Multiple Solutions by Means of Genetic Programming: A Collision Avoidance Example

摘要

Seldom is it practical to completely automate the discovery of the Pareto Frontier by genetic programming (GP). It is not only difficult to identify all of the optimization parameters a-priori but it is hard to construct functions that properly evaluate parameters. For instance, the "ease of manufacture" of a particular antenna can be determined but coming up with a function to judge this on all manner of GP-discovered antenna designs is impractical. This suggests using GP to discover many diverse solutions at a particular point in the space of requirements that are quantifiable, only a-posteriori (after the run) to manually test how each solution fares over the less tangible requirements e.g. "ease of manufacture" . Multiple solutions can also suggest requirements that are missing. A new toy problem involving collision avoidance is introduced to research how GP may discover a diverse set of multiple solutions to a single problem. It illustrates how emergent concepts (linguistic labels) rather than distance measures can cluster the GP generated multiple solutions for their meaningful separation and evaluation.