jev-rl
Reinforcement-learning lab where cached Jev judgments act as reward signals for DQN agents on four classic control and grid games.
Overview
JevRL is a runnable reward-model laboratory: DQN agents learn CartPole, MountainCar, Acrobot and FrozenLake using rewards derived from Jev's judgments, compared with other reward conditions. The README reports 36 training runs over 3 seeds and 3 reward conditions, 3.24 million environment steps, and 129 Jev judgments cached and reused across 12 Jev training runs, with a recorded Jev API cost of about $0.0024 to build the score cache (training compute excluded).
A checkpoint gallery shows each game's policy at successive training stages under the Jev reward, from untrained to final. Experiment records and cost records are published, and a website and blog post explain the approach. It is MIT-licensed with a Chinese README.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How good is this game state or behaviour (used as reward)?(inferred) | — |
How it works
- Ask Jev to judge game states and cache the scores
- Use cached scores as reward for DQN training
- Train across seeds and reward conditions
- Record checkpoints and replays
Strengths
- Extremely low judgment cost through caching
- Reproducible experiment records
- Visual replays of learning progress
Caveats
- Classic toy environments only
- Reward caching limits judgments to seen states
- Compares only a few reward conditions
Getting started
See the Run it section of the README.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- RL training scripts with web viewer
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
JEV Reinforcement Learning: four classic games trained with JEV-powered rewards, reproducible experiments and checkpoint replays.
logicrw/awesome-jev-projects
Frequently asked questions
How is Jev used?
Its judgments are cached and used as reward signals for DQN agents.
How much did Jev cost?
About $0.0024 for 129 cached judgments, per the README.
Which games?
CartPole, MountainCar, Acrobot and FrozenLake.
Related projects
Real-time bot battles where you write a fighter's strategy in plain English and a System One model pilots it several times a second.
ProjectbetaTypeScriptBot that plays Red Alert 2 web skirmishes, with code handling mechanics and a decision model making stance, threat and timing calls.
ServicebetaPythonMultiplayer 3D science-fantasy version of Houston's River Oaks District with Jev-powered NPC companion, chauffeur and smart drive.
ProjectbetaTypeScriptBrowser 3D planetary rover sandbox where a TypeSafe controller chooses the next action from complete candidates, versus a baseline.
ProjectbetaTypeScriptBilingual turtle-soup (lateral-thinking) puzzle site where Jev acts as host, judging player questions and grading solutions.
AppbetaTypeScriptDoom runs as a spacecraft payload over real flight software; Jev pilots from the ground every half second while Claude reviews.
ProjectprototypePython