This project is a part of:
Deep Reinforcement Learning Nanodegree
The project uses DDPG algorithm to solve 'Reacher Arm' environment.
Goal is to move 20 double-jointed arms to be in target position as long as possible. A reward of +0.1 is provided for each step that the agent's hand is in the goal location.
- There were 20 arms.
- Each arm contain 33 state observations
- Action space containd 4 contunuous values in range from -1 to 1.
- Each frame with arm in good location yields 0.1 reward
- Each episode has 1000 frames
- Overall goal is to train agent to keep all arms in target location for 100 epochs with average total score of 30
Below you can find a list of requirements required to run train.py & play.py scripts
- python 3.6
- Reacher app (this is delivered by Udacity Team) - I was using headless linux client to be able to run it seamlesly on any environment (even without display)
- torch
- numpy
- tqdm
- matplotlib
- unityagents
python3 train.pyImplementation approach & details and metrics can be found in report file.
Here is quick look for the results:
Link to full report