Tensorflow implementation of the REINFORCE policy gradient algorithm, developed as a class project for the course "Stochastic Processes and Reinforcement Learning". See the PDF project report for details.
Train a REINFORCE agent on Gym MountainCar-v0 environment (feel free to stop training when the desired episode reward is achieved, it tops at around -105 to -110 on average, but it may take a long time):
python -m train_mountaincarMonitor the training progress in Tensorboard:
tensorboard --logdir .experiments/Model with the best average reward will be saved to the filesystem, use the following command to visualize the policy behavior:
python -m enjoy_mountaincar
Train on Gym CartPole-v0 environment (trains much faster than MountainCar, mere 200 episodes should be enough for a good policy):
python -m train_cartpoleMonitor the training progress:
tensorboard --logdir .experiments/Use the following command to visualize the behavior:
python -m enjoy_cartpole
If you have any questions please feel free to reach me: apetrenko1991@gmail.com