Bayesian Optimization methods

Bayesian Optimization methods are often employed in a hardware design setting where simulation of a given hardware configuration can be computationally very expensive. The problem gets more difficult when there are multiple objective functions to balance against each other, for instance, the power consumption and the accuracy of the hardware accelerator. The set of optimal configurations that are not outperformed in both objective functions is called the Pareto front. This project uses Deep Gaussian Process models in order to give better prediction of the Pareto front of Multiobjective Optimization problems. The state-of-the-art algorithms use Gaussian Processes to model each objective function. This approach can be inefficient because it is unable to capture the correlation between the different objectives. Deep Gaussian Processes can improve on this by sharing multiple layers of Gaussian Processes between the objective functions. By capturing these dependencies, we can increase the accuracy of the model and therefore increase the accuracy of the resulting Pareto front.