search

nazaruka / gym-http-api

NSGA2-based Sonic agent + experimental code
MIT License
1 stars 1 forks source link

Organize the code to have a central access point #8

Closed schrum2 closed 5 years ago

schrum2 commented 5 years ago

For this code, I would like to have a main class that can be easily launched form the command line, and will accept command line parameters. The program should, at a minimum, accept parameters defining which domain to learn in (MountainCar, CartPole, etc) and which learning algorithm to use (DQN, DDQN, A2C, etc).

This will require separating out the domain specification from the learning algorithm, which is a problem with the current code. Therefore, this will be a fairly big issue. You will need to split up the existing code into more organized classes in a way that makes it easy to run any algorithm on any domain.

schrum2 commented 5 years ago

This can currently run either MountainCar-v0 or CartPole-v1 using wither A2C or DQN.

However, before the issue can be closed, support for more algorithms and domains needs to be added.

schrum2 commented 5 years ago

We may want to start incorporating Open AI baselines in addition to or instead of what we have been using: https://github.com/openai/baselines

schrum2 commented 5 years ago

Our code repo is a bit of a mess ... which makes this issue a bit pointless. We're currently focused so much on Sonic that the idea of having a broader system that can be applied to multiple domains with multiple algorithms is a bit out the window. Therefore, I'm going to close this issue ... let's just focus on Sonic.