Learning Transferable Variation Operators in a Continuous Genetic Algorithm

The notion of experience has often been neglected within the domain of evolutionary computation while in machine learning a large variety of methods has emerged in the recent years under the umbrella of transfer learning. Notably, realizing experience-based methods suffers from a variety of conceptual key problems. The first one being in regards to what constitutes problem-similarity from an algorithm perspective and the second one being what constitutes the transferable experience by itself. Ideally, one would envision that a learning optimization algorithm could be expected to act similarly to a human-problem solver who tackles novel tasks initially without any preconceptions. Experience only comes into play until sufficient similarity to known problems is established. Our paper therefore has two aims. First, to outline existing related fields and methodologies and highlight their insufficiencies. Second, to make the case for experience-based optimization by a demonstration using a novel and statistics-based approach with a real-coded genetic algorithm as a case study. In this paper we do not claim to construct universal problem solvers, but instead propose that from an algorithm-specific-view, problem characteristics can be learned and harnessed to improve future performance of similarly-structured optimization tasks.