0 1
w = 4, v = 40 ub = 76

with 1

w = 0, v = 0 ub = 100

without 1 2
w = 0, v = 0 ub = 60

with 2
w = 11

without 2
w = 4, v = 40 ub = 70

4 without 3
w = 4, v = 40 ub = 64

3 not feasible 5 with 4 7

w = 12

with 3
w = 9, v = 65 ub = 69

inferior to node 8

6
inferior to node 8 17

without 4
w = 9, v = 65 ub = 65

not feasible

optimal solution

Bounding

A bound on a node is a guarantee that any solution obtained from expanding the node will be:

Greater than some number (lower bound) Or less than some number (upper bound)

If we are looking for a maximal optimal (knapsack), then we need an upper bound

For example, if the best solution we have found so far has a profit of 12 and the upper bound on a node is 10 then there is no point in expanding the node

The node cannot lead to anything better than a 10 18

Bounding

Recall that we could either perform a depth-first or a breadth-first search

Without bounding, it didnt matter which one we used because we had to expand the entire tree to find the optimal solution Does it matter with bounding?

Hint: think about when you can prune via bounding

19

Bounding

We prune (via bounding) when:

(currentBestSolutionCost >= nodeBound)

This tells us that we get more pruning if:

The currentBestSolution is high And the nodeBound is low

So we want to find a high solution quickly and we want the highest possible upper bound

One has to factor in the extra computation cost of computing higher upper bounds vs. the expected pruning savings
20

Enumeration in a search tree

root node Level 0

child nodes

Level 1

child nodes

...
...

Level 2

each node is a partial solution, i.e. a part of the solution space

21

The assignment problem

We want to assign n people to n jobs so that the total cost of the assignment is as small as possible (lower bound)

22

Example: The assignment problem

Select one element in each row of the cost matrix C so that: no two selected elements are in the same column; and the sum is minimized For example: Job 1 Job 2 Job 3 Job 4 Person a 9 2 7 8 Person b 6 4 3 7 Person c 5 8 1 8 Person d 7 6 9 4 Lower bound: Any solution to this problem will have total cost of at least: sum of the smallest element in each row = 23
10

Assignment problem: lower bounds

24

State-space levels 0, 1, 2

25

Complete state-space

26

Traveling Salesman Problem

Can apply branch & bound if we come up with a reasonable lower bound on tour lengths. Simple lower bound = finding smallest element in the intercity distance matrix D and multiplying it by number of cities n Alternate solution:

For each city i, find the sum si of the distances from city i to the two nearest cities; compute the sum s of these n numbers; divide the result by 2 and if all the distances are integers, round up the result to the nearest integer;
27

Traveling salesman example:

lb = [(1+3)+(3+6)+(1+2)+(3+4)+(2+3)]/2 = 14

28

Summary: Branch and bound

Feasible solution Optimal solution Breadth-First Search Best-First Search (with branch-and-bound pruning)

29

Summary: Branch and bound

Branch and Bound is:

a general search method. minimize a function f(x), where x is restricted to some feasible region. a means of computing a lower bound on an instance of the optimization problem. a means of dividing the feasible region of a problem to create smaller subproblems. a way to compute an upper bound (feasible solution) for at least some instances. 30

To apply branch and bound, one must have:

Summary: Branch and bound

The method starts by:

considering the original problem with the complete feasible region (called the root problem). The lower-bounding and upper-bounding procedures are applied to the root problem. If the bounds match, then an optimal solution has been found. Otherwise, the feasible region is divided into two or more regions, which together cover the whole feasible region. These subproblems become children of the root search node. The algorithm is applied recursively to the subproblems, generating a tree of subproblems. 31

Summary: Branch and bound

If an optimal solution is found to a subproblem,

it is a feasible solution to the full problem, but not necessarily globally optimal. Since it is feasible, it can be used to prune the rest of the tree:

if the lower bound for a node exceeds the best known feasible solution, no globally optimal solution can exist in the subspace of the feasible region represented by the node. Therefore, the node can be removed from consideration.

The search proceeds until all nodes have been solved or pruned, or until some specified threshold is meet between 32 the best solution found and the lower bounds on all unsolved subproblems.