Consistencies for Ultra-Weak Solutions in Minimax Weighted CSPs Using the Duality Principle Arnaud Lallouet1, Jimmy H.M. Lee2, and Terrence W.K. Mak2 1Université de Caen, GREYC, Caen, France Arnaud.Lallouet@unicaen.fr 2The Chinese University of Hong Kong, Shatin, N.T., Hong Kong {jlee,wkmak}@cse.cuhk.edu.hk 1 Introduction • Motivation • Minimax Weighted CSPs – Ultra-weakly solved, weakly solved, and strongly solved • Consistency Techniques 1. Lower Bound formulations 2. Upper Bounds using duality principle 3. Strengthening lower and upper bounds by adopting WCSP consistencies • Performance Evaluations • Conclusion & Future Work 2 Radio Link Frequency Assignment Problems • Soft Constrained Problems – Model: Weighted CSPs/COPs • CELAR Problem [Cabon et al., 1999]: – Given a set S of radio links located between pairs of sites – Assign frequencies to S: • Prevent/Minimize interferences – Involves two types of constraints 3 CELAR Problem: http://www.inra.fr/mia/T/schiex/Doc/CELAR.shtml Radio Link Frequency Assignment Problems D Communication from A to B A B Communication from B to A C 4 Radio Link Frequency Assignment Problems D Technological constraints |fAB - fBA| = constant B C fAB fBA A between two sites 5 Radio Link Frequency Assignment Problems D B fAB fBA fBC fCB C between links close to each other A Constraints to prevent interferences: e.g. |fAB - fBC| > threshold 6 Sometimes the problem is unfeasible… Radio Link Frequency Assignment Problems D B fAB fBA fBC fCB C between links close to each other A Soft constraints to minimize interferences: e.g. max(0, threshold - |fAB - fBC|) 7 Radio Link Frequency Assignment Problems insecure region Subject to control by adversaries fBD D fDB Minimize interferences? B C Minimize interferences? A 8 Radio Link Frequency Assignment Problems • Nature of the problem: – Optimization: Minimizing interferences – Adversaries: Controlling parts of the links • We can solve: 1. Many COPs/WCSPs • Each perform optimization on one combination of adversary’s frequency adjustment 2. Multiple QCSPs • Reducing into a decision problem 9 Radio Link Frequency Assignment Problems • Viewing in game theory: – Two-person zero-sum turn-based game • Allis [1994] proposes three solving levels: – Ultra-weakly solved • Best-worst case for a player – Weakly solved Stronger • Strategies for a player to achieve his/her best against all possible moves by his/her opponent – Strongly solved • Strategies for a player to achieve his/her best against all legal moves 10 Radio Link Frequency Assignment Problems insecure region Assume worst case adversary fBD fDB Minimize interferences a priori? Minimize interferences a posteriori? A D B Minimize interferences a posteriori? C Minimize interferences a priori? Finding frequency assignments for the worst possible case! 11 Minimax Weighted CSPs To avoid multiple sub-problems, we propose: Minimax Weighted CSPs = Min/Max + Soft Constraints + CSPs Quantifiers ≈ Weighted CSPs + Quantified CSPs 12 Minimax Weighted CSPs • Minimax Weighted CSP [Lee et al., 2011] Soft constraints – Variables: • x1, x2, x3 – Domains: • D1=D3 ={a,b,c}, D2 = {a,b} – Soft Constraints: – Global Upper Bound k: 11 – Valuation structure: • ([0..k] , ⊕, ≤ ) – Quantifier Sequence: • Q1 = max, • Q2 = min, • Q3 = max x1 Cost x2 Cost x3 Cost a 4 a 0 a 5 b 0 b 2 b 0 c 0 c 0 Unary constraint x2 x3 Cost x1 x2 Cost a a 1 a a 0 a b 1 a b 0 a c 0 b a 1 b a 0 b b 0 b b 2 c a 0 b c 0 c b 1 Binary constraint 13 A-Cost for Sub-problems max x1 min x2 max x3 a a x1 Cost x2 Cost x3 Cost a 4 a 0 a 5 b 0 b 2 b 0 c 0 c 0 b a b x3 Cost x1 x2 Cost a 1 a a 0 a b 1 a b 0 a c 0 b a 1 b a 0 b b 0 b b 2 c a 0 b c 0 c b 1 x2 ba a b c a b c 10 5 4 11 8 6 7 2 4 ⊕ 0 ⊕ 5 ⊕ 1 ⊕ 0 = 10 1 7 a b c a b c c 4 2 a a b 6 1 c 0 b a b c 8 5 3 14 max(10,7,6)=10 max x1 a min(10,11) = 10 a min x2 max x3 7 a b max(10,5,4) = 10 11 4 11 8 6 a 7 6 a b c a b c 7 1 7 2 6 b 7 a b c a b c 10 5 c b 4 a b 2 6 1 b 8 c 0 a b c 8 5 3 15 A-Cost for Sub-problems max(10,7,6)=10 max x1 a min(10,11) = 10 min x2 a max x3 7 a b max(10,5,4) = 10 11 a b c a b c c b 7 a b 6 a b 7 c a b c b 8 6 a b c a b c 10 5 4 11 8 6 7 2 1 7 4 2 6 1 0 8 Best-worst case (ultra-weak solution): {x1 = a, x2 = a, x3 = a} A-cost for the problem: 10 5 3 16 Algorithms for Ultra-Weak Sol. Previous Work [Lee et al., 2011]: 1. Alpha-beta prunings – Maintains two bounds • • Alpha lb: Best costs for max players Beta ub: Best costs for min players 2. Suggest Two sufficient conditions to perform prunings and backtracks Computing the exact A-cost is hard! Theorem: (NP-hard) For the set S of sub-problems P ’, where vi is assigned to xi: ∀P ’ ∈ S, A-cost(P ’) ≥ ub (Condition 1), or ∀P ’ ∈ S, A-cost(P ’) ≤ lb (Condition 2) We can prune or backtrack according to the table: A-cost(P ’) ≥ ub ≤ lb Qi = min prune vi backtrack Qi = max backtrack prune vi 17 Sufficient Conditions for Prunings How to compute efficiently? Corollary: For the set of sub-problems P ’ obtained from P, where vi is assigned to xi: A-cost(P ’) ≥ lbaf(P,xi = vi) ≥ ub (Condition 1), or A-cost(P ’) ≤ ubaf(P,xi = vi) ≤ lb (Condition 2) We can prune or backtrack according to the table below: lbaf(P,xi = vi) ≥ ub Qi = min prune vi Qi = max backtrack ubaf(P,xi = vi) ≤ lb backtrack prune vi 18 Consistencies • Local consistency enforcement – Make implicit costs information explicit • E.g. bounds, prunings/backtracks • Consistencies composes of 3 parts: 1. Lower bound estimation: lbaf(P,xi = vi) – NC & AC version 2. Upper bound estimation: ubaf(P,xi = vi) – Two dualities: DC & DQ 3. Strengthening lower & upper estimation by projections/extensions – Adopt WCSP consistencies: NC*, AC*, FDAC* – Naming convention: – DC-NC[proj-NC*], DQ-AC[proj-FDAC*] 19 Lower Bound Estimation • Lower bound estimation: lbaf(P,xi = vi) • Consider a simplified problem: – Only unary constraints, i.e. no binary Lemma: The A-cost of an MWCSP P with only unary constraints is equal to: Q1C1 ⊕ Q2C2 ⊕ … ⊕ QnCn x1 Cost x2 Cost a 4 a 8 b 1 b 6 c 2 c 1 Q1 = max ⊕ Q2 = min ⊕ x3 Cost a 1 b 3 Q3 = max =8 20 Lower Bound Estimation • Lower bound (NC version): nclb(P,xi = vi) CØ ⊕ (⊕j<i min Cj ) ⊕ Ci(vi) ⊕ (⊕i<j QjCj ) • Example: – nclb(P,x1 = b) x1 Cost x2 Cost a 4 a 8 b 1 b 6 c 2 c 1 Q1 = max – nclb(P,x2 = a) For all sub-problems where x2 = a Q2 = min x1 Cost x2 Cost a 4 a 8 b 1 b 6 c 2 c 1 Q1 = max Q2 = min x3 Cost a 1 b 3 Q3 = max x3 Cost a 1 b 3 Q3 = max 21 Lower Bound Estimation • Lower bound (AC version): aclb[Cij](P,xi = vi) – nclb(P,xi = vi) + a binary constraint Cij • Example: – aclb(P,x1 = b) x1 x2 Cost a a 17 5 a b 13 3 b a 11 2 b b 16 9 x3 Cost x1 Cost x2 Cost a 4 a 4 a 8 b 1 b 1 b 6 c 2 Q1 = max Q2 = min Q3 = max 22 Upper Bound Estimation • Upper bound ubaf(): Duality of Constraints Definition of Dual Problem: Given an MWCSP P = (X,D,C,Q,k). The dual problem of P is PΤ = (X,D,CΤ,QΤ,k) where: 1. Quantifier: Qi = max → QΤi= min & Qi = min → QΤi= max 2. Cost: For a complete assignment l, cost(l) = -1*costΤ(l) Construction Method: Q1 = max Q1 = min x1 Cost a 4 b 1 Q2 = min Q2 = max -1 x1 Cost a -4 b -1 x1 x2 Cost a a -7 a b -3 1 b a -1 6 b b -6 x1 x2 Cost a a 7 a b 3 b a b b -1 23 Upper Bound Estimation • Upper bound: Duality of Constraints (DC) – Corollary: A lbaf(PΤ,xi = vi) on the dual multiply by -1 is an ubaf(P,xi = vi) for the original problem 2 max x1 Upper bound ub : 10 Lower bound lb : 1 -2 c b min x1 1 2 b a min x2 2 max x3 10 a b c 0 1 2 a b c 0 10 10 1 0 max x2 11 a b c 1 0 a b c 0 11 10 2 = b) ≤ -11 -2 Τ,x = b) ≥-1 11 lbaf(P 2 b b a a -2 min x3 c bΤ lbaf(P ,x → -1 * b a Upper bound ub : -1 Lower bound lb : -10 a b c 0 -1 -2 -1 -10 a b c 0 -10 -10 -11 a b c 0 -1 0 a b c 0 24 10 -11 Upper Bound Estimation • Following the corollary: • We implement ubaf(P,xi = vi) by: – NC version: nclb(PΤ,xi = vi) – AC version :aclb[Cij](PΤ,xi = vi) • Advantage for Duality of Constraints (DC) – Reuse the same lbaf() – New lbaf() can be used as ubaf() 25 Upper Bound Estimation • Upper bound: Duality of Quantifiers (DQ) • Creating/Writing new ubaf() via: • Flipping quantifiers of existing lbaf() • Example: – nclb(P,x2 = a) For all sub-problems where x2 = a, guarantee a lower bound – ncub(P,x2 = a) For all sub-problems where x2 = a, guarantee an upper bound x1 Cost x2 Cost a 4 a 8 b 1 b 6 c 2 c 1 Q1 = max Q2 = min x1 Cost x2 Cost a 4 a 8 b 1 b 6 c 2 c 1 Q1 = max Q2 = min x3 Cost a 1 b 3 Q3 = max x3 Cost a 1 b 3 Q3 = max 26 Upper Bound Estimation • Upper bound: Duality of Quantifiers (DQ) • Creating/Writing new ubaf() via: • Flipping quantifiers of existing lbaf() • Immediate attempt: CØ ⊕ (⊕j<i min Cj ) ⊕ Ci(vi) ⊕ (⊕i<j QjCj ) min to max CØ ⊕ (⊕j<i max Cj ) ⊕ Ci(vi) ⊕ (⊕i<j QjCj ) • Problem: Binary constraints add costs! 27 Upper Bound Estimation • Upper bound: Duality of Quantifiers (DQ) • Creating/Writing new ubaf() via: • Flipping quantifiers of existing lbaf() • To fix: CØ ⊕ (⊕j<i max Cj ) ⊕ Ci(vi) ⊕ (⊕i<j QjCj ) • Further add maximum costs for constraints which are not covered in the function • For implementation: 1. We pre-compute and add these maximum costs before search 2. We maintain the added sum during search 28 Consistencies • We have methods to compute: – lbaf(): NC & AC version • Standard approximation analysis – ubaf(): Two dualities • Inspired from QCSP consistencies and algorithms • [Bordeaux and Monfroy, 2002] • [Gent et al., 2005] 29 Consistencies • Can we further strengthen both estimation functions? • Utilize projections & extensions conditions – WCSP consistencies: NC*, AC*, and FDAC* [Cooper et al., 2010] • For Duality of Constraints (DC) consistencies – Conditions are enforced in both the original and dual problem 30 Performance Evaluation • Compare and study different consistency notions – DQ-NC[proj-NC*], DQ-AC[proj-AC*], DQ-AC[proj-FDAC*] – DC-NC[proj-NC*], DC-AC[proj-AC*], DC-AC[proj-FDAC*] • Benchmarks: 1. Randomly Generated Problems 2. Graph Coloring Game 3. Generalized Radio Link Frequency Assignment Problem • Each set of parameters: – 20 instances & taking average result – If there are unsolved instances, we state the #solved besides runtime • Compare our results against: – Alpha-beta pruning – QeCode: A solver for solving QCOP+ 31 Performance Evaluation Randomly Generated Problems [Lee et al.,2011] – (n,d,p): (# of vars, domain size, constraint density) – Integer costs of a binary constraint • Generated uniformly in [0…30] for each tuple of assignments – Probability of 50%: a min (max resp.) quantifier Duality of Constraints Extracts costs from900s two different – Time limit: copies of constraints (original and dual) and resolve the issue Stronger projection/extension We may: •Strengthening lbaf() (ubaf() resp.) •Weakening ubaf() (lbaf() resp.) 32 Conclusion • Define and implement various consistency notions for MWCSPs 1. Lower bound by costs estimations 2. Upper bound by duality principle 3. Strengthening lower & upper bound estimation functions: • Adopting projection/extension conditions in WCSP consistencies • Discussions on our solving techniques on the two other stronger solutions 33 Related Work • Related CSP frameworks tackling adversaries: – Stochastic CSPs [Walsh, 2002] – Adversarial CSPs [Brown et al., 2004] – QCSP+/QCOP+ [Benedetti et al., 2007] [Benedetti et. al, 2008] • Other related frameworks: – Bi-level Programming – Plausibility-Feasibility-Utility framework [Pralet et al., 2009] 34 Future Work • Consistency algorithms: – High-arity Soft Table Constraints, and – Global Soft Constraints • Theoretical comparisons on different consistency notions • Algorithms tackling stronger solutions • Online & Distributed Algorithms • Value ordering heuristics – ICTAI 2012 35 Q&A 36 Performance Evaluation Graph Coloring Game [Lee et al.,2011] – Two player zero-sum games • Writing numbers of nodes – (v,c,d): (# of vertices, # allowed numbers, edge density) – Turns: Similar results • Odd/Even numbered turns - Player 1/Player 2 → A series of alternating quantifiers – Time limit: 900s 37 Performance Evaluation Generalized Radio Link Frequency Assignment – Designed according to two CELAR sub-instances – Minimize interference beforehand – (i,n,d,r): (CELAR sub-instance index, # of links, # of allowed frequencies, ratio of adversary links) – Time limit: 7200s Projection/extension in FDAC* • Slightly improves the search only • Quantifier info. not considered 38 Algorithms for Stronger Sol. • Solution Size – Ultra-weak: O(n) – Weak: O((n - m)dm) – Strong: O(dn) • Where: – # of variables: n – # of adversary variables: m – Maximum domain size: d • Ultra-weak solutions are linear 39 Algorithms for Stronger Sol. • Pruning Conditions – • A sound pruning condition when solving a weaker solution may not hold in stronger ones Invalid: •When finding weak solutions •Adversary max player Reason: – Theorem: Removal of the assumption of optimal/perfect plays Invalid: •When finding weak solutions •Adversary min player For the set S of sub-problems P ’, where vi is assigned to xi: ∀P ’ ∈ S, A-cost(P ’) ≥ ub (Condition 1), or ∀P ’ ∈ S, A-cost(P ’) ≤ lb (Condition 2) We can prune or backtrack according to the table: A-cost(P ’) ≥ ub ≤ lb Qi = min prune vi backtrack Qi = max backtrack prune vi 40 Relations with complexity classes • Weighted CSPs: – NP-hard Assumption: P ≠ PSPACE • Quantified CSPs: – PSPACE-complete Theorem: – Finding the truthfulness of QCSPs can be reduced (by Karp reduction) to finding the A-Cost of MWCSPs →MWCSPs: – PSPACE-hard 41 Transforming MWCSP to QCOP • Theorem: – An MWCSP P can be transformed into a QCOP P ’. The A-cost of P can be found by solving the optimal strategy of P ’. • Proof (Sketch): – Using ‘Soft As Hard’ approach [Petit et. al, 2001] • Transform soft constraints into hard constraints 42 Graph Coloring Game (GCG) Owned by B Maximize costs Owned by A Owned by A Owned by B Player A Minimize costs Player B Owned by B Owned by A Owned by A Owned by B How do they play the game? 43 Graph Coloring Game 5/B Write number 3 on node 1 Player A 6/A Write number 6 on node 2 2/B 1/A Player B 6 3 8/A 7/B 3/A 4/B Game Cost: |3 - 6| = 3 44 Graph Coloring Game 5/B 6/A 2/B 1/A Maximize costs 6 3 Place 0 Gain a cost of 3 Place 3 No cost gain Player A What should I do? 7/B 3/A 4/B 8/A so on… 45 Graph Coloring Game 5/B When the game terminates… 6/A 5 2 1/A 2/B 3 6 What we want to study… 9 1 7/B 5 3/A 8/A 0 4/B Final Game Cost: 55 46 5/B Approach 1: 6/A 1/A 2/B 8/A 7/B 3/A 4/B so on… 6/A 0 0 1/A 0 Modeled and solved by COP/ Weighted CSP 1/A 0 Modeled and solved by COP/ Weighted CSP 0 3/A 0 Modeled and solved by COP/ Weighted CSP 1 1 0 0 6/A 1/A 6/A 8/A 0 3/A 8/A 3/A 1 8/A 47 Approach 2: Modeling GCG 5/B 6/A 1/A 2/B 7/B 8/A 3/A 4/B 1. Guess a threshold: 56 2. Generate a Quantified CSP [Bordeaux and Monfroy, 2002] which asks: – Can player A finds numbers against player B’s moves – s.t. Player A gets costs < 56? 48 Modeling GCG • Approach 1: – Number of COPs/ Weighted CSPs constructed is exponential to the possible numbers player B can write • Approach 2: – Generate Quantified CSPs based on the objective function 49 x1 x2 Cost a a 5 a b 3 b a 10 b b 9 x2 Cost a a 5 a b 3 b a 10 b b 9 x3 Cost x1 Cost x2 Cost a 4 a 4 a 8 b 1 b 1 b 6 c 2 Q1 = max x1 NC Q2 = min Q3 = max AC Merge x1 Cost x2 Cost a 4 a 8 b 1 b 6 x1 x2 Cost x3 Cost a a 17 a 4 a b 13 b 1 b a 19 c 2 b b 16 Q3 = max Q1 = max Q2 = min 50 Original Problem x1 Cost a 4 b 1 x1 x2 Cost a a 7 a b 3 b a 1 b b 6 Q1 = maxQ2 = min Dual Problem DC x1 Cost a -4 b -1 x1 x2 Cost a a -7 a b -3 b a -1 b b -6 Q1 = min Q2 = max 51