FeReRA: A Multi-agent Approach to Constraint Satisfaction

Abstract

This work introduces the FeReRA (Feedback, Reinforcement, and Reactive Agents) algorithm, an extension to the ERA framework [3]. ERA (Environment, Reactive Rules, and Agents) introduced a non-deterministic, self-organising, multi-agent approach to solving Constraint Satisfaction Problems (CSP) whereby a problem is broken up into smaller sub-problems and each sub-problem is to be solved by an independent agent. Each agent inhabits a local environment which represents the domain of its respective variables, and will try to find solutions to its sub-problem by seeking positions in the environment that translates to consistent value assignments for the variables in the sub-problem. ERA’s key strength is that it is capable of finding solutions to CSPs without much computational overhead. Its Achilles’ heel, however, is its inconsistent performance resulting from inbuilt random behaviours which are relied on to escape local optimums. To overcome this weakness, FeReRA extends ERA by replacing the random decisions with a deterministic feedback mechanism that helps the algorithm to decide which agents are to make non-improving moves necessary to take it out of a local optimum. This feedback process is structure dependent, and it takes into account the cumulative effects of individual agents’ behaviours on the global state of the system. This is in contrast with the approaches taken in similar work [2][4], where emphasis is placed on individual agents detecting and escaping quasi-local minimums . Preliminary results from experiments on graph colouring instances are as follows: First, on critically constrained instances, the time taken to find solutions on average was equal or better with FeReRA compared to ERA. Secondly, on over constrained instances the quality of solutions found was consistent with FeReRA and was on average better or equal to ERA’s performance. An extended version of this abstract is available at [1].