Some thoughts on the use of sampled fitness functions for the multilevel Darwinist brain

When attempting to evolve models of the real world through the information obtained by interacting with it, we always come across the same problem: the fitness function for the problem, that is, the real world, can only be known in a sampled manner. In this article we study the effect of different parameters and algorithmic strategies when working with sampled fitness functions in evolutionary processes. The results presented here correspond to a study of the effect of the size of the Short Term Memory and the number of generations between updates for a generic genetic algorithm that will be operating within the Multilevel Darwinist Brain. From these results, several critical points may be considered in order to define the limits to which computations can be simplified when working with sampled fitness functions while maintaining the same representational power. We provide a study of different proposals for the construction of Short Term Memories and their replacement strategies in order to obtain the maximum information with the minimum use of resources.