Approximation Algorithms for 2-Stage Stochastic-Weighted Matching Problem INEN 689: Large-Scale Stochastic Optimization Dec 6, 2005 Balabhaskar Balasundaram Svyatoslav Trukhanov Graphs & Matching G = (V,E), |V| = n, |E| = m Matching is a subset of edges such that no two of them are incident at the same vertex. Graphs & Matching G = (V,E), |V| = n, |E| = m Matching is a subset of edges such that no two of them are incident at the same vertex. 1 1 3 2 5 4 1 3 2 5 4 3 2 5 4 Maximum Weighted Matching G = (V,E) with non-negative edge weights ce on every edge e in E Find a matching M such that the sum of the weights of edges in the matching is a maximum 1 1 5 3 1 2 3 5 4 5 3 4 3 2 2 3 1 3 2 5 4 4 0,1-Formulation Max∑ ce xe e∈E subject to : x ∑ δ e∈ ( v ) e ≤ 1, ∀v ∈ V xe ∈ {0,1}, ∀e ∈ E Applications Marriage Problem (Bipartite Matching) Roommate Problem (General Graph Matching) Employee-task assignment Processor-job assignment Several other time-tabling and assignment type problems 2-Stage Stochastic Extension (Kong & Schaefer, 2004) First stage weights are known (ce) Second stage weights are (des) with probability ps for each scenario s=1,…,r First Stage: Find a matching Second Stage: Augment the matching Maximize sum of first stage weights and expected sum of second stage weights Interpretation First stage decision presents you with the first stage profit on Day 1 Tomorrow is uncertain ! We wish to maximize tomorrow’s expected profit Eg. Marriage problem…compatibilities change after marriage ! 2-Stage SIP Formulation (Kong & Schaefer, 2004) 2 SWMP : r Max∑ ce xe + ∑ ps ∑ d es yes e∈E s =1 e∈E subject to : x ∑ δ e∈ ( v ) e + y ∑ δ e∈ ( v ) s e ≤ 1, ∀v ∈ V , s = 1, … , r xe ∈ {0,1}, ∀e ∈ E yes ∈ {0,1}, ∀e ∈ E , s = 1, … , r Complexity Results Maximum Weighted Matching is polynomial time solvable. (Edmonds, 1965, see also Cook et al., 1998) Complexity Results Maximum Weighted Matching is polynomial time solvable. (Edmonds, 1965, see also Cook et al., 1998) Maximum Stochastic-Weighted Matching is NP-Hard. (Kong & Schaefer, 2004) Even when G is bipartite. ) Complexity Results Maximum Weighted Matching is polynomial time solvable. (Edmonds, 1965, see also Cook et al., 1998) Maximum Stochastic-Weighted Matching is NP-Hard. (Kong & Schaefer, 2004) Even when G is bipartite. Surprising ? ) Complexity Results Maximum Weighted Matching is polynomial time solvable. (Edmonds, 1965, see also Cook et al., 1998) Maximum Stochastic-Weighted Matching is NP-Hard. (Kong & Schaefer, 2004) Even when G is bipartite. Surprising ? Is it the #scenarios (r) ? No. NP-Hard even when r=1. ) Combinatorial Restatement of 2SWMP Are there matchings M0, M1,…,Mr in G such that for s=1,…,r, M0∩ Ms = ∅ , M0∪Ms is a matching in G r and ∑ c + ∑ p ∑d e e∈M 0 s =1 s e∈M s s e is a maximum ? Approximation Algorithms KS-Algorithm (Kong & Schaefer, 2004) 1. Solve Max Wt Matching on G,ce → Z0,x0 2. For each s, solve Max Wt Matching on G,des → Zs, ys 3. If Z0 ≥ ∑rs=1psZs 1. 2. Then return ZAPX = Z0 and SOLAPX = (x0,0,…,0) Else return ZAPX = ∑rs=1psZs and SOLAPX = (0,y1,…yr) Theorem: ½ ZOPT ≤ ZAPX ≤ ZOPT Mix & Match Algorithm Define mixed weight wes = max (ce , des) M&M-Algorithm 1. For each s, solve Max Wt Matching on G,wes → Mc,s ∪ Md,s 2. 3. Find Mc = ∩s=1..r Mc,s Return and Z APX = r ∑ ce +∑ ps e∈M c s =1 s d ∑ e e∈M d ,s SOLAPX = ( M c , M d ,1 , … , M d ,r ) Mix & Match Bounds ? Z APX = ∑ e∈M c r ce + ∑ ps s =1 s d ∑ e e∈M d ,s ⎛ ⎞ s ZUB = ∑ ps ⎜ ∑ ce + ∑ d e ⎟ ⎜ e∈M ⎟ ∈ s =1 e M d ,s ⎝ c ,s ⎠ ⇒ Z APX ≤ Z OPT ≤ ZUB r r and ZUB − Z APX = ∑ ps s =1 ∑c e e∈M c ,s \ M c = E [weight of c − edges lost w.r.t mixed matching ] Mix & Match Bounds ? Z APX = ∑ e∈M c r ce + ∑ ps s =1 s d ∑ e e∈M d ,s ⎛ ⎞ s ZUB = ∑ ps ⎜ ∑ ce + ∑ d e ⎟ ⎜ e∈M ⎟ ∈ s =1 e M d ,s ⎝ c ,s ⎠ ⇒ Z APX ≤ Z OPT ≤ ZUB r r and ZUB − Z APX = ∑ ps s =1 ∑c e e∈M c ,s \ M c = E [weight of c − edges lost w.r.t mixed matching So What ?! ] Current Project Work Survey literature on approximation algorithms for stochastic combinatorial optimization problems Implement KS-Algorithm Implement M&M-Algorithm Compare performance on randomly generated instances Ratio 2 and (NO) PTAS ? Future Work Better Approximation ratios ? Best Possible ? Faster Algorithms ? Matching takes O(mn+n2logn) … use linear-time approximations. Multistage ? Random Graphs ? Future Work Better Approximation ratios ? Best Possible ? Faster Algorithms ? Matching takes O(mn+n2logn) … use linear-time approximations. Multistage ? Random Graphs ? Approximating Stochastic Combinatorial Optimization Problems is a new and hot area !