AlphaEvolve再进化，DeepMind用A“养殖”算法，碾压所有人类设计

Core Insights - DeepMind's latest paper introduces AlphaEvolve, which treats algorithm source code as a genome and uses Gemini as a genetic operator to evolve new game theory algorithms through a process akin to natural selection [1][5][20] - The evolved algorithms outperform human-designed optimal solutions in various tests, utilizing mechanisms that human researchers had not previously considered [1][22] Group 1: AlphaEvolve and Its Mechanism - AlphaEvolve is described as an evolutionary coding agent that uses the source code as a genome, with LLM acting as a genetic operator to mutate the code [5][20] - The process involves evaluating the fitness of each "offspring algorithm" based on its exploitability in a set of benchmark games, allowing the best-performing algorithms to survive and evolve further [5][6][20] Group 2: Target Algorithms and Historical Context - The focus of AlphaEvolve is on two core algorithm families in multi-agent reinforcement learning: Counterfactual Regret Minimization (CFR) and Policy Space Response Oracles (PSRO) [6][7] - Historically, researchers have manually tuned and designed variations of these algorithms, but AlphaEvolve automates this process, significantly expanding the search space for potential solutions [10][16][20] Group 3: Implications of AI-Driven Algorithm Design - The paper highlights that the design of game theory algorithms has traditionally been a challenging task due to the complexity of incomplete information games [12][13] - AlphaEvolve's approach allows for meaningful mutations in code, leading to the discovery of effective strategies that human experts had not conceived [17][25] - The results indicate a paradigm shift where AI not only executes algorithms but also invents them, achieving superior performance compared to human-designed methods [22][25] Group 4: Future Directions - DeepMind plans to apply this evolutionary framework to the complete design of deep reinforcement learning agents and explore mechanism discovery in cooperative games [25]