timbernard/drlnd-p2-continuous-control

Deep Reinforcement Learning Nano Degree: Project 2, Continuous Control

Udacity Deep Reinforcement Learning: Project 2, Continuous Control

Environment

• Reacher Environment
• The goal of this problem is for the agent (a double jointed arm) to track a goal location, where it gives a reward of +0.1 for each timestep the hand is in the goal location
• For the purposes of this project, the environment is considered solved when the agent achieves an average score (accumulated reward over an episode) of +30.0 over 100 episodes.

State Space:

• 33-Dimensional continuous observation space, (includes agent position, orientation and there respective velocities)

Action Space:

• 4-Dimensional continuous space (torque applied to two joints), with each element bounded to be within (-1,1)

Installation and Usage

• This notebook can be used by simply opening it in the provided Udacity Project Workspace
• Alternatively, you can run this locally:
  1. Clone this repository
  2. To get the python/conda environment setup, follow these instructions: https://github.com/udacity/deep-reinforcement-learning#dependencies
  3. Make sure you have jupyter properly installed, and you may have to get an archived version of pytorch
  4. Download the environment for Linux: https://s3-us-west-1.amazonaws.com/udacity-drlnd/P2/Reacher/one\_agent/Reacher\_Linux.zip
  5. In the repository directoy root mv <path/to/Reacher_Linux.zip\> .
  6. Then run conda activate drlnd
  7. In your local directory, to see the trained agent in action: python main.py "./Reacher\_Linux/Reacher.x86\_64"
  8. To train the agent yourself, use the ipynb notebook and use the Udacity Project Workspace as stated above so that you can take advantage of the GPU, hit "run all cells"