BoatAttack with ML-Agents (Windows Build - 1.0)

This branch contains the windows build version of BoatAttack with ML-Agents repo.

⚠ Warning:
This repo contains LFS tracked files. So contact me if you are facing any issues while cloning, forking, or downloading this repository because it's probably that I have exhausted GitHub's LFS quota. (Issues like not able to fetch all the files or some of the files that look missing)

Note:

• If you want to try out the BoatAttack with ML-Agents, then fork and clone this repo.
• You will find various build versions on different branches.
• Now checkout the branch suitable to your OS and run the executable file. 😎

YouTube Videos

Check out the above youtube video on "Navigation strategies learned by an ML-Agent" on a custom environment made with Unity Engine.

Environment Details

• Set-up: Environment where the agent needs to complete the lap in a minimum amount of time
• Goal: Reach the finishing line in a minimum amount of time
• Agents: The environment contains one agent (boat)
• Agent Reward Function (independent):
  • Speed * 0.001 at each step
  • -1.0 when agent crashes
  • checkpoint value * 0.00002
• Behavior Parameters:
  • Vector Observation space: 171 variables corresponding to
    • 56 ray-casts stacked 3 times to capture motion (Total 168 ray-casts) each detecting one of two possible objects (boundary & ground)
    • 1 variable stacked 3 times to capture the speed (Total 3 variable) of the agent
  • Vector Action space (Continuous): Size of 2 correspondings to agents rotation and forward/backward movement
  • Visual Observations (Optional): None
• Benchmark Mean Reward: 13.4

License

Licensed under the MIT License.

Acknowledgements

• Boat Attack environment by Unity Technology.