Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
自主水下导航强化学习中的任务专属子网络发现
[Daily Paper] Task-specific Subnetwork Discovery in Reinforcement Learning for Autonomous Underwater Navigation
AI summary
论文提出用梯度归因在训练好的强化学习策略中发现任务专属子网络,多任务水下导航智能体在无显式模块化架构下自发形成分工不同的神经元子集。该方法可定位引发特定子任务失败的模块,实现定向干预而非整体重训;单个子网络还能隔离测试以验证安全规范,受损时智能体表现为可预测的局部退化而非整体崩溃。
Read originalfailurefirst.org