Parallelization of TSP Solving in CP

Abstract

The currently best CP method for solving the Travelling Salesman Problem is the Weighted Circuit Constraint associated with the LCFirst search strategy. The use of Embarrassingly Parallel Search (EPS) for this model is problematic because EPS decomposition is a depth-bounded process unlike the LCFirst search strategy which is depth-first. We present Bound-Backtrack-and-Dive, a method which solves this issue. First, we run a sequential solving of the problem with a bounded number of backtracks in order to extract key information from LCFirst, then we decompose with EPS using that information rather than LCFirst. The experimental results show that we obtain almost a linear gain on the number of cores and that Bound-Backtrack-and-Dive may considerably reduce the number of backtracks performed for some problems.