Data-informed active design space reduction and optimisation using self-organising maps

For black-box optimisation problems with high simulation costs, efficient global optimisation methods based on surrogate models and space reduction are widely used to reduce modelling costs and improve optimisation efficiency. However, regional reduction in unknown design spaces remains challenging in practical engineering applications. In this work, a new space reduction method is proposed. This method uses surrogate models and self-organising maps (SOM) to address the optimisation of dense black-box objective functions. During the reduction process, the interesting weight extreme value of the self-organising map network is used as an alternative space reduction scheme, forming a set of candidate schemes. Statistics are performed within the set of candidate schemes to determine the space reduction scheme, while promising sampling points are selected from the interesting weights of the SOM. Tests on ten representative benchmarks and two engineering design optimisation problems show that the newly proposed design space reduction method has improved capabilities regarding irreversible loss of attractive spaces due to over-reduction when early model accuracy is poor and cost wastage caused by heuristic algorithm searches for additional samples. It exhibits good search efficiency and robustness in identifying the global optimum.