Evaluation Function Modeling using Action Sequences in RoboCup Soccer

Abstract

In RoboCup Soccer Simulation 2D League, almost all teams use an evaluation function in action selection. Team performance highly depends on the evaluation function. The aim of this paper is to propose a method that improves the performance of a team by imitating a target team. For this purpose, a neural network is employed to model the target team’s decision making. The neural network is trained by using positive and negative episodes of a target team’s action sequences that are extracted from game logs. Though our computational experiments, it is shown that the performance (e.g., win rate, scored goal, and so on) is improved by mimicking the winner of RoboCup 2017 soccer simulation 2D league.