Multi-agent Dynamic Pricing Using Reinforcement Learning and Asymmetric Information

Alexander Kastius, Nils Kiele, Rainer Schlosser

Abstract: Self-learning agents can be used in numerous ways for dynamic pricing nowadays. It has been shown, that reinforcement learning can serve as a toolkit to efficiently develop pricing strategies in dynamic environments. In many real-world situations, it can be expected that multiple market participants rely on such self-learning agents to implement pricing decisions. From the view of one agent, this violates the fundamental Markov property, which leads to instability in the learning process. Past publications proposed to rely on asymmetric information to achieve equilibria and usually focused on tabular solutions or solvers. We use multi-agent learning and asymmetric information with function approximation tools for high-dimensional state spaces by exchanging policy information between multiple actors. We discuss possible problems and their solutions and propose a simulation environment for further evaluation of the developed system.