The aim of this task was to solve the tennis environment using multiple agent reinforcement learning methods.
The environment has a state space of 8, including positions and velocities of the ball and racket. The environment returns a stack of 3 states, hence the input of the networks are of size 24. The action space is 2 continuous action, including moving forward and backward, and up and down. The environment is considered solved when the average score is greater than 0.5 for 100 episodes, however, this is still no where near optimal play, so in the notebook, the average score is set to 1 for 100 episodes.
The dependencies that are required can be installed using the following:
First, install conda: https://www.anaconda.com/distribution/#download-section
Next, create a new conda environment and activate
conda create -n Tennis python=3.6.3 anaconda
conda activate Tennis
Next install pytorch using:
conda install pytorch=0.4.0 cuda80 -c pytorch
And ml-agents ugin:
pip install mlagents==0.4.0
Finally, the environment and scripts are downloaded from
git clone https://github.com/SamJCKnox/P3_Tennis_Submission.git
The P3_Tennis_Submission notebook is the header which calls all scripts required to run. Run all sections to train the agent. Outputs will show how the agent is performing. The last section shows the agent evaluation.
The networks trained in the current outputs of the Jupyter Notebook are in BenchmarkNetworks, copy these into the root directory to view in the evaluation section.
Report.md shows the architecture of the networks with the hyperparamteres.
The directory D4PG is the attempt at using D4PG with multiple branches with different structures and implementations, discussed fuyrther in Report.md.