This repository provides source code for running the experiments described in the following work: (link here). The following modules are implemented:
-
pg_agent.py
module contains the code for training a policy gradient agent with entropy regularization -
fcnn_policy.py
module contains a PyTorch implementation of a multi-layer perceptron -
gridworld
module contains an implementation of the GridWorld as described in Sutton & Barto. The source code is taken from the UC Berkeley (http://ai.berkeley.edu) -
train.py
is a scripts that trains a policy gradient agent to play the gridworld. To run the training execute the following from thesrc
directory:python3 train.py --grid SmallGrid --iters 10001 --episodes 32 --entropy_reg 1.0
Logging information with history from the training is saved inside a
logs
directory. -
plot.py
is a script that creates plots from the training history. To run this script execute:python3 plot.py
Before running this script the agent must be trained on both
SmallGrid
andConfuseGrid
with and without entropy regularization.