Local Maxima in ADCOPs via Side Payments

Distributed constraints optimization problems (DCOPs) are composed of multiple constrained agents and their solution is an assignment of values by all agents that has the maximal global social welfare. DCOPs are known to be NP-Hard and therefore need incomplete search algorithms. Local search algorithms for DCOPs are incomplete distributed algorithms that start from some initial state and attempt to find improved solutions ( [10, 13]). When the constraints have different values for the constrained agents, the problems are termed Asymmetric DCOPs, or ADCOPs. Standard local search algorithms, such as DSA or MGM, are not guaranteed to converge on ADCOPs. When convergence does happen, the solution is not necessarily an extremum of the global social welfare. This is true also for local search algorithms that were designed specifically for asymmetric DCOPs [3]. The present paper proposes an iterative local search algorithm for ADCOPs, that relies on the analogy between ADCOPs and multi-agents games. In each iteration of the algorithm agents can propose side-payments to their neighbors, in return to their choice of an assignment that is preferred by the agent offering the side payment. It is shown that the proposed protocol produces a behavior of the distributed algorithm that is an extension of the best-response dynamics in muti-agent games. These properties of the proposed Bidding Efficient Equilibria Contracts (BEECon) algorithm guarantee convergence towards a local optimum of the global social welfare. Extensive experimental evaluation on randomly generated ADCOPs demonstrates that convergence is fast and that the resulting solutions are of higher social welfare than those of the best former ADCOP local search algorithm.