Adaptive Optimization of Urban Layout for Air Quality using a GPU Cluster

In prior work, a fast GPU (Graphics Processing Unit) implementation of the Quick Urban and Industrial Complex (QUIC) dispersion system was used to optimize a simplified urban layout of minimal geometry with regard to pollutant concentrations. Yet as more parameters, such as locations or volumes for additional buildings are added to the optimization landscape, the size of the search space rises exponentially. The work presented here shows several approaches taken to reduce optimization time allowing for optimization of more complex urban scenarios. In particular, a stochastic optimization approach, Sampling Importance Resampling (SIR), is used, which maintains a population of samples in the search space and resamples them based on their fitness in the optimization. SIR has several useful properties for this type of problem, including scaling well with coarse-grain parallelism afforded by a cluster of GPU-based workstations. As each population is generated, rapid completion of the populations' simulations is achieved by pushing the dispersion simulations to multiple GPUs. Additional acceleration methods, such as adaptively varying the number of pollution particles in the simulations can significantly decrease execution time while providing rough, yet low-precision approximations during problem exploration in the initial phases. At the conference, our techniques for GPU-based simulation and adaptive optimization will be presented along with results from more complex urban layouts optimized for lowered pollution concentrations.