Reinforcement Learning in the Pokémon Trading Card Game

Date of Award

8-2026

Document Type

Thesis

Degree Name

Bachelor of Arts (BA)

Department

Natural Sciences

First Advisor

Roy, Tania

Area of Concentration

Computer Science

Abstract

Reinforcement learning (RL) has shown strong performance in decision-making environments, but applying it to a trading card game presents challenges due to large, rule-constrained action spaces, stochastic transitions, and delayed rewards. This thesis looks into the application of RL algorithms on a simplified version of the Pokémon Trading Card Game (PTCG), which is designed to keep core gameplay mechanics while also providing a more efficient environment. A custom environment was developed by using a Node.js-based game engine with a Python RL training pipeline using a Gymnasium-compatible interface. The environment models core aspects of the game such as turn-based play, energy management, bench development, and prize-based win conditions. To support stable learning, the system incorporates an action legality mask to restrict invalid actions and a shaped reward function to provide overall learning signals. This thesis covers two RL algorithms: Proximal Policy Optimization (PPO), a policy-gradient method, and Deep Q-Networks (DQN), a value-based approach. PPO and DQN were both trained using self-play and evaluated using metrics including episodic reward, episode length, invalid action rate, and prize collection. Results show that both agents successfully learn rule-compliant behavior and effective strategies in order to win the game. While DQN demonstrates faster learning in the beginning, PPO produces more stable and consistent performance. This work presents a complete reinforcement learning framework for a structured card game environment, including environment design, training evaluation, and interactive visualization tools, and provides a foundation for future research in more complex game systems.

Rights

The author has not granted New College of Florida the nonexclusive right to archive, make accessible, and distribute for educational purposes this work in whole or in part in all forms of media, now or hereafter known. The copyright of this work remains with the author.

This document is currently not available here.

Share

COinS