Extending dynamic backtracking for distributed constraint satisfaction problems

Abstract

A recent constructive search technique called dynamic backtracking (DB) achieves a systematic and complete search while allowing significant movement in the search space. The algorithm constructs tuples of inconsistent variable assignments called nogoods. An important issue is managing the number of nogoods constructed and remembered (cached) during the search. The nogood caching scheme for DB limits the size of the cache as the search proceeds through the 0space. Recently a new constructive search algorithm for the distributed constraint satisfaction problem (DCSP) was described called asynchronous backtracking (AB). In this method, agents construct nogoods and convey them to other agents to effect the backtrack search. A obvious question to ask if whether the nogood caching scheme employed by DB can be extended for the DCSP In this paper, we briefly analyse the existing DB caching scheme from this perspective and suggest two new improved caching algorithms. Finally we provide some preliminary experimental evidence that our new caching algorithms outperform dynamic backtrackinq in the multiaqent context.