| This work explores strategy learning through genetic programming applied to artificial ants that must learn to navigate the San Mateo Trail. The individuals in the problem arise from linearly-structured Genetic Programming, or L-GP, and their performance is compared to the traditional tree-based GP solution to the problem. The work examines a number of properties of the L-GP ants, including memory size, maximum instruction set size, constraints applied to the crossover operator (called “paging”), and how active the ants are. Metrics for evaluation of the fitness accumulated and exploration done by particular parts of the genome during evolution are presented. The analysis motivates the creation of an implementation robust to solving particularly hard problems where no strategy is readily available. A numerical and graphical analysis of the context of instructions in the ants' genome is also undertaken, and it is argued that paging in L-GP produces better solutions through retaining context within page boundaries. |
