JevMade hello@JevMade.com
← Back to experiments

Benchmarks & research

mario-play

An original platformer and reinforcement-learning lab testing whether frozen Jev risk features help PPO learn faster.

Source screenshot of mario-play
SOURCE SCREENSHOT · source ↗ · captured 2026-09-22Full screenshot ↗

What it does

The reported single-seed pilot found that Jev's frozen risk features did not help PPO reach the target sooner—a useful negative result rather than a claimed win.

Maker-reported (not independently measured by JevMade): 476 Mario and 500 Taxi Jev API requests for the frozen feature tables (author-reported) · Baseline PPO first met the learning target at 600k interactions versus 900k with frequent Jev features and 1.1M with infrequent ones, single seed and single level (author-reported)

Primitives
choice, noul
Platform
Python
Added
Project created
GitHub stars
0 · snapshot 2026-09-22

Source checked 2026-09-22 — opened the primary source directly.