Representation and Learning Methods for Situation Evaluation in RoboCup Soccer Simulation

Abstract

In this paper, representations of a situation evaluation model and their learning methods in RoboCup soccer simulation 2D are reviewed. A soccer game is commonly known as an example of multi-agent systems in an uncertain and dynamic environment. First, this paper discusses the similarities and differences between the RoboCup soccer simulation and other benchmark tests of game AI research. Next, a search method based on action chains is presented as a mechanism of action selection by a soccer player. This method is based on the look-ahead of field situations by the combination of a search tree and a situation evaluation model, which is the case in the chess and shogi programs. Then, various methods of constructing the situation evaluation models that are used in the generation of action chains by the machine learning framework are reviewed along with example models that have been proposed.