让机器人像生物一样快速适应新问题，靠的是动态优化行为搜索。
Life, uh, Finds a Way: Hyperadaptability by Behavioral Search
- 用认知图谱和自修改搜索机制，按简单有效排序连续行为。
- 在复杂迷宫中快速实现强鲁棒导航，且在难题上获得高奖励。
- 适合研究自主学习、机器人技能掌握与异常情况应对的学者。
生物能在极少或仅一次经历的情况下解决各种问题，这种能力称为超适应性（hyperadaptability）。本文提出一种理论，将行为视为自我修改的搜索过程的物理表现。系统不随机探索，而是通过简化与有效性动态排序无限连续行为，行为从认知图谱路径中采样，其顺序由行为执行与图谱修改的紧密反馈回路决定。我们采用赫布学习和一种新型谐波神经表示实现认知图谱，支持灵活信息存储。仿真实验验证了该方法在复杂迷宫中迅速获得强鲁棒导航能力，并在经典强化学习难题的困难扩展上取得高奖励。该框架为发展性学习提供新理论模型，推动机器人自主掌握复杂技能并应对异常情境。
原文摘要 · Abstract (English)
Living beings are able to solve a wide variety of problems that they encounter rarely or only once. Without the benefit of extensive and repeated experience with these problems, they can solve them in an ad-hoc manner. We call this capacity to always find a solution to a physically solvable problem $hyperadaptability$. To explain how hyperadaptability can be achieved, we propose a theory that frames behavior as the physical manifestation of a self-modifying search procedure. Rather than exploring randomly, our system achieves robust problem-solving by dynamically ordering an infinite set of continuous behaviors according to simplicity and effectiveness. Behaviors are sampled from paths over cognitive graphs, their order determined by a tight behavior-execution/graph-modification feedback loop. We implement cognitive graphs using Hebbian-learning and a novel harmonic neural representation supporting flexible information storage. We validate our approach through simulation experiments showing rapid achievement of highly-robust navigation ability in complex mazes, as well as high reward on difficult extensions of classic reinforcement learning problems. This framework offers a new theoretical model for developmental learning and paves the way for robots that can autonomously master complex skills and handle exceptional circumstances.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。