Solving nonogram puzzles by reinforcement learning

We study solvers of nonogram puzzles, which are good examples of constraint-satisfaction problems. Given an optimal solving module for solving a given line, we compare performance of three algorithmic solvers used to select the order in which to solve lines with a reinforcement-learningbased solver. The reinforcement-learning (RL) solver uses a measure of reduction of distance to goal as a reward. We compare two methods for storing qualities (Q values) of stateaction pairs, a lookup table and a connectionist function approximator. We find that RL solvers learn near-optimal solutions that also outperform a heuristic solver based on the explicit, general rules often given to nonogram players. Only RL solvers that use a connectionist function approximator generalize their knowledge to generate good solutions on about half of unseen problems; RL solvers based on lookup tables generalize to none of these untrained problems.