Load balancing strategies for parallel forward search algorithm with conflict based backjumping

Constraints satisfaction problems (CSP) is a well studied example of the NP-complete family which is usually solved by backtracking search algorithms. Performance comparisons of different algorithms led us to consider FC or FC-CBJ as one of the best sequential algorithms to solve a great family of CSP's. In this paper we give an explicit formulation for a generic parallel search algorithm based on typical parallel behaviors, such as load balancing policy. From this paradigm, we exhibit parallel FC-CBJ. The experimental study is up to now restricted to the performance of this within the scope of phase transition phenomenon for randomly generated binary CSPs and more particularly in the mushy region where the hardest problems are encountered. We discuss the performances of our algorithm relatively to a server initiated load balancing strategy with four techniques of tree-search splitting and four partner selection policies.