Experimental Comparison of BTD and Intelligent Backtracking: Towards an Automatic Per-instance Algorithm Selector

We consider a generic binary CSP solver parameterized by high-level design choices, i.e., backtracking mechanisms, constraint propagation levels, and variable ordering heuristics. We experimentally compare 24 different configurations of this generic solver on a benchmark of around a thousand instances. This allows us to understand the complementarity of the different search mechanisms, with an emphasis on Backtracking with Tree Decomposition (BTD). Then, we use a per-instance algorithm selector to automatically select a good solver for each new instance to be solved. We introduce a new strategy for selecting the solvers of the portfolio, which aims at maximizing the number of instances for which the portfolio contains a good solver, independently from a time limit.