BabaIsRLAgent

• Create a virtual environment and run pip install -r requirements.txt.
• You can run things by running run.py.

  • See the code in if __name__ == "__main__" to see how things are run. Essentially, just specify a model to train, and evaluate it. The run_manually function is only used for debugging purposes (to make sure the enviroment and rewards were working correctly).

  • In the environment parameters, if train=True then randomization will be used. If object_to_shuffle is not set, then "complete" randomization is used; otherwise, only the specified object will be randomized.

  • To train with different reward weights, set reward_config= on the BabaWorldEnv object. The set of configs can be found in envs/reward_schemes.py

  • Command-line arguments:

    • --alg: one of 'dqn', 'a2c', or 'ppo'
    • --rewards: a space-separated list of "winlose", "nochange", "movetext", "distance", or "all". See envs/BabaWorldEnv.py for implementation of each reward scheme

Sample command: python -m run 1 --alg dqn --rewards winlose movetext