Reinforcement Learning Plays Catan
Feb 2023 - Mar 2023
What?
The project aimed to optimize initial settlement placement in the strategic board game Catan.
It sought to develop an algorithm for improving strategic decisions made by human players in the game's initial phase.
How?
Utilized deep Q-learning, a model-free reinforcement learning method.
Leveraged existing Catan simulator for testing and training
Defined specific state and action spaces (Sec IV. C of the report).
Implemented a neural network for action value function approximation.
Employed ϵ-greedy method for balancing exploration and exploitation during training.
Results
The algorithm outperformed random policy players but not heuristic-based players.
Demonstrated better-than-random decision making in initial placement.
Suggested further iterations and exploration of different neural network architectures for potential improvement.
Check out the report below for more detail!