MooZi: A High-Performance Game-playing System that Plans with a Learned Model

• Author / Creator

  Wang, Zeyi

• The intent of this thesis is to develop a high-performance open-source system that plans with a learned model and to understand the algorithm through extensive analysis. We formulate the problem of maximizing accumulated rewards in Markov Decision Processes, and we frame playing games as such problems. We develop the MooZi system to solve these problems. MooZi includes (1) a MuZero-based algorithm that plans with a learned model (2) a distributed architecture that trains and evaluates the algorithm efficiently, and (3) a collection of tools to visualize and understand the algorithm. We empirically show that MooZi outperforms PPO and AC in MinAtar environments. We also show that MooZi learns to play the two-players board game Breakthrough. We use our tools to analyze the learned model by visualizing search trees and learned representation. We make MooZi publicly available to accelerate future research.

• Subjects / Keywords

  • planning
  • deep learning
  • reinforcement learning
  • MuZero

• Graduation date

  Spring 2023

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-vkjq-h964

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.