Evaluating memetic building spatial design optimisation using hypervolume indicator gradient ascent

In traditional, single objective, optimisation local optima may be found by gradient search. With the recently introduced hypervolume indicator (HVI) gradient search, this is now also possible for multi-objective optimisation, by steering the whole Pareto front approximation (PFA) in the direction of maximal improvement. However, so far it has only been evaluated on simple test problems. In this work the HVI gradient is used for the real world problem of building spatial design, where the shape and layout of a building are optimised. This real world problem comes with a number of constraints that may hamper the effectiveness of the HVI gradient. Specifically, box constraints, and an equality constraint which is satisfied by rescaling. Moreover, like with regular gradient search, the HVI gradient may overstep an optimum. Therefore, step size control is also investigated. Since the building spatial designs are encoded in mixed-integer form, the use of gradient search alone is not sufficient. To navigate both discrete and continuous space, an evolutionary multi-objective algorithm (EMOA) and the HVI gradient are used in hybrid, forming a so-called memetic algorithm. Finally, the effectiveness of the memetic algorithm using the HVI gradient is evaluated empirically, by comparing it to an EMOA without a local search method. It is found that the HVI gradient method is effective in improving the PFA for this real world problem. However, due to the many discrete subspaces, the EMOA is able to find better solutions than the memetic approach, albeit only marginally.