The next Shocklab Seminar will be online on Wednesday, 8 October 2025, at 16:00 PM. Baraah Sidahmed will be presenting "Game-Aware Optimization for Multi-Agent Reinforcement Learning". Please come along if you can, it’s sure to be an engaging talk!

Title: Game-Aware Optimization for Multi-Agent Reinforcement Learning
Speaker: Baraah Sidahmed
Date: Wednesday, 8 October 2025
Time: 16:00-17:00 (GMT +2)
Zoom Meeting Link: https://uct-za.zoom.us/j/92750361177?pwd=QzNiRzBJRjRITVlwa2k5SVNkVmx5UT09
 

Abstract: Training multiple competing RL agents is fundamentally an equilibrium-seeking problem not a pure minimization task - standard updates often oscillate instead of converging. This talk presents a game-aware formulation for actor–critic and multi-agent RL using variational inequalities (VIs), a framework for computing equilibria in interacting systems. I will explain why gradient dynamics can cycle and show how simple, plug-in optimization methods - such as lookahead (and related extragradient variants) - stabilize learning in competitive settings along with a practical guidance for choosing lookahead hyperparameters. Classical exact benchmarks from multi-agent environments (such as rock–paper–scissors and predator–prey) demonstrate that VI-based training yields steadier dynamics, faster convergence, and more reliable policies.

Bio: A PhD candidate at the relational ML group at the CISPA Helmholtz Center for Information Security. Previously worked on optimizing multi-agent reinforcement learning using ideas from game theory. Currently working on a general framework that enables a wide range of agents - from simple nodes to complex models - to interact efficiently.

See you there!

 

Housekeeping: