Super Auto Pets AI
This project has 3 major components:
- Game Engine (90% complete, 320 of 340 tests completed)
- Reinforcement Learning AI
- DQN Complete
- AQN In Progress
- A2C In Progress
- A3C In Progress
- Future Revisions TBD
- Selenium Webdriver (to be completed)
The game engine mimics the engine which drives the strategy auto-battler game Super Auto Pets (c. Team Wood Games) as near as I could through reverse engineering the minutiae of the logic behind the game mechanics.
The engine has its own program interface with intuitive controls analogous to the discrete actions possible in the real game. Results are non-deterministic but discrete. The program interface has a builtin save and load method for easily changing contexts.
The engine is also packaged as an Open AI Gym environment.
Three RL AIs are to be trained on this game, one DQN, one AQN, and one A3C network.
The webdriver is to be an implementation of the AI, used so players may play against the AI in the real game in the 1v1 versus mode. Image analysis is used to convert the game image to the state representation used by the AI and project's engine, and the webdriver performs the actions based on the AI's decisions.
The image analysis will be completed using numerous fixed position crops followed by basic feature matching to determine which unit is in which position, if it is holding an item, and what its stats are.