The Benefits of Adaptive Parameterisation in Multi-Objective Tabu Search
Tiziano Ghisu, Geoffrey Thomas Parks, J. P. Jarrett, P. John Clarkson · UNICA IRIS Institutional Research Information System (University of Cagliari) · 2009
Real-world optimisation problems are often characterised by large design spaces and a large number of objectives and constraints, generating a highly fragmented and multi-modal landscape that can prove particularly challenging for any optimisation algorithm. Tabu Search has proved particularly efficient in the solution of such optimisation problems, thanks to the reliance on a local search algorithm (Hooke & Jeeves pattern search in the case of our implementation) that, by applying small changes to the design vector, is able to navigate the complex design space efficiently. In this work, we present a modification of the local search algorithm that, based on an adaptive change of the design space parameterisation, allows a re-direction of the search towards the most energetic directions of the design space, which are more likely to produce improvements in the figures of merit. The same approach allows also the dimensionality of the problem to be temporarily reduced, leading to a greater effectiveness of the local search (not only because of the smaller number of designs to be evaluated at every optimisation step, but also because of the elimination of the noisy, less important design variables).