Mining Markov Network Surrogates for Value-Added Optimisation

Abstract

Surrogate fitness functions are a popular technique for speeding up metaheuristics, replacing calls to a costly fitness function with calls to a cheap model. However, surrogates also represent an explicit model of the fitness function, which can be exploited beyond approximating the fitness of solutions. This paper proposes that mining surrogate fitness models can yield useful additional information on the problem to the decision maker, adding value to the optimisation process. An existing fitness model based on Markov networks is presented and applied to the optimisation of glazing on a building facade. Analysis of the model reveals how its parameters point towards the global optima of the problem after only part of the optimisation run, and reveals useful properties like the relative sensitivities of the problem variables.