joint work with Mahdi Aziz, Warren Hare, Majid Jaberipour
We model the road design problem as a bilevel problem. The top level calculates the road path from a satellite's viewpoint while the lower level is a mixed-integer linear problem (solving the vertical alignment i.e. where to cut and fill the ground to design the road and the earthwork i.e. what material to move to minimize costs) for which we build a variable-fidelity surrogate using quantile regression. We compare two variable-fidelity algorithms from the literature (a generalized pattern search with adaptive precision vs. a trust-region with controlled error) and report significant speedups.
joint work with Nicolas Calvet, Peter R. Armstrong
The reflectors of a beam-down solar concentrator are adjusted to maximize the net power collected at the receiver. The performance of the solar plant is predicted with a Monte Carlo ray-tracing model as a function of a feasible reflector geometry. Optimization is carried out with NOMAD, an instance of the mesh-adaptive direct search (MADS) blackbox algorithm. Challenges include reducing the number of optimization variables, dealing with the stochastic aspect of the blackbox, and the selection of effective MADS optimization parameters.
joint work with Dounia Lakhmiri, Christophe Tribes
The performance of deep neural networks is highly sensitive to the choice of the hyper-parameters that define the structure of the network and the learning process. Tuning a deep neural network is a tedious and time consuming process that is often described as a "dark art". This work introduces the HYPERNOMAD package based on the MADS derivative-free algorithm to solve this hyper-parameter optimization problem including the handling of categorical variables. This approach is tested on the MNIST and CIFAR-10 datasets and achieves results comparable to the current state of the art.