摘要: Most CSP algorithms are based on refinements and extensions of backtracking, employ one two simple "branching schemes": 2-way branching or d-way branching, for domain size d. The schemes not equivalent, but little is known about their relative power. Here we compare them in terms how efficiently they can refute an unsatisfiable instance with optimal choices, by studying variants the resolution proof system, denoted C-RES NG-RES, which model reasoning algorithms. tree-like restrictions, tree-C-RES tree-NG-RES, exactly capture power backtracking respectively. We give a family instances require exponential sized search trees have O(d2n) branching. also natural strategy finds refutations these time O(d2n2). unrestricted NG-RES simulate incorporate learning k-consistency enforcement. show separations between as well versions each system. All given nearly optimal.