google-deepmind/meltingpot

What it solves

Melting Pot provides a standardized benchmark for multi-agent reinforcement learning (MARL). It addresses the challenge of evaluating whether AI agents can generalize their social behaviors—such as cooperation, competition, and trust—to novel situations and unfamiliar individuals they did not encounter during training.

How it works

The project offers a suite of multi-agent games called "substrates" (over 50 available) for training agents. To test generalization, it provides over 256 unique test scenarios. By evaluating agents on these held-out scenarios, researchers can quantify how well agents interact with unfamiliar partners and perform in interdependent social situations, allowing for the ranking of different MARL algorithms.

Who it’s for

It is designed for researchers working on multi-agent reinforcement learning and social AI who need a rigorous way to test the robustness and social generalization of their agents.

Highlights

• Over 50 training substrates (multi-agent games).
• Over 256 unique test scenarios for evaluation.
• Tests a wide range of social interactions including deception, reciprocation, and stubbornness.
• Built on top of DeepMind Lab2D.