The Collective Learning Procedure

This paper proposes the collective learning procedure. Collective learning uses a genetic algorithm to quickly sample the search space and allows each individual to perform local optimization steps. This hybrid scheme yields very good performance in the eld of continuous parameter optimization and is especially suited for optimizing multimodal functions. Furthermore, this paper shows that collective learning performs independent of a rotation of the coordinate system whereas other high-performance algorithms exhibit a drastic performance loss. In a second application domain, collective learning is used to simultaneously evolve and train minimal neural network topologies. Within a reasonable number of generations, the procedure evolves smaller networks than other existing approaches. In addition to practical results, the theoretical part explains and justiies certain design decisions.