reinforced-greediness
"Dermatoplastic shepard scratches its reinforced greediness by formulating without shock." — random text on the web
This is a bot for Legends of Code and Magic submitted to the IEEE CEC 2020's Strategy Card Game AI Competition. Made by Ronaldo Vieira, Luiz Chaimowicz and Anderson Tavares from Universidade Federal de Minas Gerais and Universidade Federal do Rio Grande do Sul.
Dependencies
Our bot requires Python 3.6+ and the numpy, scipy and sortedcontainers libraries. They can by installed by:
pip install numpy scipy sortedcontainers
Usage
python agent.py
Draft strategy
We use neural networks to choose cards. They are trained by reinforcement learning in a self-play setting, and a separate network is used when playing as first player and second player. This is part of Ronaldo's master's thesis.
Playing strategy
We find the best combination of actions with a best-first search on the current turn, using a simplified version of the forward model available in the gym-locm project. Our state evaluation function is formed by a linear combination of hand-made features, with coefficients found via Bayesian optimization.