BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Understanding the Behaviour of Inverse Reinforcement Learning Agen
 ts
DTSTART:20180829T111500
DTEND:20180829T131500
DTSTAMP:20260916T055257Z
UID:34604b0b09cce6e30b932b1366e10ddab9738c02e8d8562a2461db9e
CATEGORIES:Conferences - Seminars
DESCRIPTION:Teresa Yeo\nEDIC candidacy exam\nExam president: Prof. Boi Fal
 tings\nThesis advisor: Prof. Volkan Cevher\nThesis co-advisor: Prof. Pierr
 e Dillenbourg\nCo-examiner: Dr. Mathieu Salzmann\n\nAbstract\nProgramming 
 autonomous agents in a Markov decision\nprocess setting typically requires
  designing a reward function.\nThis is a challenging problem in many areas
  that do not have\na well-defined score\, such as control\, locomotion and
  navigation\ntasks\, among many others. In inverse reinforcement learning\
 n(IRL)\, the agent learns this function from expert demonstrations.\nNumer
 ous IRL methods have been developed\, each with their\nown strengths and w
 eakness. However\, a less studied area\, is on\nunderstanding such a model
 ’s behavior. We would like models\nthat not only perform well but are al
 so explainable as it is\nessential for establishing trust in a system or f
 or debugging.\nOur goal is to be able to explain why an IRL agent behaves 
 a\ncertain way\, by identifying which of the expert’s trajectory was\nmo
 st responsible for that behavior. As the method used has close\nconnection
 s to generating adversarial attacks\, we also discuss\nhow this can be app
 lied to IRL.\n\nBackground papers\nApprenticeship Learning via Inverse Rei
 nforcement Learning\, by Pieter Abbel and Andrew Ng [ICML04] \nModel-fre
 e Imitation Learning with Policy Optimization\, by Jonathan Ho\, Jayesh Gu
 pta and Stefano Ermon [ICML16]\nUnderstanding Black-box Predictions via I
 nfluence Functions\, by Pang Wei Koh and Percy Liang [ICML17]\n 
LOCATION:RLC D1 661 https://plan.epfl.ch/?room==RLC%20D1%20661
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR