Reinforcement Learning for Solving Yahtzee
Reinforcement Learning for Solving Yahtzee
Minhyung Kang & Luca Schroeder
Stanford University
2018
Abstract
for Solving Yahtzee
Kang & Schroeder, 2010
---> DOWNLOAD <---
This paper presents a reinforcement learning approach for solving the dice game Yahtzee. The authors outline the challenges involved due to the massive state-action space, which makes traditional model-based or online solution techniques infeasible. They implement two reinforcement learning agents - one using perceptron Q-learning with eligibility traces, and another employing hierarchical reinforcement learning with the MAXQ value function decomposition. The hierarchical learning agent performed better, consistently outperforming random and naive greedy benchmark agents, though it remained suboptimal compared to limited lookahead search methods. The perceptron Q-learning agent showed limited success against the benchmarks. The study highlights the difficulties in solving Yahtzee optimally and provides avenues for future improvement, such as exploring deep reinforcement learning or alternative hierarchical reinforcement learning methods. The key challenges arose from the probabilistic and structural complexity of the game despite its simple ruleset.
