Risk Based Negotiation of Service Agent Coalitions Bastian Blankenburg, Matthias Klusch Minghua He, Nick Jennings DFKI University of Southampton 1 Collaboration of Service Agents Service Requesters Service Provider Agents • Independent • Rational sra1 spa1 ws1 spa2 ws2 Deadline: t1 Plan: <ws2,ws1> spa3 ws3 Deadline: t2 Plan: <ws3,ws1,ws2> 2 Service Agent Coalition Formation Coalition negotiation • Set of requests, set of composition plans • Which plans to execute? – Do the agents have enough resources? – Is a plan profitable? – What about the costs in case of failure? • How to share the profit (or loss)? – Stability: avoid that agents break their coalitions 3 Example: Medical Information Provision Request diagnosis, spa1 offer: 250€, ws1 deadline: 10min Coalition Proposal C1 reward: 250€ my costs: 10€ deadline: 10min my runtime: 5-6min spa2 ws2 spa3 ws3 Coalition Proposal C2 reward : 150€ my costs: 15€ deadline: 10min my runtime: 1-2min C1 my runtime: 3-5min my costs: 40€ Might fail! C2 my runtime: 1-2min my costs: 10€ On the safe side! If C2 then I can afford to risk C1! 6 Assessing Coalition Risk (1) Financial Risk Measures • Informal Definition – Combination of the probability of undesirable outcomes and their net results • Coherency (Artzner et al. 1999) – Translation invariance, positive homogenity, monotonicity, subadditivity risk ( X Y ) risk ( X ) risk (Y ) – Tail Conditional Expectation TCE TCE E X | X inf x : P X x • Expected loss in α worst cases • Based on Value-at-Risk 8 Assessing Coalition Risk (2) • Service instances in a plan are executed sequentially • Probability functions for instance runtimes • Composed service runtime – Sum of random variables: convolution of PDFs – Equal to point-wise multiplication of Fourier Transforms – Fast approximation with FFT • Probability of Failure/Success spa2 spa1 Composition Plan: 0,25 0,45 0,4 0,2 0,35 0,3 0,15 0,25 0,2 0,1 0,15 0,1 0,05 0,05 0 0,00 0,50 1,00 1,50 2,00 2,50 3,00 3,50 4,00 4,50 5,00 5,50 6,00 6,50 7,00 7,50 8,00 8,50 9,00 9,50 10,00 0 0,00 0,50 1,00 1,50 2,00 2,50 3,00 3,50 4,00 4,50 5,00 5,50 6,00 6,50 pdf tPlan( x) pdf tA ( y ) pdf tB ( x y )dy 0 F 1 ( F ( pdf tA ) F ( pdf tB )) PoF ( Plan, tStart, DL ) pdf tPlan ( x)dx DL tStart PoS 1 PoF 9 Fuzzy Coalition Model ~ C ~ C ({spa1 / mem1;; span / memn }, sra, Plan) • Fuzzy Coalition – Bound to request and plan – Coalition membership degree in [0,1] – Fraction of resources per time – Determines service instance runtimes, PoF and PoS • Values of a fuzzy coalition – Reward r is paid only if of successful – Expected reward – Expected value ~ r PoS (C ) r ~ C v(C ) r C~ cost i instances • Fuzzy coalition structure – Set of fuzzy coalitions – Feasibility wrt. resources spa : mem ~ CS ~ C spa 1 10 Stability in SPA Fuzzy Coalitions Membership vs. PoS • Existing approaches (Aubin; Bunariu;Nishizaki,Sakawa) Mean Runtime PoS = P (Runtime < 4) PoS = P (Runtime < 7) 80 1,0 0,9 70 0,8 60 0,7 Minutes • Shapley value, Core, Nucleolus and others • Assumption: coalition value is proportional to membership degrees – does not hold – runtime is 1/x. – PoS/PoF and expected value not proportional – PoS must not be overestimated! Single-agent coalition, normal distribution with min. mean runtime = 3, σ=1/mem. 50 0,6 40 0,5 0,4 30 0,3 20 0,2 10 0,1 0 0,0 100% 0% 20% 40% 60% 80% Membership 12 Stability in SPA Fuzzy Coalitions (2) e(C, u) v(C ) • Recall: excess of a coalition: • Excess of a fuzzy coalition – Any amount of membership can be transferred – Coalition structure might be too risky for a member aC u (a) • Should such coalitions be considered a feasible threat? • Mutual dependency of risk and payoff – How is an agent‘s payoff affected by withdrawing a certain amount of membership? – Consider conditional expected values ~ ~ ~ e(C , u) v|TCE (C ) aC~ min. payoff attenuatio n|TCE (a, C ) 13 Stability in SPA Fuzzy Coalitions (3) • Kernel – Surplus • „I can gain more without you, than you without me“. • max. excess of coalitions excluding the other agent • With fuzzy coalitions, it is possible to transfer membership to multiple other coalitions at the same time – Kernel-stable solution: equilibrium of surplusses – Computation: transfer scheme 14 Complexity • Computation of surplus depends on computation of TCE and vice versa • Both have exponential computation time • How to do it (highly) polynomial: – Compute upper bounds for TCE: • Consider minimum individual rational payoffs • Use subadditivity when forming additional coalitions • Refine bounds while there is time – Add some constraints to the game to compute surpluses • Bound the max. coalition size, number of plans per coalition and number of coalitions that an agent can join 15 Rational Service Agent Model • Service Request Agent – Represents a SWS request – Specifies a deadline – Provides a monetary reward for timely execution • Service Provider Agent – Offers one SWS – Has an SWS composition planning module – Has Bounded resources, – May split resources among multiple service instance executions, – Computes probabilistic estimations of service instance execution times, by e.g. • Learning • Stochastic process modeling (Manolache et al. 2004) – Produces a fixed cost for any service execution 16 RFCF Outline Each agent performs in parallel: • Composition Planning • Coalition Negotiation 1. Proposal generation i. Minimize memberships s.t. risk is acceptable ii. Maximize payoff / membership 2. Proposal evaluation: form feasible coalitions with i. acceptable risk ii. maximal payoff / membership 3. Payoff distribution and risk bound update i. Transfer Scheme ii. Compute single-coalition TCE and add to coalition structure TCE • Risk Measure Computation 1. Compute exact TCE for new random subset of coalitions until service execution start time 18 Conclusions • Adavantages – Anytime approach – Guaranteed risk bounds wrt. individual risk averseness – Gradually improvement of • risk assessment • coalition structure • Drawbacks/Simplifications – Complexity: • Exact solution has exponential runtime • Constrained solution still has highly polynimial runtime – Independent service runtime assumption – Static setting • service execution start time • for the dynamic case: when to stop negotiation? 20