On Using Surrogates with Genetic Programming.

Area of Science:

• Computational Intelligence
• Evolutionary Computation
• Machine Learning

Background:

• Evolutionary algorithms with expensive fitness evaluations can be accelerated using surrogate models.
• Traditional surrogate models require numerical representations, limiting their use with genetic programming's (GP) tree-based genotypes.
• Genetic programming (GP) often involves computationally intensive fitness evaluations.

Purpose of the Study:

• To develop a novel method for integrating surrogate models with genetic programming (GP).
• To overcome the limitation of numerical representations in surrogate models for GP.
• To improve the efficiency and effectiveness of GP for problems with expensive fitness functions.

Main Methods:

• Proposed using phenotypic characterization as input for surrogate models in GP, instead of direct genotype representation.
• Developed efficient methods for computing phenotypic characterization and defining approximate equivalence and similarity measures.
• Applied the approach to a stochastic, dynamic job shop scheduling problem using simulation-based GP.

Main Results:

• Demonstrated the successful construction of surrogate models using phenotypic characterization in GP.
• Showcased significant improvements in convergence speed compared to standard GP.
• Achieved enhanced solution quality for the job shop scheduling problem.

Conclusions:

• Phenotypic characterization offers a viable and effective way to integrate surrogate models with genetic programming (GP).
• This novel approach accelerates evolutionary algorithms with expensive fitness evaluations.
• The method improves both the speed and quality of solutions generated by GP.