Berlin 2018 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

SOE: Fachverband Physik sozio-ökonomischer Systeme

SOE 17: Social Systems, Opinion and Group Dynamics II

SOE 17.2: Vortrag

Mittwoch, 14. März 2018, 18:30–18:45, MA 001

A Comprehensive Analysis of Reaction to Disturbances through Social Learning in Networks — •Takuro Yamazaki and Hirotada Ohashi — University of Tokyo, Department of Systems Innovation, Tokyo, Japan

There are many social tasks involving decision making of multiple people who have their own interests, e.g. election, traffic and investment. In such situations, dilemmas between social and individual interests often occur, and the mechanism and dynamics of dilemmas are widely studied in evolutionary game frameworks. In these frameworks, agents behave according to predetermined rules, however, in real-world setting, people learn optimal behavior from their own experience and interactions with neighbors. We study repeated matrix games played by many agents with reinforcement learning, which corresponds to social learning frameworks. Each agent interacts with all neighboring agents in each round and learns his strategy from his and neighbor's payoffs. We analyze the learning processes of agents by changing learning parameters and connection structures among agents. We have particular interest in how reinforcement learning agents react to disturbances to conditions. We incorporate changes of payoffs during the learning process and observe the variation of agent's actions. Furthermore, we investigate the effect of network topology among agents. Simulation results show that the speed of reaction to disturbances changes with network topology and reinforcement learning parameters. And sharing information by sharing q value among neighbors and observing neighbor's action both increase robustness to disturbances.