I was trying to solve a partially initialized sudoku puzzle (the kind that appears in newspapers) with the ‘Drools Planner’ package. While it can generate a (random) puzzle from scratch in 3 seconds, it gets stuck in a loop solving a partially initialized puzzle .
Question: Are heuristics such as tabu search and simulated annealing fundamentally a bad choice for sudoku? I am talking of completeness (will the solution be reached) and efficiency (is it overkill).
My doubt comes from the fact that a sudoku puzzle always has an exact and single solution and heuristic algorithms are (AFAIK) not designed to “reach them”.
My answer to your yes/no question, if heuristics such as tabu search and simulated annealing are a bad choice for solving Sudoku, is yes.
The problem has far too many constraints for local search strategies to be efficient.
Sudoku is a good example for a constraint satisfaction problem (CSP), and CSP solvers are very good at solving it. That doesn’t mean that local search won’t work or that heuristics are a bad idea in general, but this problem can be solved pretty easily by propagating constraints.