Date of Creation
7-2025
Document Type
Departmental Honors Thesis
Department
Computer Science
First Advisor
Farhad Mohsin
Abstract
This work focuses on communication strategies within a cooperative multi-agent reinforcement learning system. The goal is to explore how communication can be used among agents to potentially improve performance. The research operates within the scope of “learning tasks with communication,” where the primary aim is to solve domain-specific tasks through information exchange using explicit communication protocols. Three distinct communication strategies were implemented and explored: Combinatorial Ghost, Feature Sharing Ghost, and Move Sharing Ghost. Fully Centralized Training and Execution and Centralized Training with Decentralized Execution training approaches were based on the different communication strategy used. Performance metrics were recorded for these strategies based on the last 10% of training episodes. The Pacman Training showed a Win % of 77% and Average Score of 1012.96. The Combinatorial Ghost had Win % of 40.3% and Average Score of 544.134, Feature Sharing Ghost observed Win % of 68.2% and Average Score of 923.704, and Move Sharing Ghost recorded Win % at 30.0% and Average Score of 437.099. The various communication strategies had differing levels of impact on how well the ghosts acted in the framework of the Pacman game.
Recommended Citation
Kalarickal, Matthew, "Exploring Communication in Multi-agent Cooperative Reinforcement Learning" (2025). Math and Computer Science Honors Theses. 65.
https://crossworks.holycross.edu/math_honor/65