Supporting custom state representations
cdvv7788 opened this issue · comments
Is your feature request related to a problem? Please describe.
The current state representations is nice if you want to work with the whole image. If you want to try something simple, like Q-learning, the state representation needs to change.
Describe the solution you'd like
I would like to have a way to encode the state into something different than the current image (http://cs229.stanford.edu/proj2012/LiaoYiYang-RLtoPlayMario.pdf)
Describe alternatives you've considered
If all the necessary information could be encoded in the info field, that information could be used to encode the state instead of the raw state provided, without affecting the current API.
What I believe you're describing is manual feature engineering at the game logic level. While a totally valid technique, it is not a primary objective of this project which aims to be a simple model of human interaction with the game. However, feel free to create a fork! This idea has been explored by @ppaquette and is also referred to in #85.