AI_Gym

A collection of reinforcement learning approaches to games from the OpenAI gym

https://gym.openai.com/

We have written some convenience wrappers around the source code developed by OpenAI, to make it easier for the user to play around with network architectures in the context of a time-limited hackathon.

The HLML_RL class is defined in rl_class.py and is the main object used for writing quick scripts such as run.py with custom PyTorch models. To build a custom model, the user must write their own ActorCritic in customActorCritic.py, following the documentation provided by the HLML_RL.ac_help() method.

Minimal user steps for training your own agent:

1. Choose an OpenAI RL environment: https://github.com/openai/gym/wiki/Table-of-environments. Default: 'LunarLander-v2'.
2. Choose a compatible RL training algorithm (see compatibility_checks.py). Default: 'vpg'.
3. In customActorCritic.py, specify your chosen algorithm as TRAINING_ALG
4. In customActorCritic.py, specify your default algorithm hyperparameters in CUSTOM_AC_DEFAULT_KWARGS
5. In customActorCritic.py, write an ActorCritic following the documentation provided by the HLML_RL.ac_help() method.
6. Train the model using run.py as described below.

Training the model:

• Modify user_input in run.py, which is used to initialize an instance as HLML_RL(**user_input)
• Modify train_input in run.py, which is used to set hyperparameters for training in HLML_RL.train(**train_input)
• Run run.py

Hardware Case 1: CPU only, multiple threads

• Modify the default parameters for the training algorithm being used
• Specifically, need to override ncpu: 1 in user_input defined in run.py.
• Only some algorithms (e.g. VPG) support MPI for multi-processing (use ncpu=1 otherwise).

Hardware Case 1: GPU

• TODO