Skip to content

Advanced topics in Artificial Intelligence - Course of my third year at Ecole polytechnique

Notifications You must be signed in to change notification settings

felixchalumeau/INF581-project

Repository files navigation

INF581-project

Advanced topics in Artificial Intelligence - Third year course at Ecole polytechnique

Repo for our final project: training an autonomous agent in the Gym Car-Racing environment.

Requirements

The project needs gym box2D environments. To install it:

  • you must have installed swig (http://www.swig.org/download.html), which might need Visual Studio Tools for C++ if you're on Windows
  • then do "pip install box2d-py" and it should works fine The project also needs keras and pytorch.

Files

  • play.py : allows to play the game (up for accelerating, left or right for steering). When a trial is succesful, the data is stored in data>data.pkl.zip. You can change the path at your convenience in the play.py file if necessary.
  • dqn.ipynb: contains the DQN, DDQN and DP (neural network policy) agents. Only the DQN works for the moment and obtains an average score of 814.0 with the model dqn_imitation_learning.h5 that is in the repo. You can train yourself the model by using agent.train(memory) where the memory comes from the function read_data() that reads the file data>data.pkl.zip
  • policy_gradient.ipynb: implements a policy gradient algorithm REINFORCE (see report for further documentation).
  • evolutional.ipynb: implements an evolutionary optimisation process.

Performances

Here is a video showing the behavior of the agent after a training period of 1000 epochs of 100 training samples. (~30min training)

Useful links:

About

Advanced topics in Artificial Intelligence - Course of my third year at Ecole polytechnique

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •