Virtual Coffee with an Expert – Reinforcement learning with probabilistic safety guarantees

Speaker: Jana Tumová 

Abstract:

Reinforcement learning (RL) is a powerful approach to decision making under uncertainty. However, it is often difficult to validate the resulting policy with high confidence. In this talk, we overview a recently popular two-step approach to safe reinforcement learning: in the first step, formal verification identifies the set of actions that guarantee safety with desired probability, and the second step integrates this into RL algorithm to constrain the exploration of the agent. Probabilistic guarantees thus get transferred to the trained policy during training and execution. We will discuss our results in problems inspired by urban autonomous driving involving multiple traffic participants, where desired safety properties are expressed in linear temporal logic (LTL). We will demonstrate in simulation of an unsignalized intersection that the resulting policy exhibits strong safety guarantees and scales well.

This event will be held via Zoom. Register for the event below. A link to the meeting room will be sent out to attendees before the event starts.  

Zoom link here: https://kth-se.zoom.us/j/66805342181