Project: #93 Incorporating Uncertainty for Reinforcement Learning of Agent Policies Progress Report - Reporting Period Ending: Sept. 30, 2018 Principal Investigator: Daniel Lee Status: Overdue Project Start Date: Jan. 1, 2018 End Date: Dec. 28, 2018 Research Type: Advanced Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2017 Mobility21 UTC Progress Report (Last Updated: Sept. 30, 2018, 12:09 p.m.) % Project Completed to Date: 66 % Grant Award Expended: 0 % Match Expended & Document: 0 USDOT Requirements Accomplishments We have introduced a new Bayesian approach to Reinforcement Learning using off-policy TD methods and Assumed Density Filtering. This allows for updates on action-values (Q) through an online Bayesian inference method that explicitly takes into account uncertainty. Our empirical results show our proposed algorithms outperform comparable algorithms in several task domains, and has been submitted for publication at the AAAI conference. Impacts The current work will enable more efficient training of autonomous systems for intelligent transportation systems. In particular, it will allow for better exploration of policies in decision making processes. This will be critical in systems that need to navigate in uncertain and dynamic environments. Other Open source software available through Penn. Outcomes New Partners N/A Issues N/A