Scaling PCG training levels improves Mario PPO performance on unseen levels
What happened
Shinichi Arimura's paper trains PPO agents on progressively scaled, procedurally generated Super Mario Bros. levels and tests their generalization to unseen levels with 1,000 held-out test levels. The results show that increasing the number of non-repeating PCG training levels generally improves agent performance on unseen levels. The same study also fine-tunes the policy with DRAIL, which the author says yields more human-like gameplay.
AI-written from the coverage · updated 9 d ago
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Game AITraining PPO on more PCG levels and fine-tuning with DRAIL yields more human-like Super Mario Bros. play
The study trains PPO agents on Super Mario Bros. levels produced by PCG at increasing scale, and evaluates generalization to unseen levels on 1000 held-out test levels. Results show that increasing the number of non-repeating PCG training levels generally improves performance on unseen levels.