In this project we will use reinforcement learning to train a single agent to complete a simplified version of the Banana Collector using a Deep Q Network (DQN). This project will utilize Unity-ML to create the environment for learning.
The agent is navigating a 3-dimensional square world to collect yellow bananas for a reward of +1 and to avoid blue bananas give -1 reward. The state space of the environment is 37 dimensions and includes the agent’s velocity and ray-based perception of bananas forward of the agent. Using this, the agent must select the best actions to maximize the returned award. The four actions are:
- 0 - forward
- 1 - backward
- 2 - left
- 3 - right To solve the environment the agent must score 13 over 100 episodes.
Grab this repo using:
Mac:
conda create --name drlnd_navigation python=3.6
source activate drlnd_navigation
conda install -y python.app
conda install -y pytorch -c pytorch
pip install torchsummary unityagents
- Mac OSX: link Unzip in toplevel directory
To train an agent open up Navigation.ipynb and run all the cells until the Training Agent header. This will load the environment and display the state space and action space with which the agent will interact. To experiment with different training parameters run one of the cells under in the Training Agent header to train an agent that is able to solve the environment and visualize some graphs that show the agent’s learning process. Normally it takes 130 - 250 episodes for the agent to solve the environment.
Lastly to visualize a trained agent interact with the environment run the last cell which will load the weights from a success training run and display the agent for 1 episode.