Autonomous intelligent decision-making system based on Bayesian SOM neural network for robot soccer

The complex confrontation in robot soccer match requires the decision-making system to learn the priori-knowledge given by humans and learn from its own experience. The two learning issues are usually addressed in two phases: off-line learning and on-line learning. Though lots of methods have been developed to address the two issues separately, the construction of a fully autonomous intelligent decision-making system remains challenging because of the difficulty of connecting the two phases. Most existing intelligent decision-making systems focus on only one of the two phases consequently. The model and algorithms of the Bayesian SOM neural network are proposed in this paper, based on which a fully autonomous intelligent decision-making system for robot soccer is built. This model provides a knowledge structure which can be shared by the off-line learning and on-line learning algorithms. By integrating the Bayesian classifier into each neuron, the whole neural network is equivalent to a multi-agent decision-making system. In the on-line learning phase, the Bayesian method is used to update each neuron's beliefs and the whole network's estimation of the state space. In matches with different opponents, this Bayesian SOM intelligent decision-making system showed outstanding learning ability and great adaptivity.