Investigating coevolutionary algorithms For expensive fitness evaluations in cybersecurity

Abstract

Coevolutionary algorithms require evaluating fitness of solutions against adversaries, and vice versa, in order to select high quality individuals to generate offspring and evolve the population. However, some problems require computationally expensive fitness evaluations, which makes it hard to generate solutions in a feasible amount of time. In this thesis, we devise coevolutionary algorithms and methods that achieve good results with fewer fitness evaluations, and we present methods for selecting a solution to deploy after running experiments with multiple coevolutionary algorithms. Comparing our new algorithms presented with baselines, we found that MEULockstepCoev performs relatively well, especially for attackers.