On the Use of Surrogate Models for Derivative Free Optimization on Real Consumer Circuits
Surrogate models such as Support Vector Machines and Neural Network have become popular in the fields of continuous function regression and series analysis.
In most applications objective functions are not available is closed form and only samples of such functions can be used in order to perform the optimization. For these reasons, it is quite natural to use surrogate models to approximate these functions and then use derivative free methods in order to perform the optimization.
In this seminar we talk about the application of surrogate models together with derivative free methods for solving the problem of optimal design of real world circuits. The relations between the design parameters of the circuits and its performances are given by computationally expensive computer simulations. In other words, the objective function that must be optimized is not in closed form. Furthermore, there are strict temporal limitations in industrial design due to the time to market. This motivates the effort to speed up the circuit analysis phase by resorting to the use of surrogate models rather than simulations.
The method applied to problem is an efficient derivative-free optimization routine for mixed-integer nonlinear problems , as the parameters of the circuit can be both continuous and discrete.
Finally, we also present the result of a novel derivative free robust optimization strategy in order to find an optimal design point robust to statistical variations.