Policy Gradient Methods in Repeated Games Joint work with Domenico Mergoni and Ed Plumb

Thumbnail

Event details

Date 30.10.2025
Hour 11:0012:00
Speaker Dr Galit Ashkenazi-Golan, Assistant Professor at the department of mathematics, LSE, UK
Location
Category Conferences - Seminars
Event Language English
Abstract:
Policy gradient methods have emerged as powerful tools for optimising complex decision-making processes.
These methods utilise gradient ascent to iteratively enhance policies based on observed rewards.
While they frequently achieve local convergence, global convergence guarantees remain elusive outside of specific classes of games.
We illustrate this phenomenon through examples, illuminating the challenges surrounding global optimisation.
Following that, we explore the utilisation of policy gradient methods in repeated games and define the set of policies that are learnable using these methods.
Lastly, we provide a preview of our follow-up presentation, wherein we will present a Folk theorem result for learning in repeated games.

 
Short Bio
Galit Ashkenazi-Golan is an assistant professor at the department of mathematics, London School of Economics and Political Science (LSE), United Kingdom.
Her PhD is from the school of mathematical sciences – Tel-Aviv University, and she spent more than two years being a visiting student at the Ecole Polytechnique, Paris.
Her topics of research are repeated games, stochastic games, opinion dynamics and learning in games.
Recently, she has focused on the game dynamics resulting from multi-agent reinforcement learning.

Practical information

  • General public
  • Free

Organizer

  • Prof. Maryam Kamgarpour

Share