Epigenetic crossover for multi-objective Evolutionary Algorithms

Abstract

There are hundreds of different Evolutionary Computation approaches developed to solve multi-objective optimisation problems. Among these approaches there are only two fundamental evolutionary concepts, split between the two main areas of the field: Evolutionary Algorithms with genetic inheritance, and Swarm Intelligence with cultural inheritance. Modern evolutionary biology has since continued to study further evolutionary and non-genetic mechanisms that is relatively unexplored in Evolutionary Computation.

In this thesis, a framework to analyse existing Evolutionary Computation algorithms is developed, based on evolutionary concepts from the Extended Evolutionary Synthesis. The gap in epigenetic inheritance is identified through this analysis as an approach with high potential due to its fast partially-genetic adaptability to dynamic changes in the environment. Furthermore, a detailed benchmarking suite is used to test and compare existing Genetic Algorithms and Particle Swarm Optimisation algorithms to determine their differences and suitability to incorporate a new epigenetic mechanism. 

Genetic Algorithms are therefore chosen for epigenetics because the increased diversity balances the convergence properties of epigenetic fast adaptations. Next, a novel epigenetic blocking mechanism based on gene silencing is developed and tested. Performance on static and dynamic multi-objective problems show the improvement the epigenetic mechanism can make, with improved performance across the  duration of the optimisation process. Further study and comparison of the hyperparameters and a gradient-based approach indicate the mechanism can be both problem and algorithm specific. The choice of blocking variables with all positive or negative gradients achieve the best general results, and utilising different hyperparameters specifically tuned for problems with changing Pareto sets achieve the best performance.

Finally, the epigenetic mechanism is applied to a real-world voyage optimisation system. Faster convergence is demonstrated for voyages in calm weather conditions, and savings on fuel consumption are found for more complex voyages in severe weather conditions.

More information

Identifiers

ORCID for Sizhe Yuen: orcid.org/0000-0001-5552-2074

ORCID for Adam Sobey: orcid.org/0000-0001-6880-8338

ORCID for Tom Ezard: orcid.org/0000-0001-8305-6605

Catalogue record

Export record