Machine Learning Transfer Research: Realistic Prototype

Realistic prototype for the research project: Evaluating transfer learning for reinforcement learning agents across same-genre video games
Research project carried out in partnership between PulluP Entertainment and ISART Digital Paris

Table of Content

• Presentation
• Setup
• Technology
• Credits

Presentation

This project was developed in Unity by a team of four to prototype a realistic racing game in order to train on it, an agent with machine learning and transfer it to the other prototype we made, an Arcade prototype.

Setup

Warning: Make sure to have installed the python environment

Start a training session

In Anaconda Prompt or MiniForge Prompt

1. Go to the project folder
2. Activate the python environment

conda activate mlagents

3. Start training

mlagents-learn config/AgentRealisticBehavior_config.yaml --run-id=Training_Name

In Unity

1. Open the scene Assets/Scenes/Training
2. Launch the scene

Test trained agents

In Unity

1. Open the scene Assets/Scenes/MainCircuit
2. Get an .onnx file from the folder Result in the Assets/Results
3. In the Hierarchy Car Pickup GameObject
4. In the Inspector Behavior Parameters Script
5. Change the parameters Model with the neural network you want to test. Make sure that Behavior Type is set to Inference Only
6. Launch the scene

Play agent

---

1. Open either the scene MainCircuit or Training
2. On the car pickup, in the component Behavior Parameters
   • Change the Behavior Type to Heuristic Only

Technology

• Unity 6 v6000.3.9f1
• ML Agents Plugin (version 4.0.2)

Credits

• Bryan BACHELET
• Vincent DEVINE
• Matéo ERBISTI
• Omaya LISE
• Aurelien CHAMBON
• Aurélien LHERBIER

Assets

• Vehicle Physics Pro
• Dreamteck Spline by Dreamteck