An Opponent Formation Classifier for Simulated Robot Soccer

Abstract

This paper contributes with a novel opponent formation classifier for simulated robot soccer. Modeling opponent positioning strategy can help the development of dynamic soccer behaviors and increase the team’s performance. We used tools from RoboCup 2D Soccer Simulation League (RoboCup 2D) for running soccer matches. Several researchers of RoboCup 2D developed classifiers for soccer formations but were limited to defense or preestablished formations. To acquire the formation categories of the opponents, we obtained positioning data from game logs against 11 opponents and detected mean formation patterns with the Gaussian Mixture Model clustering algorithm. Later, we trained a neural network and a random forest to predict four formation categories. The random forest outperformed the neural networks on average for unseen teams during training. However, the neural network achieved the lowest error rate for four unseen teams during training. Last, we introduced a method to decrease the number of game cycles required for correct classification by choosing the predicted label with the highest frequency over time.

Supported by CNPq.