Multiagent reinforcement learning in Markov games : asymmetric and symmetric approaches

摘要

Modern computing systems are distributed, large, and heterogeneous. Computers, other information processing devices and humans are very tightly connected with each other and therefore it would be preferable to handle these entities more as agents than stand-alone systems. One of the goals of artificial intelligence is to understand interactions between entities, whether they are artificial or natural, and to suggest how to make good decisions while taking other decision makers into account. In this thesis, these interactions between intelligent and rational agents are modeled with Markov games and the emphasis is on adaptation and learning in multiagent systems.\n\nMarkov games are a general mathematical tool for modeling interactions between multiple agents. The model is very general, for example common board games are special instances of Markov games, and particularly interesting because it forms an intersection of two distinct research disciplines: machine learning and game theory. Markov games extend Markov decision processes, a well-known tool for modeling single-agent problems, to multiagent domains. On the other hand, Markov games can be seen as a dynamic extension to strategic form games, which are standard models in traditional game theory. From the computer science perspective, Markov games provide a flexible and efficient way to describe different social interactions between intelligent agents.\n\nThis thesis studies different aspects of learning in Markov games. From the machine learning perspective, the focus is on a very general learning model, i.e. reinforcement learning, in which the goal is to maximize the long-time performance of the learning agent. The thesis introduces an asymmetric learning model that is computationally efficient in multiagent systems and enables the construction of different agent hierarchies. In multiagent reinforcement learning systems based on Markov games, the space and computational requirements grow very quickly with the number of learning agents and the size of the problem instance. Therefore, it is necessary to use function approximators, such as neural networks, to model agents in many real-world applications. In this thesis, various numeric learning methods are proposed for multiagent learning problems.\n\nThe proposed methods are tested with small but non-trivial example problems from different research areas including artificial robot navigation, simplified soccer game, and automated pricing models for intelligent agents. The thesis also contains an extensive literature survey on multiagent reinforcement learning and various methods based on Markov games. Additionally, game-theoretic methods and methods originated from computer science for multiagent learning and decision making are compared.