{"type":"ScholarlyArticle","title":"Reconciling Rewards with Predictive State Representations","authors":[{"name":"Andrea Baisero","slug":"andrea-baisero","given":"Andrea","family":"Baisero","affiliation":""},{"name":"Christopher Amato","slug":"christopher-amato","given":"Christopher","family":"Amato","affiliation":""}],"abstract":"Predictive state representations (PSRs) are models of controlled non-Markov\r\n  observation sequences which exhibit the same generative process governing\r\n  POMDP observations without relying on an underlying latent state.  In that\r\n  respect, a PSR is indistinguishable from the corresponding POMDP.  However,\r\n  PSRs notoriously ignore the notion of rewards, which undermines the general\r\n  utility of PSR models for control, planning, or reinforcement learning.\r\n  Therefore, we describe a sufficient and necessary accuracy condition\r\n  which determines whether a PSR is able to accurately model POMDP rewards, we\r\n  show that rewards can be approximated even when the accuracy condition is not\r\n  satisfied, and we find that a non-trivial number of POMDPs taken from a\r\n  well-known third-party repository do not satisfy the accuracy condition.  \r\n  We propose reward-predictive state representations (R-PSRs), a\r\n  generalization of PSRs which accurately models both observations and rewards,\r\n  and develop value iteration for R-PSRs.  We show that there is a mismatch\r\n  between optimal POMDP policies and the optimal PSR policies derived from\r\n  approximate rewards.  On the other hand, optimal R-PSR policies perfectly\r\n  match optimal POMDP policies, reconfirming R-PSRs as accurate state-less\r\n  generative models of observations and rewards.","doi":"10.24963/ijcai.2021/299","url":"https://ijcai.org/proceedings/2021/299","pdf":"https://ijcai.org/proceedings/2021/0299.pdf","markdown":"https://ijcai.org/proceedings/2021/299.md","bibtex":"https://ijcai.org/proceedings/2021/bibtex/299","volume":{"slug":"2021","title":"Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence","short_title":"IJCAI 2021","url":"https://ijcai.org/proceedings/2021/","year":2021,"location":"Montreal","dates":"19-27 August 2021","editors":"Prof. Zhi-Hua Zhou","isbn":"978-0-9992411-9-6","publisher":"International Joint Conferences on Artificial Intelligence Organization","doi_prefix":"10.24963/ijcai.2021"},"section":"Machine Learning","track":"Main Track","number":299,"pages":{"first":2170,"last":2176},"keywords":["Machine Learning: Reinforcement Learning","Uncertainty in AI: Uncertainty Representations","Planning and Scheduling: POMDPs"],"videos":[],"references":[],"license_note":"Copyright International Joint Conferences on Artificial Intelligence."}