Comparative Study of Evolutionary Algorithms for a Hybrid Analog Design Optimization with the use of Deep Neural Networks

Abstract

Analog design optimization is the process of optimizing the circuit parameters to achieve specific performance metrics. In order to choose the best optimization methodology, a comparative study between different methodologies is needed. This work introduces hybrid design optimization method that combines Evolutionary Algorithms (EA) such as Particle Swarm Optimization (PSO) or Genetic Algorithm (GA) with a multi-output Deep Neural Network (DNN) to obtain both fast and accurate circuit optimizer. A CMOS Miller op-amp is used as an example of the optimization problem. Training data for the DNN is extracted with Mentor Analog Fast Spice (AFS) and using TSMC 90nm PDK. This work gives important insights on how to choose the best DNN structure by showing that using Adadelta optimizer in the DNN training phase is the best compared to Adagrad and Gradient Descent(GD). Moreover, it is proven that there is an optimum size of the DNN to achieve the least prediction error. Finally, a comparative study between PSO and GA algorithms proved that PSO has less failure rate for all test iterations.