tdardinier / InterGP

• agent.py: Defines what the methods an agent (a policy) should define

• agents: This folder contains several agents (policies). The "random" agent is currently the only one that fully works.

• baselines: The baselines written by OpenAI.

• controller.py: Takes an agent (policy) and an environment, and runs the policy on the environment.

• definitions.py: Defines all the variables that can be used with main.py: the enviromnents, the agent(s), the predictors, the default values.

• evaluator.py: Takes a ReplayBuffer file (from replays/) and a predictor, and evaluates it. Saves the result in a file (in results/).

• main.py: "Demo" file which implements the 3 main steps that can be done.

  1. Collect data from an environment and save it in a ReplayBuffer file.
  2. Evaluate a predictor on a ReplayBuffer and save it in a Result file.
  3. Visualize the results of the predictors.

• modelWrapper.py: A meta-agent (policy), which takes another agent (policy) and runs it, while evaluating a predictor or saving a ReplayBuffer.

• models: This folder will contains in the future the DeepQ policies learned by the DeepQ baseline.

• policyExamples.py: Obsolete demo file for RL examples (LQR, discrete Q-Learning, ...)

• predictor.py: Defines the methods a Predictor should define.

• predictors: This folder contains several "Predictors", ie. model that can be trained and used to predict the dynamics. Currently contains:

  • A "full" neural net
  • A "linear" neural net (Xt+1 = A(Xt) + W(Xt) * Ut)
  • A Gaussian Process
  • An "identity" predictor (ie. Xt+1 is predicted to be Xt)

• replayBuffer.py: The class that handles the replay files to record episodes and load them.

• result.py: The class that handles the result files to store predictions made by predictors, and to visualize errors.

• results: This folder contains the results generated by evaluator.py.

• simulation.py: Obsolete file.

• tools.py: A bunch of useful classes and functions are coded here (LQR, normalizer, Gaussian Processes).

• visualisator.py: The class that handles the visualisation of the results.