Project Mission
This project aims to design and implement Bloomward, a competitive hex based restoration strategy game for reinforcement learning research. The game focuses on flower placement, spirit activation, seasonal changes, and inward spreading corruption on a shared board. By formalising these mechanics into a structured digital environment, Bloomward will support both strategic human play and future intelligent agent training.
Project Documentation and Presentations
Term 1: The project focused on analysing and defining the foundation of Bloomward. This included refining the game concept, identifying the core mechanics, and formalising the requirements needed for the project to function as both a competitive strategy game and a reinforcement learning environment.
Term 2: The project focused on moving Bloomward from a design concept into a working prototype. This included implementing the core game mechanics, creating a playable Pygame interface, setting up the Gymnasium compatible environment structure, and preparing the training, testing, and evaluation files needed for future reinforcement learning experiments.
Term 3: The project focused on integrating reinforcement learning into the playable Bloomward prototype and evaluating agent performance. This included refining the game mechanics and user interface, implementing action masking with MaskablePPO, training agents across different configurations, and comparing their performance against a random-agent baseline. Preliminary results showed improved agent performance, while a structured testing plan was developed to assess gameplay functionality, environment behaviour, and reinforcement learning effectiveness.
The Team
Honours Student
4269355@myuwc.ac.za
Supervisor
mghaziasgar@uwc.ac.za