Mimicking Dribble Trajectories by Neural Networks for RoboCup Soccer Simulation

This paper proposes a framework for acquiring a low-level behavior of a soccer agent. The task of a learning agent is to mimic the behavior of a target agent with a well-trained behavior. Neural networks are used to represent the behavior of the target agent. In order to obtain a set of training data, we convert game logs of the target agent into a set of input-output pairs for the learning of neural networks. We consider two implementations of neural networks. The first implementation maps a situation of a dribbling agent at a certain time step to an action to be conducted at the next time step. In the second implementation three neural networks are used for three possible action such as turn, dash, and kick. Each neural network presents the activation degree of the correspondent action at the next time step. We show the effectiveness of the proposed framework through the computational experiments.