Competitive coevolutionary training of simple soccer agents from zero knowledge

A new competitive coevolutionary team-based particle swarm optimisation (CCPSO) algorithm is developed to train multi-agent teams from zero knowledge. The CCPSO algorithm uses the charged particle swarm optimiser to train neural network controllers for simple soccer agents. The training performance of the CCPSO algorithm is analysed. The analysis identifies a critical weakness of the CCPSO algorithm in the form of outliers in the measured performance of the trained players. A hypothesis is presented that explains the presence of the outliers, followed by a detailed discussion of various biased and unbiased relative fitness functions. A new relative fitness function based on FIFA´s league ranking system is presented. The performance of the unbiased relative fitness functions is evaluated and discussed. The final results show that the FIFA league ranking relative fitness function outperforms the other unbiased relative fitness functions, leading to consistent training results.