arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1602.02523stat.MLcs.LGcs.SYeess.SY
收藏

连续状态POMDPs中的数据高效强化学习

Data-Efficient Reinforcement Learning in Continuous-State POMDPs

University of Cambridge(剑桥大学)

Rowan McAllister, Carl Edward Rasmussen

AI总结:

本文提出一种抗观测噪声的数据高效强化学习算法,通过扩展PILCO算法至POMDPs,利用过滤过程提升策略评估性能,实现在Cartpole摆动任务中更优的非线性控制效果。

摘要:

我们提出了一种抗观测噪声的数据高效强化学习算法。我们的方法通过在策略评估过程中考虑过滤过程,将高度数据高效的PILCO算法(Deisenroth & Rasmussen, 2011)扩展至部分观测马尔可夫决策过程(POMDPs)。PILCO进行策略搜索,通过首先预测可能系统轨迹的解析分布来评估每个策略。我们还预测轨迹相对于过滤过程,从而在结合过滤器与由原始(未过滤)框架优化的策略时实现了显著更高的性能。我们的测试设置是带有传感器噪声的Cartpole摆动任务,该任务涉及非线性动态并需要非线性控制。

英文摘要:

We present a data-efficient reinforcement learning algorithm resistant to observation noise. Our method extends the highly data-efficient PILCO algorithm (Deisenroth & Rasmussen, 2011) into partially observed Markov decision processes (POMDPs) by considering the filtering process during policy evaluation. PILCO conducts policy search, evaluating each policy by first predicting an analytic distribution of possible system trajectories. We additionally predict trajectories w.r.t. a filtering process, achieving significantly higher performance than combining a filter with a policy optimised by the original (unfiltered) framework. Our test setup is the cartpole swing-up task with sensor noise, which involves nonlinear dynamics and requires nonlinear control.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑