Discovering Multiagent Learning Algorithms with Large Language Models
An AI agent evolved two novel algorithms that outperform state-of-the-art baselines in game theory.
Researchers from DeepMind and collaborators built AlphaEvolve, an evolutionary coding agent powered by large language models, to automate the discovery of multi-agent reinforcement learning (MARL) algorithms. It evolved two new algorithms: VAD-CFR for iterative regret minimization and SHOR-PSRO for population-based training. These algorithms introduce non-intuitive mechanisms like volatility-sensitive discounting and hybrid meta-solvers, outperforming established baselines like Discounted Predictive CFR+ in imperfect-information games.
Why It Matters
This automates a core, human-intensive research process, potentially accelerating breakthroughs in game theory, economics, and multi-agent AI systems.