The next D5 seminar will be held on Thursday, October 7, 2021, at 11am in LORIA, Amphi C. It is open to everyone interested!
Our invited speaker will be Joerg Hoffmann, Professor of CS at Saarland University, Saarbrücken (and former Directeur de Recherche at INRIA Nancy).
Action policies in sequential decision making map states (or state sequences) to actions. Learned action policies, in particular neural action policies, are gaining more and more traction in dealing with complex environments under real-time action-decision constraints. But how to ensure trust in such a policy? Within the broad movement in AI towards addressing this question (and many related ones), my research group in Saarbrücken explores approaches towards explanation (elucidating trade-offs between different user objectives), verification (proving that the policy never enters an unsafe state), testing (trying to find cases where the policy is unsafe or otherwise undesirable), and visualization (enabling human users to inspect policy behavior). Our research on all these topics is in the early stages; I will outline ideas and first results, and I’m very interested in collaborations. The talk thus is intended to outline our agenda and start discussions about possible collaborations between my group and LORIA researchers.
Joerg Hoffmann has been a Directeur de Recherche at INRIA Nancy from
2009–2012, and is now a Professor of CS at Saarland University,
Saarbrücken. He is a Fellow of AAAI and EurAI. His core research area is
AI automated planning, with connections to related fields such as
verification, and a recent focus on explainable and trusted AI through
combinations of neural and symbolic methods.