Semantic Approach in Genetic Programming
Evolutionary algorithms are stochastic optimization techniques based on the principles of natural evolution, and Genetic Programming (GP) belongs to this family. In recent years the study of GP systems has been extended to phenotypic aspects, while previously it was mainly focused on genotypic and syntactic aspects. Phenotypic properties, or semantics, is used with the aim of optimizing the ability of GP algorithms to explore the solution space in an effective way, increasing the probability of finding an optimal solution and escaping local optima. Currently, semantic GP is strictly related to the evaluation of individual behavior in the candidate population: this kind of evaluation is mainly obtained through the fitness function itself. This work introduces a new way of measuring semantic similarity between individuals that is more independent from the fitness itself, allowing a fair comparison even when the fitness values involved are far away from each other. This new measure enables a new series of techniques to be used to tackle the open problems in GP, like bloat and overfitting, and also targeting the phenotype variety preservation, thereby enhancing performance. Preliminary results will be provided. A new theoretical GP algorithm based on this new semantic measure is also introduced, showing the potential advantages. Very early results coming from a first naive implementation show interesting insight on this potential, when compared with other cutting edge algorithms.