Using a hybrid evolutionary-A* approach for learning reactive behaviours

Abstract

A hybrid approach for learning reactive behaviours is presented in this work. This approach is based on combining evolutionary algorithms (EAs) with the A* algorithm. Such combination is done within the framework of Dynastically Optimal Forma Recombination, and tries to exploit the positive features of EAs and A* (e.g., implicit parallelism, accuracy and use of domain knowledge) while avoiding their potential drawbacks (e.g., premature convergence and combinatorial explosion). The resulting hybrid algorithm is shown to provide better results, both in terms of quality and in terms of generalisation.