Google DeepMind has developed a new technology called AlphaEvolve, which enables a large language model (LLM) to automatically improve game theory algorithms. This method has demonstrated performance surpassing that of human experts.

AI explained

How does AlphaEvolve improve game theory algorithms?

AlphaEvolve is a large language model developed by Google DeepMind that automatically mutates and improves Multi-Agent Reinforcement Learning algorithms. It has created new algorithm variants that outperform human-designed methods in game theory tasks.

Summary: AlphaEvolve evolves the source code of game theory algorithms like CFR and PSRO, discovering more efficient update rules without manual tuning.
Why it matters: It surpasses expert-designed algorithms and can accelerate AI development in areas requiring strategic multi-agent decision-making.
Key point: AlphaEvolve found novel mechanisms such as volatility-adaptive updates and improved exploration-exploitation transitions, tested successfully on multiple games.

Illustration of Google DeepMind Develops LLM That Can Enhance Game Theory Algorithms — AI-generated illustration

AlphaEvolve Automates Algorithm Design in Game Theory

AlphaEvolve is an evolutionary code agent that leverages LLM technology to replace manual iteration in developing algorithms for Multi-Agent Reinforcement Learning (MARL). Previously, researchers relied on intuition and trial-and-error to create algorithms such as Counterfactual Regret Minimization (CFR) and Policy Space Response Oracles (PSRO). Instead of tuning hyperparameters, AlphaEvolve actually mutates the source code of MARL algorithms, discovering new update rules that are more efficient than existing methods. In experiments, it found algorithms that outperformed the best hand-designed alternatives, using the OpenSpiel framework for testing.

The research shows that AlphaEvolve can uncover non-intuitive mechanisms, such as a hard restart at iteration 500 and asymmetric boosting of positive regret. The new variant of CFR, called Volatility-Adaptive Discounted CFR (VAD-CFR), adapts to volatility in the learning process, making it more responsive in dynamic game environments. AlphaEvolve also developed a new variant of PSRO, named Smoothed Hybrid Optimistic Regret PSRO (SHOR-PSRO), which improves the transition between exploration and exploitation. Both algorithms were tested across various games and proved competitive against existing methods.

Implications for AI Development in the U.S. Market

AlphaEvolve’s ability to automate the design of game theory algorithms could accelerate AI innovation in the U.S., particularly in sectors like gaming, simulations, and strategic decision-making systems. By reducing the need for extensive manual tuning and expert intervention, American developers and companies can more rapidly deploy advanced AI solutions that adapt to complex multi-agent environments.

Source: Marktechpost

⚡ Tool recommendation

Automate AI workflows without coding → Try free