Field State Evaluation from Image Information in RoboCup Soccer Simulation

Abstract

We propose a method that evaluates the field from images without numerical information in RoboCup soccer simulation. Training data sets are generated by converting numerical field information into images and assigning corresponding target values according to the number of frames by the next goal. A convolution neural network is used for regression modeling where images are mapped into a scalar value. We discuss the way training data sets are generated and also the performance of the trained neural networks.