Jev as a reinforcement-learning policy

By Yi Ma and colleagues

This paper tests whether Jev’s advice can help another AI learn games with less practice.

Bookmark Keep this in your collection.

Leave a noteWhat would you try with this?

Only you can see your notes.

Source screenshot of Jev as a reinforcement-learning policy — SOURCE SCREENSHOTFull screenshot ↗

What it does

For nine MiniGrid tasks and three Atari games, Jev helps choose moves, guide exploration and rate previous attempts. The authors report gains within fixed interaction budgets, not reliable control of a game by Jev alone. Jev stays untrained, with numerical value calculations still used. The OpenReview submission is under review; the public arXiv version names its authors.

How you can use it

Keep the game player learning from what succeeds and what fails. It can consult Jev to pick moves for practice or decide which past attempts to revisit. Jev’s choices guide that learning without changing Jev itself. This is help for a learning game player, rather than a method for Jev to play alone.