AlphaSnake: Policy Iteration on a Nondeterministic NP-Hard Markov Decision Process (Student Abstract)

Reinforcement learning has been used to approach well-known NP-hard combinatorial problems in graph theory. Among these, Hamiltonian cycle are exceptionally difficult analyze, even when restricted individual instances of structurally complex graphs. In this paper, we use Monte Carlo Tree Search (MCTS), the search algorithm behind many state-of-the-art reinforcement algorithms such as AlphaZero, create autonomous agents that learn play game Snake, a centered on properties cycles grid The Snake can be formulated single-player discounted Markov Decision Process (MDP), where agent must behave optimally stochastic environment. Determining optimal policy for defined maximizes probability winning -- or win rate with higher priority and minimizes expected number time steps lower priority, is conjectured NP-hard. Performance-wise, compared prior work game, our first achieve over 0.5 (a uniform random achieves < 2.57 x 10^{-15}), demonstrating versatility AlphaZero tackling problems.