Mittagsseminar (in cooperation with A. Steger, D. Steurer and B. Sudakov)

Date and Time: Thursday, April 18, 2024, 12:15 pm

Duration: 30 minutes

Location: CAB G51

Speaker: Cesare Carissimo

How Can We Better Understand Complex Learning Dynamics - The Case of Q-Learning and Braess's Paradox

Economic analysis can be made more complete through the lens of Learning Dynamics. Indeed, it is shown in recent work that Nash Equilibria are incomplete descriptors of game dynamics (Milionis, Papadimitriou, Piliouras, Spendlove, 2023). In this talk we will explore some notable cases where learning dynamics driven by Q-Learning deviate "greatly" from the Equilibrium Behaviour, and in such a manner that it is beneficial for the system welfare. The purpose of this talk is to explore and discover with you which questions to ask when faced with these kinds of dynamics which are interesting in their complexity. I will first present the base model, Q-learners in the Braess Paradox, and explain the ways in which we can create their complex and coordinated behaviour. I will then present an extension of this model with which we can begin to imagine Learning Dynamic Control, in the control theoretic sense of influencing the dynamical behaviour of the multi-agent learning system towards a target behaviour. This work is thus far empirical. I set forth the following question to you: what theoretical questions do you think we can purposefully ask and answer?

Upcoming talks     |     All previous talks     |     Talks by speaker     |     Upcoming talks in iCal format (beta version!)

Previous talks by year:   2024  2023  2022  2021  2020  2019  2018  2017  2016  2015  2014  2013  2012  2011  2010  2009  2008  2007  2006  2005  2004  2003  2002  2001  2000  1999  1998  1997  1996  

Information for students and suggested topics for student talks

Automatic MiSe System Software Version 1.4803M   |   admin login