Abstract: |
Genetic Programming uses trees to represent chromosomes. The user defines the representation space by defining the set of functions and terminals to label the nodes in the trees. The sufficiency principle requires that the set be sufficient to label the desired solution trees, often forcing the user to enlarge the set, thus also enlarging the search space. Structure-preserving crossover, STGP, CGP, and CFG-based GP give the user the power to reduce the space by specifying rules for valid tree construction, based on types, syntax, and heuristics. These rules in effect change the representation. However, in general the user may not be aware of the best representation, including heuristics, to solve a particular problem. Last year, ACGP methodology was introduced for extracting local problem-specific heuristics, that is for learning a local model of the problem domain. ACGP discovers representation, in the space of probabilistic representations, one that improves the search itself and that provides the user with heuristics about the domain. This paper discusses and illustrates the probabilistic representation. |
