Crossover context in page-based linear genetic programming

This work explores strategy learning through genetic programming in artificial ´ants´ that navigate the Son Mateo trail, We investigate several properties of linearly structured (as opposed to typical tree-based) GP including: the significance of simple register based memories, the significance of constraints applied to the crossover operator, and how ´active´ the ant are. We also provide a basis for investigating more thoroughly the relation between specific code sequences and fitness by dividing the genome into pages of instructions and introducing an analysis of fitness change and exploration of the trail done by particular parts of a genome. By doing so we are able to present results on how best to find the instructions in an individual´s program that contribute positively to the accumulation of effective search strategies.