In this project we are solving the "Reacher" environment using the open-source Unity plugin Unity Machine Learning Agents (ML-Agents). We are using the variant with only a single agent. The main use-case of those environments is to use trained agents for multiple purposes. But we are using it for designing, training and evaluating the performance of our own agent.
The goal in the environment "Reacher" is to keep a double-jointed arm at a moving target location which is a ball moving in a circle around the arm. This goal is definined by the reward rule +0.1 for each time step the hand of the arm in the goal location.
The continuous action space is a 4-dimensional vector, each entry between -1 and 1 and corresponding to torque applicable to two joints. The state space consists of the position, rotation, velocity, and angular velocities of the arm. It has 33 dimensions.
The environment is considered to be solved after having achieved an average score of 30 in 100 consecutive episodes.
- Clone the repository https://github.com/fjonck/Project_2_Continuous-Control
- Open the notebook Continuous_Control.ipynb with Jupyter Notebook
- In the menu, click on Cell -> Run All
- Enjoy watching the Reacher Agent learn :)