Jev-Like Model Learns to Cook
An experiment in which a decision model learns cooperative cooking using only basic game controls
SITE CAPTUREBased on information published on the service's website. Selection and review policy
WHAT IT DOES
What does it do?
OpenJev takes a text description of the situation and possible actions and calculates a probability for each action. In the experiment, 2 agents each chose 1 of 6 basic controls based on their own observations, and served 6 soups over 512 ticks in the recorded evaluation. This result came from 1 seed and layout; they did not complete a soup when moved to a new layout.
KEY FEATURES
Key features
- Calculates probabilities for each action from a situation description and candidate actions.
- Each agent chooses 1 of 6 basic controls from its own observation.
- Learns from game observations, actions, and rewards without demonstration data.
- 2 agents with shared weights independently cook and deliver.
EDITOR'S NOTE
Why did it stand out?
The experiment lets 2 models each choose 1 button from their own observations without assigning cooking steps or roles in advance. Their alternating roles are interesting, but the lack of success on another kitchen layout is a clear limitation.