This repository was created for the semesterproject in the second semester of the Interactive Media master at the University of Applied Sciences Media Department in Hagenberg.
Additional Settings:
- 20 Physics frames per Second
- Agent Desicion Period: 15 Physics Frames = 1 Agent Frames The agent currently works with the following principles:
- If agent dies: Reset Car Position of Agent to start instead of reloading scene (would reset ML-Agent) -> Unloading & Reloading Terrain (Resetting Gamemanager manually)
- Agent can use the following:
- Fuel state
- Fuel consumption
- Current Velocity (normalized)
- Velocity history (last 5 Physics frames)
- Current Rotation
- Trajectory (Landing point calculation: Terrain-Angle, Distance to Landing Point)
- Coins (Distance to Coin, Angle to next Coin)
- Rewards:
- for each 10 meters: Reward of 10 * Speedbonusfactor
- based on fuel state: (1 - CurrentFuel)/10000
- based on collected coins: CollectedCoinsValue/100
- based on Acceleration: if hold for 3 Agent-Frames
- Punishment:
- Driving backward: 0.05 per Physics Frame
- Episode End Conditons:
- Fuel Empty
- Head Hit (Neck broken)
- Underneath Terrain (Error)
- Reaches Goal (9900) -> Reward = 1000
This project is ... Work in Progress!
Changes compared to original: Procedural Terrain Generation with Customizable Fourier-Transformation, Main Menu Changes, Moon Map, ML-Agent added
Project adjusted from original source: https://github.com/choijoohee213/HillClimbRacing
