LesenForschungRadarAnlageframework
Anmelden / Registrieren
← Zurück zur Lektüre
NewsThema · 进化算法价值重估Wesentlich2 Min. LesezeitDisconfirmAI Redaktion

MIT PhD Student Akarsh Kumar Argues Evolutionary Search Is Not Random Fallback

In a Machine Learning Street Talk podcast episode description, MIT PhD student Akarsh Kumar argues against the common AI perception of evolutionary algorithms as inefficient random search fallbacks. He posits that selection preserves partial solutions, meaning mutations only need to be useful about 1% of the time for progress to continue. Kumar describes his ASAL method, which runs simulations and asks a foundation model what happened, discovering an island of open-endedness among 262,144 Life-like rules.

2026-10-10 13:48 GMT+8

AI research often treats evolutionary algorithms as an inefficient fallback, used only as random search when gradients cannot be computed. Akarsh Kumar, a PhD student at MIT and researcher at Sakana AI, refutes this view in a Machine Learning Street Talk podcast episode description. He argues that selection preserves partial solutions, meaning mutations only need to produce useful improvements about 1% of the time for the entire search process to make continuous progress. This mechanism repositions evolutionary algorithms as a primary tool for complex, non-differentiable problems.

Kumar is a PhD student working with Phillip Isola at MIT and the first author of the Fractured Entangled Representation paper, collaborating with Kenneth Stanley, Jeff Clune, and Joel Lehman. In the episode, he explained to host Tim Scarfe that the path a learner takes may shape the structure of what it learns, offering a view of intelligence distinct from the statistical one.

To illustrate the mechanism concretely, the episode description introduces Kumar's lead work on ASAL (Artificial Life via Simulation and Language). Traditional approaches attempt to predict rule behavior, whereas ASAL directly runs simulations and then asks a foundation model what happened. The team applied the method to map all 262,144 Life-like rules and found that the most interesting and open-ended worlds are not uniformly distributed but concentrated on a small island.

Anmelden / Registrieren
LesenForschungRadarAnlageframework

The above content comes from the podcast episode description rather than a full transcript. The specific derivation of the 1% figure, complete ASAL methodology details, and specific metrics for the island finding are not expanded in the readable material. This is an attributed explanation; the core mechanism (selection preserving partial solutions) has broad consensus in evolutionary computation, but specific numerical claims and experimental conclusions require independent verification against the original paper (arXiv:2412.17799).

Quellen

  • podcasters.spotify.com↗
Thema

进化算法价值重估

【背景】长期以来,进化算法(EA)常被AI领域视为梯度下降之外的“随机搜索备胎”,主要用于超参数调优等离散或不可微场景。然而,近期研究正在重塑这一认知:一方面,以MIT博士生Akarsh Kumar为代表的学者通过ASAL等项目证明,结合基…

4 Schritte1 BeiträgeAktualisiert 2026-10-13 03:29

Wie es dazu kam

  1. 2026-10-13MIT博士生Akarsh Kumar在播客中反驳EA是随机搜索备胎的观点,介绍ASAL方法利用基础模型在26万类规则中定位开放性世界。podcasters.spotify.com ↗
  2. 2024-12-01Kumar等人发布ASAL论文,首次成功将视觉语言基础模型应用于人工生命搜索,发现了新的Lenia和Boids生命形式。arxiv.org ↗
  3. 2025-12-20Philosophy & Technology期刊发表文章,主张将进化计算视为“算法隐喻”,强调其在公共话语中应避免过度拟人化或生物化。link.springer.com ↗

Was noch ungeklärt ist

  • 基础模型与进化算法的结合(如ASAL)是否揭示了EA在连续高维空间中的新潜力,还是仅依赖于FM的先验知识?
  • 在理论计算机科学中,如何更精确地界定“支配关系”失效的边界,以便指导新一代EA的设计?
  • 将EA定义为“算法隐喻”是否会削弱公众对其实际技术能力的理解,或者有助于建立更负责任的AI叙事?
Vollständiges Thema lesen →
Recherche fortsetzen →