Scholarlink

Effectively leveraging action priors from diverse tasks in reinforcement learning remains a significant challenge, particularly in task-agnostic datasets where capturing task relevance is crucial. This work addresses these challenges by introducing a modified Set Transformer, cosine similarity, and an existing importance network to derive suggested weights for each implicit action prior based on a normalizing flow architecture. These suggestions are then dynamically combined using Multi-Head Attention enhanced by Gated Recurrent Units (GRUs) together with residual networks, generating the final weights for each normalizing flow. This approach not only enhances task relevance and robustness against input order sensitivity but also improves few-shot adaptation by ensuring the selected priors are both sparse and diverse through weight regularizations. Our experiments across four challenging environments show the superiority of this approach, with a 37.96% improvement in total performance, reducing the dependency on large-scale task-agnostic datasets, and enhancing resilience to sequence variations and noise.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Combining Explicit Priors and Set Attention Driven Implicit Priors for Demonstration-Guided Reinforcement Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Shujiong Tang,
  • Irwin King,
  • Zenglin Xu

摘要

Effectively leveraging action priors from diverse tasks in reinforcement learning remains a significant challenge, particularly in task-agnostic datasets where capturing task relevance is crucial. This work addresses these challenges by introducing a modified Set Transformer, cosine similarity, and an existing importance network to derive suggested weights for each implicit action prior based on a normalizing flow architecture. These suggestions are then dynamically combined using Multi-Head Attention enhanced by Gated Recurrent Units (GRUs) together with residual networks, generating the final weights for each normalizing flow. This approach not only enhances task relevance and robustness against input order sensitivity but also improves few-shot adaptation by ensuring the selected priors are both sparse and diverse through weight regularizations. Our experiments across four challenging environments show the superiority of this approach, with a 37.96% improvement in total performance, reducing the dependency on large-scale task-agnostic datasets, and enhancing resilience to sequence variations and noise.