MechE PhD Prospectus Defense - Xiao Li
- Starts: 11:30 am on Tuesday, October 16, 2018
- Ends: 1:30 pm on Tuesday, October 16, 2018
TITLE: A FORMAL METHODS APPROACH TO INTERPRETABLE AND GENERALIZABLE REINFORCEMENT LEARNING. COMMITTEE: Prof. Calin Belta (ME/SE/ECE) (Advisor) Prof. Roberto Tron (ME/SE) Prof. Rebecca Khurshid (ME/SE) Prof. Sean Andersson (ME/SE) ABSTRACT: Reinforcement learning (RL) has shown potential to learn control policies that are otherwise diﬃcult to manually design. However, current (deep) reinforcement learning techniques suﬀer from a number of drawbacks including low sample eﬃciency, diﬃcult to learn tasks with long horizons, low transferability and explainability. These problems are especially crucial in robotic applications where data is expensive and safety is critical. Thus, learning techniques have yet to be widely adopted by the robotics industry. Formal methods on the other hand puts its focus on semantic rigor, theoretical proofs and guarantees. The (temporal) logic based tools that formal methods provide are comprehensible by humans and more easily transferable among problem domains. Preliminary results suggest that temporal logic is much suited for learning and reasoning symbolically on a high level over long horizons whereas reinforcement learning excels at learning low-level motor controls and perception from raw sensor data. This also intuitively aligns with human intelligence. Given the complementary nature of reinforcement learning and formal methods, the goal of this project is to combine RL with techniques in formal methods to advance state-of-the-art in robotics control and reasoning. We aim to develop an eﬃcient data driven control system that is able to self-improve as it interacts with the environment. The skills that the system acquire should satisfy certain levels of safety requirements. The system is to be able to learn to perform logically complex tasks while forming skill abstractions eﬀective for transferring to other problem domains.
- EPIC B29 Conference Room, 750 Commonwealth Ave