Genetic Programming explores the problem search space by means of operators and selection. Mutation and crossover operators apply uniformly, while selection is the driving force for the search. Constrained GP changes the uniform exploration to pruned non-uniform, skipping some subspaces and giving preferences to others, according to some heuristics. Adaptable Constrained GP is a methodology for discovery of such useful heuristics. Both methodologies have previously demonstrated their surprising capabilities using only first-order (parent-child) heuristics. Recently, they have been extended to second-order (parent-children) heuristics. This paper describes the second-order processing, and illustrates the usefulness and efficiency of this approach using a simple problem specifically constructed to exhibit strong second-order structure.
Categories and Subject Descriptors
General TermsDesign, Experimentation.
KeywordsGenetic Programming, Heuristics, Search Space.
BackgroundGenetic Programming (GP) is an evolutionary computation method bringing together concepts from computer science and nature. It solves a problem at hand by using a population of candidate solutions, represented as chromosomes, and by manipulating the solutions via simulated mutation and crossoverwhile driven by selection to explore better solutions.Even though GP methods have been devised to work with a broad range of possible representations for the candidate solutions, the most common representation is that of a tree [1,6]. These trees are labeled with functions and terminals representing problemspecific elements: functions, connectors, constants, sensors, etc. The actual search space, called genotype space, searched by GP is uniquely determined by the labels, and only constrained by limits on tree size or depth -the trees can be labeled in any arityconsistent manner (the closure property [6]). The corresponding solution space, called phenotype space, depends on the interpretations of the labels -the interpretations provide a mapping from the search space to the solution space or from genotype to phenotype space. Somewhere in the search space, GP attempts to find a point mapped to the actual solution in the solution space, which will provide the solution to the problem at hand. The quality of a single point in the search space is determined by evaluating the mapped solution through a provided black-box fitness function.There are some important issues to consider when designing GP, similar to those of other evolutionary methods yet specific to GP. If a given solution does not have a search space point mapped into it, it will never be discovered. Therefore, the mapping must be onto. To accomplish this, in the absence of detailed information about the problem or solution, the search space needs to be enlarged (a part of the sufficiency principle [6]). This leads to many-to-one mappings, with large redundancy in the representation. To handle these problems, some properties need to be there, among them many-to-one mappings to the better solut...