Date of Award

Fall 2024

Project Type

Thesis

Program or Major

Computer Science

Degree Name

Master of Science

First Advisor

Wheeler Ruml

Second Advisor

Alexander Dockhorn

Third Advisor

Marek Petrik

Abstract

Duelyst II is an online collectible card game (CCG) that features a 9 × 5 grid board,making it a cross between the popular CCG Hearthstone and chess. It is a partially- observable stochastic game (POSG) with a large branching factor and the ability to take several actions in a time-limited turn, making it a challenging domain for game-playing AI. On top of that, successor generation is very slow, limiting the effectiveness of traditional planning approaches such as Monte Carlo tree search (MCTS).

In this thesis, I build on my previous work on Duelyst II AI, creating a stronger game-playing agent than the built-in rule-based one that is also practical to integrate into the live game. I accomplish this through enhanced planning, then beat that agent with one that uses action abstractions. I also investigate different methods of behavior cloning from game logs for enhancing abstraction-based planners. I find that learning leads to more intelligent selection of certain types of actions, but that those action types are not critical to an agent’s strength.

Share

COinS