In this paper we demonstrate how evolutionary computation can be used to acquire difficult to solve combinatorial problem instances, thereby stress-testing the corresponding algorithms used to solve these instances. The technique is applied in three important domains of combinatorial optimisation, binary constraint satisfaction, Boolean satisfiability, and the travelling salesman problem. Problem instances acquired through this technique are more difficult than ones found in popular benchmarks. We analyse these evolved instances with the aim to explain their difficulty in terms of structural properties, thereby exposing the weaknesses of corresponding algorithms.

We show how an evolutionary algorithm can successfully be used to evolve a set of difficult to solve symmetric travelling salesman problem instances for two variants of the Lin-Kernighan algorithm. Then we analyse the instances in those sets to guide us towards deferring general knowledge about the efficiency of the two variants in relation to structural properties of the symmetric travelling salesman problem.

This paper introduces a generator that creates problem instances for the Euclidean symmetric travelling salesman problem. To fit real world problems, we look at maps consisting of clustered nodes. Uniform random sampling methods do not result in maps where the nodes are spread out to form identifiable clusters. To improve upon this, we propose an evolutionary algorithm that uses the layout of nodes on a map as its genotype. By optimising the spread until a set of constraints is satisfied, we are able to produce better clustered maps, in a more robust way. When varying the number of clusters in these maps and, when solving the Euclidean symmetric travelling salesman person using Chained Lin-Kernighan, we observe a phase transition in the form of an easy-hard-easy pattern.

We present a study on the difficulty of solving binary constraint satisfaction problems where an evolutionary algorithm is used to explore the space of problem instances. By directly altering the structure of problem instances and by evaluating the effort it takes to solve them using a complete algorithm we show that the evolutionary algorithm is able to detect problem instances that are harder to solve than those produced with conventional methods. Results from the search of the evolutionary algorithm confirm conjectures about where the most difficult to solve problem instances can be found with respect to the tightness.