Unit - V: Branch and Bound - The Method
Unit - V: Branch and Bound - The Method
Unit - V: Branch and Bound - The Method
UNIT V
BRANCH AND BOUND -- THE METHOD
The design technique known as branch and bound is very similar to backtracking (seen in unit 4) in that it searches a tree model of the solution space and is applicable to a wide variety of discrete combinatorial problems. ach node in the combinatorial tree generated in the last !nit defines a problem state. All paths from the root to other nodes define the state space of the problem. Solution states are those problem states 's' for which the path from the root to 's' defines a tuple in the solution space. The leaf nodes in the combinatorial tree are the solution states. Answer states are those solution states 's' for which the path from the root to 's' defines a tuple that is a member of the set of solutions (i.e."it satisfies the implicit constraints) of the problem. The tree organi#ation of the solution space is referred to as the state space tree. A node which has been generated and all of whose children have not yet been generated is called a live node. The live node whose children are currently being generated is called the E$node (node being e%panded). A dead node is a generated node" which is not to be e%panded further or all of whose children have been generated. Bounding functions are used to kill live nodes without generating all their children. Depth first node generation with bounding function is called backtracking. &tate generation methods in which the E-node remains the E$node until it is dead lead to branch-and-bound method. The term branch$and$bound refers to all state space search methods in which all children of the $node are generated before any other live node can become the $node. 'n branch$and$bound terminology breadth first search(()&)$ like state space search will be called )')* ()irst 'n )irst *utput) search as the list of live nodes is a first $in$first $out list(or queue). A D-search (depth search) state space search will be called +')* (+ast 'n )irst *ut) search" as the list of live nodes is a list$in$first$out list (or stack).
(ounding functions are used to help avoid the generation of sub trees that do not contain an answer node. The branch$and$bound algorithms search a tree model of the solution space to get the solution. -owever" this type of algorithms is oriented more toward optimi#ation. An algorithm of this type specifies a real $valued cost function for each of the nodes that appear in the search tree. !sually" the goal here is to find a configuration for which the cost function is minimi#ed. The branch$and$bound algorithms are rarely simple. They tend to be quite complicated in many cases. %ample ..,/4$queens0 +et us see how a )')* branch$and$bound algorithm would search the state space tree (figure 1.2) for the 4$queens problem.
'nitially" there is only one live node" node,. This represents the case in which no queen has been placed on the chessboard. This node becomes the $node. 't is e%panded and its children" nodes2" ,." 34 and 45 are generated. These nodes represent a chessboard with queen, in row ,and columns ," 2" 3" and 4 respectively. The only live nodes 2" ,." 34" and 45.'f the nodes are generated in this order" then the ne%t $node are node 2. 2
't is e%panded and the nodes 3" ." and ,3 are generated. 6ode 3 is immediately killed using the bounding function. 6odes . and ,3 are added to the queue of live nodes. 6ode ,. becomes the ne%t E-node. 6odes ,7" 24" and 27 are generated. 6odes ,7 and 24 are killed as a result of the bounding functions. 6ode 27 is added to the queue of live nodes. 6ow the E-node is node 34.)igure .., shows the portion of the tree of )igure 1.2 that is generated by a )')* branch$and$bound search. 6odes that are killed as a result of the bounding functions are a 8(8 under them. 6umbers inside the nodes correspond to the numbers in )igure 1.2.6umbers outside the nodes give the order in which the nodes are generated by )')* branch$and$bound. At the time the answer node" node 3," is reached" the only live nodes remaining are nodes 3. and 44.
Least Cost (LC) Search: 'n both +')* and )')* branch$and$bound the selection rule for the ne%t $node is rather rigid and in a sense blind. The selection rule for the ne%t $node does not give any preference to a node that has a very good chance of getting the search to an answer node quickly. Thus" in %ample ..," when node 35 is generated" it should have become obvious to the search algorithm that this node will lead to answer node in one move. -owever" the rigid )')* rule first requires the e%pansion of all live nodes generated before node 35 was e%panded. 3
The search for an answer node can often be speeded by using an 8intelligent8 ranking function (.) for live nodes. The ne%t E$node is selected on the basis of this ranking function. 'f in the 4$queens e%ample we use a ranking function that assigns node 35 a better rank than all other live nodes" then node 35 will become E-node" following node 27.The remaining live nodes will never become E$nodes as the e%pansion of node 35 results in the generation of an answer node (node 3,). The ideal way to assign ranks would be on the basis of the additional computational effort (or cost) needed to reach an answer node from the live node. )or any node x, this cost could be (,) The number of nodes on the sub$tree x that need to be generated before any answer node is generated or" more simply" (2) The number of levels the nearest answer node (in the sub$tree x) is from x !sing cost measure (2)" the cost of the root of the tree of )igure .., is 4 (node 3, is four levels from node ,).The costs of nodes ,. and 34"27 and 34"and 35 and 3. are respectively 3" 2" and ,.The costs of all remaining nodes on levels 2" 3" and 4 are respectively greater than 3" 2" and ,. !sing these costs as a basis to select the ne%t $node" the $nodes are nodes ," ,." 27" and 35 (in that order).The only other nodes to get generated are nodes 2" 34" 45" ,7" 24" 32" and 3,. The difficulty of using the ideal cost function is that computing the cost of a node usually involves a search of the sub$tree % for an answer node. -ence" by the time the cost of a node is determined" that sub$tree has been searched and there is no need to e%plore % again. )or this reason" search algorithms usually rank nodes only based on an estimate (.) of their cost. +et (%) be an estimate of the additional effort needed to reach an answer node from %. node % is assigned a rank using a function (.) such that (%) 9f (h(%)) : (%)" where h(%) is the cost of reaching % from the root and f(.) is any non$decreasing function. A search strategy that uses a cost function (%) 9f (h(%)) : (%)" to select the ne%t e$ node would always choose for its ne%t e$node a live node with least (.).-ence" such a strategy is called an +;$search (least cost search). ;ost function c (.) is defined as" if % is an answer node" then c(%) is the cost (level" computational difficulty" etc.) of reaching % from the root of the state space tree. 'f % is not an answer node" then c(%) 9infinity" providing the sub$tree % contains no answer node< otherwise c(%) is equals the cost of a minimum cost answer node in the sub$tree %. 4
't should be easy to see that (.) with f (h(%)) 9h(%) is an appro%imation to c (.). )rom now on (%) is referred to as the cost of %. Bo !"#!$: A branch $and$bound searches the state space tree using any search mechanism in which all the children of the $node are generated before another node becomes the $ node. =e assume that each answer node % has a cost c(%) associated with it and that a minimum$cost answer node is to be found. Three common search strategies are )')*" +')*" and +;. A cost function (.) such that (%) >9c(%) is used to provide lower bounds on solutions obtainable from any node %. 'f upper is an upper bound on the cost of a minimum$cost solution" then all live nodes % with (%)?upper may be killed as all answer nodes reachable from % have cost c(%)?9 (%)?upper. The starting value for upper can be set to infinity. ;learly" so long as the initial value for upper is no less than the cost of a minimum$ cost answer node" the above rule to kill live nodes will not result in the killing of a live node that can reach a minimum$cost answer node . ach time a new answer is found "the value of upper can be updated. As an e%ample optimi#ation problem" consider the problem of @ob scheduling with deadlines. =e generali#e this problem to allow @obs with different processing times. =e are given n @obs and one processor. ach @ob i has associated with it a three tuple ( ).@ob i requires units of processing time .if its processing is not completed by the deadline " and then a penalty is incurred. The ob@ective is to select a subset @ of the n @obs such that all @obs in @ can be completed by their deadlines. -ence" a penalty can be incurred only on those @obs not in @. The subset @ should be such that the penalty incurred is minimum among all possible subsets @. such a @ is optimal. ;onsider the following instancesA n94"( " " )9(4",",)"( " " )9(,5"3"2)"(
" " )9(B"2",)"and( " " )9(3",",).The solution space for this instances consists of all possible subsets of the @ob inde% setC,"2"3"4D. The solution space can be organi#ed into a tree by means of either of the two formulations used for the sum of subsets problem.
%#$ re &.'
)igure ..B corresponds to the variable tuple si#e formulations while figure ..1 corresponds to the fi%ed tuple si#e formulation. 'n both figures square nodes represent infeasible subsets. 'n figure ..B all non$square nodes are answer nodes. 6ode 7 represents an optimal solution and is the only minimum$cost answer node .)or this node @9 C2"3D and the penalty (cost) is .. 'n figure ..1 only non$square leaf nodes are answer nodes. 6ode 24 represents the optimal solution and is also a minimum$cost answer node. This node corresponds to @9C2"3D and a penalty of .. The costs of the answer nodes of figure ..1 are given below the nodes.
row
;ompare all effective cost and pick up the largest l. 'f two or more have same cost then arbitrarily choose any one among them. Delete (i" @) means delete ith row and @th column change (@" i) value to infinity. (!sed to avoid infinite loop formation) 'f (i"@) not present" leave it. Gepeat step , to step 7 until the resultant cost matri% having order of 2H2 and reduce it. ((oth G.G and ;.;) !se preserved cost matri% ;n" ;n$,I ;,
;hoose an edge /i" @0 having value 95" at the first time for a preserved matri% and leave that matri%. STEP 1*: EXAMP E! 22 21 4 24 . , 45 ,5 B 3, 45 35 4 1 B 24 3 ,1 ,4 !se result obtained in &tep ,, to generate a complete tour. +,-e. /rap0 + 4 24 ,7 2
, 7
MATRI+: , 4 2 24 3 45 ,1 24 1 4 3, 35 B ,5 21 24 , B 4
,7 ,4 7 45
22 . , 2 3 4 4
;, /G*= G D!;T'*6A , 5 2 5 3 ,4 ,2 ,. 5 4 B 24 4 3 2 25 5 5 4
,. ,4 3 44
,4 , , 2 3 4 4 STE) -: , 5 ,. 3 ,4 ,
;, /;olumn Geduction0 2 5 ,4 44 , 3 ,4 ,2 ,. 5 4 3 2 4 2 3 2 25 5 5 4
,. ,4 3 44
,4 , , 2 3 4 4 STE) /: +9
STE) ': ;alculation of effective cost / .;0 (,"2) 9 2:, 93 (2",) 9 ,2:3 9 ,4 (3"4) 9 2:5 92 (4"4) 9 3:5 9 3 (4"3) 9 5:,2 9 ,2 (4"4) 9 5:2 9 2 STE) &:
,5
+ having edge (2",) is the largest. STE) 0: Delete (2",) from ;, and make change in it as (,"2) if e%ists. 6ow ;ost Jatri% 9 2 3 ,4 44 , , 3 4 4 &T E ,5A The ;ost matri% 2 % 2. Therefore" go to step ,. E-A& ''A STE),A ;2(G" G) 2 3 ,3 4 , 2 5 4 5 5 5 , ,4 ,. 5 4 3 2 5 2 5 5 4
,4 44 ,. , 3 4 4 &T E 3A ;2 (;" G) 5
,,
2 ,3 43 5 3 4 4
3 ,3 ,. 5
4 , 2 5
4 5 5 5 ,
&T E 4A Ereserve the above in ;2 ;2 9 2 ,3 43 5 3 4 4 &T E BA +9 C(,"4)" (3"4)" (4"4)" (4"2)" (4"3)" (4"4)D &T E 1A calculation of .;. (,"4) 9 ,:5 9, (3"4) 9 2:5 92 3 ,3 ,. 5 4 , 2 5 4 5 5 5 ,
,2
(4"4) 9 ,.:5 9,. (4"2) 9 5:,3 9,3 (4"3) 9 5:,3 9,3 (4"4) 9 5:, 9, &T E .A + having an edge (4"4) is the largest. &T E 7A Delete (4"4) from ;2 and make change in it as (4"4) 9 if e%ists. 6ow" cost matri% 2 ,3 5 3 4 &T E ,5A T- cost matri% E-A& '''A &T E ,A ;3 (G" G) 2 ,, 5 3 4 &T E 3A ;3 (;" G) 3 ,2 5 4 5 5 , 3 ,3 5 4 , 2 ,
,3
2 ,, 5 3 4
3 ,2 5
4 5 5 ,
&T E 4A preserve the above in ;2 &T E BA +9C(,"4)" (3"4)" (4"2)" (4"3)D &T E 1A calculation of .; (,"4)9,2:59,2 (3"4)9,,:59,, (4"2)95:,,9,, (4"3)95:,29,2 &T E .A -ere we are having two edges (,"4) and (4"3) with cost 9 ,2. -ence arbitrarily choose (,"4) &T E 7A Delete (i"@) (,"4) and make change in it (4",) 9 if e%ists. 6ow cost matri% is 2 , , 5 2 3 &T E ,5A =e have got 2%2 matri% ;4 (GG)9 2 5 5 3 4 ;4 (;" G) 9 2 3 ,4 3 5 3 5
5 5 3 4 Therefore";4 9 2 5 5 3 4
3 5
,3 43 5 3 4 4 ;, 9 ,
,3 ,. 5
, 2 5
5 5 5
4 ,4
5 ,. 3 ,4 2 3 4 4 &T E ,2A i) 2 5 5 3 4
5 ,4 44 ,
,4 ,2 ,. 5
3 22 2 5
2 25 5 5
!se ;4 9 3 5
Eick up an edge ('" @) 95 having least inde% -ere (3"2) 95 -ence" T (3"2) !se ;3 9 2 ,, 5 3 4 Eick up an edge (i" @) 95 having least inde% -ere (,"4) 95 -ence" T(3"2)" (,"4) !se ;29 2 3 4 4 ,B 3 ,2 5 4 5 5 ,
,3
, 2 5
5 5 5
,3 43 ,. 5 3 4 4 5
Eick up an edge (i" @) with least cost inde%. -ere (,"4) not possible because already chosen inde% i (i9@) (3"4) not possible as already chosen inde%. (4"4) 5 -ence" T (3"2)" (,"4)" (4"4) !se ;, 9 , 5 ,. 3 ,4 2 3 4 4 Eick up an edge (i" @) with least inde% (,"2) 6ot possible (2",) ;hoose it - 6; T (3"2)" (,"4)" (4"4)" (2",) SOLUTION: ,1 2 5 ,4 44 , 3 ,4 ,2 ,. 5 4 3 22 2 5 4 2 25 5 5 ,
)rom the above list 3K2K,K4K4 This result now" we have to return to the same city where we started (-ere 3). %#!a1 res 1tA 3K2K,K4K4K3 Cost #s ,.2,.2-,2/2'3/4
,.