强化学习 颠覆级 有讲解视频
发表时间
2026-08-22
DOI
10.1038/s41467-026-75703-4

收录解读

长时程自主探索中的贝叶斯优化通常随观测数量增长而付出三次时间和二次内存代价,使有限硬件难以持续运行。

SSP-BO 用可组合的高维向量符号表示统一编码连续、离散、图和轨迹动作空间,把样本选择转化为紧致嵌入域上的连续优化,使样本选择对既往观测数保持固定资源开销。

21 页全文显示实验覆盖标准连续函数、pest control、RNA inverse folding、NAS-Bench-101、机器人 Box Reach/Belt Stretch 和多机器人月坑探索;相对标准 BO,规划总时间最高降低约 205 倍,并在 SpiNNaker 与 Loihi 2 上实现 30–188 倍能耗优势。

论文公开性能/时序数据以及 SSP-BO、机器人实验代码,形成可复用原语;但固定嵌入维度会带来表示保真度权衡,概念漂移和在线参数自适应仍未解决,因此维持 disruptive 而非 paradigm。

原始摘要

Autonomous systems that learn and explore over long horizons face a problem. Standard methods scale poorly in the number of observations, n, precluding sustained operation on bounded hardware. We show that compositional, high-dimensional vector representations inspired by neural computation address these constraints. We use these representations to construct a Bayesian optimization algorithm that operates in complex spaces and reduces time and memory requirements compared with state-of-the-art algorithms on diverse tasks.

解读视频

链接