Constraint relaxation in distributed constraint satisfaction problems

The distributed constraint satisfaction problem (DCSP) formulation has recently been identified as a general framework for formalizing various distributed artifical intelligence problems. The author extends the DCSP formalization by introducing the notion of importance values of constraints. With these values, a solution criterion is defined for DCSPs that are over-constrained (where no solution satisfies all constraints completely). It is shown that agents can find an optimal solution with this criterion by using the asynchronous incremental relaxation algorithm, in which the agents iteratively apply the asynchronous backtracking algorithm to solve a DCSP, while incrementally relaxing less important constraints. In this algorithm, agents act asynchronously and concurrently, in contrast to traditional sequential backtracking techniques, while guaranteeing the completeness of the algorithm and the optimality. Furthermore, it is shown that, in this algorithm, agents can avoid redundant computation and achieve a five-fold speed-up in example problems by maintaining the dependencies between constraint violations (nogoods) and constraints