- Setup venv or conda env.
- Install deps
pip install -r requirements.txt - Active dev dirs
/catan/bots-- create new AI bots,/catan/analysis-- analyze playout stats and bot strategies.
- Preview playouts in UI
- Run script
run_server - Run script
run_ui
- Run script
- AI player agents
- Find a way to evaluate agent performance
- Measure win rate in multi-agent playouts
- MCTS Player
- Initial implementation
- Optimal MCTS player with action pruning
- RL
- Implement custom Gymnasium environment for Catan
- Decide which RL algorithms we might need (A2C, PPO, DQN, other...)
- [ ]
- Find a way to evaluate agent performance
- Game UI
- Show agentic playouts visible in UI
- Temporal Difference over Monte Carlo methods?
- What is the action space?
- How should we architect the neural network?
- Which RL algo should we use?
- On-policy vs Off-policy methods?
- When (if so) can we utilize self-play?
- NN
- Input features for NN?
- How to model randomness?
- Is RNN suitable?
- How to model agent strategies?
- How to model dueling?
This repo uses catanator's modules for catan core game logic