A Study on Initial Population Sampling for Multi-Objective Optimization based on Differential Evolution and Bayesian Inference

Multi-objective optimization techniques are becoming an important tool for circuit sizing. The performance of these techniques has been validated on circuits with many design variables and responses that represent a highly complex task when approached using standard methodologies. Considering this, optimization techniques that tackle this challenge are usually based on evolutionary algorithms. Reducing the number of circuit simulations necessary for achieving solutions of high performance is of utmost importance, especially as simulations take a long time to execute, proportional with the circuit’s complexity. One way of approaching this is to select initial population candidates, which lead to a good solution faster. In this work we use Multi-Objective Optimization Based On Differential Evolution And Bayesian Inference (MODEBI), a state of the art algorithm with good results in circuit sizing, in order to investigate the impact of the initial population candidates on the simulation budget and number of iterations. Hence, we compare three approaches: Latin Hypercube, Sobol, and Halton against the traditional Monte Carlo used in MODEBI. Our experiments show that Latin Hypercube outperforms Monte Carlo and the number of required simulation is reduced with approximately 23%.