Project Description
Plan:
1. Design 2 similar racing games with different physics and controls.
2. Implement 2 agents in this game,
3. Transfer between games with minimal retraining to create long-term production-ready agents.
Result:
Both transfers provide convincing results. The transfer of the realistic agent to the arcade
environment reduces training costs and achieves equivalent performance compared to the specialized agent.
The inverse transfer showcases encouraging results with satisfactory initial performance despite a more complex adaptation.
My Tasks
- Analyze current state of the art.
- Design the AI agents.
- Implement Arcade Racing Prototype.
- Train the agent.
- Analyze results.
- Define evaluation metrics.
- Transfer the agent
- Write research paper.
Analyse
We are very satisfied with the results.
We were able to run some transfer tests that we had considered a bonus (e.g., transfers between agents with different inputs).
Project Context
Training an agent is time-consuming and resource-intensive. In particular,
it proves to be especially inefficient in unstable environments, such as video game development.
Problem: How does the transfer of an agent trained via reinforcement learning from one environment to another work,
with the goal of minimizing or eliminating the need for retraining?
Other
An auxiliary paper has been submitted to COG IEEE 2026. Waiting for a response.