Uma abordagem multiobjetivo usando programação genética cartesiana para projeto de circuitos digitais aproximados

Abstract

The reasons that justify studies involving Hardware Evolutionary (EHW), an area dedicated to the design of electronic devices that uses evolutionary algorithms in the design of a certain architecture like digital circuits and analog circuits, are: (i) developing models that are not attainable when traditional techniques are employed; and (ii) designing good solutions to problems and applications where specifications are incomplete or there are no optimal solutions. In this scenario, Approximate Computing appears as a new paradigm in response to the need to improve the performance or energy efficiency of electronic devices. In Approximate Computing, the accuracy of the models can be relaxed, generating designs for users who are willing to accept certain inaccuracies, such as applications imprecise arithmetic circuits, image compression processes, digital finite impulse response (FIR) and infinite impulse response (IIR) applications. Digital circuits obtained through Approximate Computing are classified as circuits. The functional requirements of the approximate digital circuits are relaxed in order to achieve: (i) energy savings, (ii) better response speed, (iii) less complexity of the circuit and (iv) smaller area occupied by the circuit components. We wish to investigate here the process of construction of approximate combinational digital circuits via Multiobjective Evolutionary Computation. For such a technique of Genetic Programming Cartesiana, based on the concept of Pareto dominance and with population size adaptive to deal with multiple objectives is proposed in this work to design approximate digital circuits by analyzing the compromise between propagation delay, power consumed and error. Such a technique is entitled CGPMO + TPA. Combinational digital circuits such as adders, multipliers, and Arithmetic Logic Units with up to 16 inputs and 370 logic gates are considered in computational experiments. The CGPMO + TPA method was compared with approaches in the literature and presented satisfactory results.