Solving Combinatorial Problems through Off-Policy Reinforcement Learning Methods

In recent years, reinforcement learning (RL) has emerged as a strong candidate to solve many decision-making problems. It fundamentally combines the benefits of both deep neural networks and control engineering practices to accomplish human-level or even superior competency in decision-concise systems. This work extends this very idea and proves the applicability of RL algorithms to combinatorial puzzles such as river-crossing puzzle and seating arrangement problems. In that regard, this work applied quality-networks (QN), deep-quality-networks (DQN), and double-deep-quality-networks (DDQN) to environments mentioned above, and afterward, evaluated their performance through analyzing reward and performance plots. The study concludes that all these methods effectively completed the puzzle tasks with unnoticeable performance differences for the given environments.