In this project you had to train two agents of tennis players, that try to keep the ball in the game as long as possible.

The base for this project and algorithm was layed down by the Udacity team from the DRLND, which we were allowed to use and to adapt. I followed the tip of the team to use the DDPG algorithm with a shared ReplayBuffer and a shared Actor Network.

To start, you need Python 3.6, PyTorch, Numpy, Matplotlib, and the Unity ML-Agents Toolkit.

With Anaconda or virtualenv you can create your python environment like: $conda create -n drlnd python=3.6 pytorch matplotlib numpy

For the Unity ML-Agent you need to download the Toolkit (https://github.com/Unity-Technologies/ml-agents) go to the ml-agents/python directory and install it via: $pip install ml-agents/python

Watch the trained agent with: $ python watch.py

and let the agent train again with: $ python train.py

The observation space for each agent consists of 8 values and the action space for each agent consists of 2 values, which are both continuous. The Environment is considered solved, after taking the maximum score of each player for each episode and then taking the mean of the last 100 scores reaches 0.5 or higher for one episode.

'Graph showing me solving the environment'

This algorithm even succeeded 0.7 over 100 consecutive episodes once and then bounced around the 0.5 mark. The environment was first considered solved at around 700 episodes.

Can be viewed in the "REPORT.md" file next to this file.

The agent so far is pretty good in a very short learning cycle. Although additional improvements can be made, like the Generalized Advantage Estimation(GAE) to improve the prediction of the expected return.