Created by W.Langdon from gp-bibliography.bib Revision:1.4659
The main contributions and results are the application of GP to a new field, and the conclusion that GP is better suited to solve this complex problem by a generate-and-test approach than an analytic one.
Three systems were implemented to evolve programs for calculating visibility spaces. The first used untyped GP and low-level operations, for maximum flexibility in evolution, but could solve the problem only for trivial cases.
The second used high-level geometric operations and typed GP, but tended to get trapped in local optima. Approaches used, unsuccessfully, to obviate this included altering the fitness cases and function set both statically and dynamically, parameter tuning, seeding the population, using program templates, and using a simpler system for modelling evolution.
The third system, which used a generate-and-test approach, evolved useful solutions. When seeded with hand-crafted partial solutions, it was able to improve them considerably.
The work shows the potential of GP to evolve or refine a region-growing generate-and-test algorithm for calculating visibility spaces, a problem not hitherto approached by the GP community.",
Genetic Programming entries for Michael Sean Grant