Please note: This PhD seminar will take place in E7 5419 and online.
Frédéric Bouchard, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Krzysztof Czarnecki
Explainable decision making is desirable to verify the safety, reliability and legal conformance of autonomous vehicles. Explicit rules are typically more interpretable than approaches that encode knowledge implicitly, and we have already deployed a rule-based behaviour planner in an autonomous car prototype that drives on public roads.
In this work, we formalize a procedure by which unmet requirements can be used to construct a rule-based behaviour planner that makes decisions with understandable causality. Starting with a set of requirements and a precise notion of their violation, we show how an expert can efficiently learn a driving policy from counterexample situations. During operation, these situations provide an explanation of why a rule applies. We demonstrate our computer-aided learning procedure using the Carla Autonomous Driving Leaderboard, which provides a set of requirements for a realistic operational design domain.
To join this PhD seminar in person, please go to E7 5419. You can also attend virtually on MS Teams.