SKM 2023 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 16: Collective Dynamics in Animal and Human Societies
SOE 16.3: Talk
Thursday, March 30, 2023, 11:30–11:45, ZEU 260
Influence of confirmation biases on collective decision-making in fluctuating environments — •Clémence Bergerot1,2, Wolfram Barfuss3, and Pawel Romanczuk2,4 — 1Charité - Universitätsmedizin Berlin, Einstein Center for Neurosciences Berlin, Germany — 2Humboldt Universität zu Berlin, Germany — 3Tübingen AI Center, University of Tübingen, Germany — 4Research Cluster of Excellence "Science of Intelligence", Berlin, Germany
In experimental studies of decision-making, it is now established that human agents tend to update confirmatory information with a higher weight than disconfirmatory information. This confirmation bias has been modeled within a reinforcement learning framework, using asymmetric updating of prediction errors. Interestingly, such a bias has been suggested to enhance individual performance in a wide range of multi-armed bandit tasks. However, little is known about the impact of the confirmation bias on collective performance. In order to characterize the circumstances that make this bias beneficial or detrimental to collective decision-making, we develop a multiagent model in which reinforcement learning agents can observe others' actions and rewards, and update this information asymmetrically. With agent-based simulations, we seek to understand how the confirmation bias affects collective performance in changing environments, and how network topology modulates this effect. We also study our multiagent system in the deterministic limit [W. Barfuss et al., Phys. Rev. E, 99(4) (2016) 043305], which allows us to gain an analytical understanding of the biased learning dynamics.