Desenvolvimento e avaliação de um framework modular para síntese automática de circuitos analógicos: aplicação do algoritmo recozimento simulado com evolução geométrica de circuitos

Abstract

This work presents the development and evaluation of a modular framework for automatic synthesis of analog electronic circuits entitled circ_autoproj. The main component of the framework is SANN-GCE, a metaheuristic driven by SPICE simulations that automatically generates the topology and dimensioning of the components of an electronic circuit based on a user-defined desired behavior. SANN-GCE is composed of the Simulated Annealing search algorithm (SANN) and the solution representation called Geometric Circuit Evolution (GCE). GCE is a new coding scheme that uses categorized degrees of freedom that allow distinct characteristics of a circuit to mutate, during the evolution of the solution, with different probabilities according to their category. This work also presents the Ngspice circuit simulator, used by SANN-GCE to evaluate candidate solutions, and the concepts of cloud computing and its use as a tool to accelerate the execution of circ_autoproj. SANN-GCE was tested in seven use cases: temperature sensor, Gaussian function, voltage reference, quadratic function, square root, cube function and cube root. Each configuration was run 50 times and performance on these circuits was evaluated using 12 metrics, compared to the ACID-MGE reference algorithm. A study was also carried out on the effect of adjustments to the algorithm parameters. In it, 7 selected parameters were changed in 41 distinct configurations. The statistical significance of the results of these adjustments was evaluated using a Permutation Test. This study revealed that, among the 41 configurations evaluated, 19 presented statistically significant results covering all use cases, with an average p value of 0.00993, approximately five times lower than the significance level usually set at 0.05. The data obtained indicated that SANN-GCE performed better and with less variability between runs when compared to ACID-MGE, as better results were obtained in 11 of the 12 metrics. The following best medians obtained were highlighted: 1.52x for success rate, 13.94x for average fitness, 64.75x for standard deviation of fitness and 7.83x for the standard deviation of the Hits percentage. Also, the median execution time of SANN-GCE was 15.9x shorter when compared to the normalized execution time of ACID-MGE.