BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Proximal Policy Optimization in the Fisher-Rao geometry - Razvan-A
 ndrei Lascu (RIKEN)
DTSTART:20251111T163000Z
DTEND:20251111T171000Z
UID:TALK238552@talks.cam.ac.uk
DESCRIPTION:PPO is one of the most widely used algorithms in reinforcement
  learning\, offering a practical policy gradient method with strong empiri
 cal performance. However\, despite its popularity\, PPO lacks rigorous the
 oretical guarantees for policy improvement and convergence. The method emp
 loys a clipped surrogate objective\, derived from linearising the value fu
 nction in a flat geometric setting. In this talk\, we introduce a refined 
 surrogate objective based on the Fisher&ndash\;Rao geometry\, leading to a
  new variant\, Fisher&ndash\;Rao PPO (FR-PPO). Our approach provides robus
 t theoretical guarantees\, including monotonic policy improvement and sub-
 linear convergence rates\, representing a substantial advance toward forma
 l convergence results for the wider class of PPO algorithms. This talk is 
 based on joint work with David Siska and Lukasz Szpruch.
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR