AI Airplane

Multi Agent Reinforcement Learning with Unity and MLAgents

Requirements

• Unity 2020.1.10f1
• ML Agents Version 1.4.0-preview - September 21, 2020 (within Unity)
• mlagents 0.21.0 (within python)
• python 3.8.5

Agent

• 3 ray sensors with 5 beams per direction
• 3 different models
  • Normal
  • Snow Hard
  • Desert Hard

Environment

• 2 areas + 2 training areas
  • training areas don't have a normal player
  • environment is multiplied 4x for faster training
  • environment contains ground, stones and water with mesh collider

Learning

• Learning config can be found in config/trainer_config.yaml
• algorithm: PPO (Proximal Policy Gradient)
• modified to work with mlagents 0.21 for python 3.8
• curricula with four lessons

References

• based on the Udemy course Reinforcement Learning: AI Flight with Unity ML-Agents