Improving the privacy of the asynchronous partial overlay protocol

Abstract

Dynamic, partial centralization has received a considerable amount of attention in the distributed problem solving community. As the name implies, this technique works by dynamically identifying portions of a shared problem to centralize in order to speed the problem solving process. Currently, a number of algorithms have been created which employ this simple, yet powerful technique to solve problems such as distributed constraint satisfaction (DCSP), distributed constraint optimization (DCOP), and distributed resource allocation.In fact, one such algorithm, Asynchronous Partial Overlay (APO), 1 was shown to outperform the Asynchronous Weak Commitment (AWC) protocol, which is one of the best known methods for solving DCSPs.One of the key differences between these two algorithms is that APO, as part of the centralization process, uses explicit constraint passing.AWC, on the other hand, passed nogoods because it tries to provide security and privacy. Because of these differences in underlying assumptions, a number of researchers have criticized the comparison between these two protocols.This article attempts to resolve this disparity by introducing a new AWC/APO algorithm called Nogood-APO that like AWC uses nogood passing to provide privacy and like APO uses dynamic, partial centralization to speed the problem solving process. Like its parent algorithms, this new protocol is sound and complete and performs nearly as well as APO, while still outperforming AWC, on distributed 3-coloring problems. In addition, this paper shows that Nogood-APO provides more privacy to the agents than both APO and AWC on all but the sparsest problems. These findings demonstrate that a dynamic, partial centralization-based protocol can provide privacy and that even when operating with the same assumptions as AWC still solves problems in fewer cycles using less computation and communication.

Distributed Constraint Problems for Utilitarian Agents with Privacy Concerns, Recast as POMDPs

“…When the upper and lower bounds meet at the root agent, a globally optimal solution has been found and the process finishes. Note that other approaches based on a graph re-arrangement have been explored, for example a cluster exploitation like Asynchronous Partial Overlay [39], [38]. d) Distributed Stochastic Algorithm (DSA): makes agents start their process by randomly selecting a value [67].…”