{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/recurrent-predictive-state-policy-networks","title":"Recurrent Predictive State Policy Networks","arxiv_id":"1803.01489","date":"2018-03-05","proceeding":"ICML 2018 7","authors":["Ahmed Hefny","Zita Marinho","Wen Sun","Siddhartha Srinivasa","Geoffrey Gordon"],"abstract":"We introduce Recurrent Predictive State Policy (RPSP) networks, a recurrent\narchitecture that brings insights from predictive state representations to\nreinforcement learning in partially observable environments. Predictive state\npolicy networks consist of a recursive filter, which keeps track of a belief\nabout the state of the environment, and a reactive policy that directly maps\nbeliefs to actions, to maximize the cumulative reward. The recursive filter\nleverages predictive state representations (PSRs) (Rosencrantz and Gordon,\n2004; Sun et al., 2016) by modeling predictive state-- a prediction of the\ndistribution of future observations conditioned on history and future actions.\nThis representation gives rise to a rich class of statistically consistent\nalgorithms (Hefny et al., 2018) to initialize the recursive filter. Predictive\nstate serves as an equivalent representation of a belief state. Therefore, the\npolicy component of the RPSP-network can be purely reactive, simplifying\ntraining while still allowing optimal behaviour. Moreover, we use the PSR\ninterpretation during training as well, by incorporating prediction error in\nthe loss function. The entire network (recursive filter and reactive policy) is\nstill differentiable and can be trained using gradient based methods. We\noptimize our policy using a combination of policy gradient based on rewards\n(Williams, 1992) and gradient descent based on prediction error. We show the\nefficacy of RPSP-networks under partial observability on a set of robotic\ncontrol tasks from OpenAI Gym. We empirically show that RPSP-networks perform\nwell compared with memory-preserving networks such as GRUs, as well as finite\nmemory models, being the overall best performing method.","url_abs":"http://arxiv.org/abs/1803.01489v1","url_pdf":"http://arxiv.org/pdf/1803.01489v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"recurrent-predictive-state-policy-networks","repo_url":"https://github.com/RPSR-DQN/paper-code","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":null},{"paper_slug":"recurrent-predictive-state-policy-networks","repo_url":"https://github.com/ahefnycmu/rpsp","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":null}],"tasks":[{"task_slug":"openai-gym","task_name":"OpenAI Gym"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1803.01489","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}