Time-Efficient Reinforcement Learning with Stochastic Stateful Policies
Firas Al-Hafez, Guoping Zhao, Jan Peters, Davide Tateo
摘要
Stateful policies play an important role in reinforcement learning, such as handling partially observable environments, enhancing robustness, or imposing an inductive bias directly into the policy structure. The conventional method for training stateful policies is Backpropagation Through Time (BPTT), which comes with significant drawbacks, such as slow training due to sequential gradient propagation and the occurrence of vanishing or exploding gradients. The gradient is often truncated to address these issues, resulting in a biased policy update. We present a novel approach for training stateful policies by decomposing the latter into a stochastic internal state kernel and a stateless policy, jointly optimized by following the stateful policy gradient. We introduce different versions of the stateful policy gradient theorem, enabling us to easily instantiate stateful variants of popular reinforcement learning and imitation learning algorithms. Furthermore, we provide a theoretical analysis of our new gradient estimator and compare it with BPTT. We evaluate our approach on complex continuous control tasks, e.g., humanoid locomotion, and demonstrate that our gradient estimator scales effectively with task complexity while offering a faster and simpler alternative to BPTT.
问 Lune
几分钟读透这篇 Paper,不用再花一下午。
Lune 能读完全文,也能读引用它的顶会 Paper,讲清它的结论、证据和局限,每一点都注明出自哪篇 Paper。不用硬啃,也不靠猜。
已有 1 篇顶会 Paper 引用了它。别等审稿人问起,先弄清它们发现了什么。
脉络
- Real-Time Recurrent Learning using Trace Units in Reinforcement Learning 引用了 Recurrent Model-Free RL Can Be a Strong Baseline for Many POMDPs
- Recurrent Model-Free RL Can Be a Strong Baseline for Many POMDPs 引用了 Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable Model
引用它的顶会 Paper1
它引用的 Paper5
- Stochastic Latent Actor-Critic: Deep Reinforcement Learning with a Latent Variable ModelAlex X. Lee, Anusha Nagabandi, Pieter Abbeel, Sergey LevineNeurIPS 2020 · 437 次引用
- Supervised Pretraining Can Learn In-Context Reinforcement LearningJonathan Lee, Annie Xie, Aldo Pacchiano, Yash Chandak 等NeurIPS 2023 · 170 次引用
- Recurrent Model-Free RL Can Be a Strong Baseline for Many POMDPsTianwei Ni, Benjamin Eysenbach, Ruslan SalakhutdinovICML 2022 · 162 次引用
- Provable Reinforcement Learning with a Short-Term MemoryYonathan Efroni, Chi Jin, Akshay Krishnamurthy, Sobhan MiryoosefiICML 2022 · 45 次引用
- LS-IQ: Implicit Reward Regularization for Inverse Reinforcement LearningFiras Al-Hafez, Davide Tateo, Oleg Arenz, Guoping Zhao 等ICLR 2023 · 1 次引用
相关 Paper
- Training Recurrent Neural Networks Online by Learning Explicit State VariablesSomjit Nath, Vincent Liu, Alan Chan, Xin Li 等ICLR 2020 · 9 次引用
- Learning Value Functions in Deep Policy Gradients using Residual VarianceYannis Flet-Berliac, Reda Ouhamma, Odalric-Ambrym Maillard, Philippe PreuxICLR 2021 · 16 次引用
- Synthesizing Programmatic Policies that Inductively GeneralizeJeevana Priya Inala, Osbert Bastani, Zenna Tavares, Armando Solar-LezamaICLR 2020 · 54 次引用
- Stabilizing Backpropagation Through Time to Learn Complex PhysicsPatrick Schnell, Nils ThuereyICLR 2024 · 2 次引用
- Reparameterization Flow Policy OptimizationHai Zhong, Zhuoran Li, Xun Wang, Longbo HuangICML 2026
