Differential Evolution under Fixed Point Arithmetic and FP16 Numbers

Abstract

In this work, the differential evolution algorithm behavior under a fixed point arithmetic is analyzed also using half-precision floating point (FP) numbers of 16 bits, and these last numbers are known as FP16. In this paper, it is considered that it is important to analyze differential evolution (DE) in these circumstances with the goal of reducing its consumption power, storage size of the variables, and improve its speed behavior. All these aspects become important if one needs to design a dedicated hardware, as an embedded DE within a circuit chip, that performs optimization. With these conditions DE is tested using three common multimodal benchmark functions: Rosenbrock, Rastrigin, and Ackley, in 10 dimensions. Results are obtained in software by simulating all numbers using C programming language.

Numerical and Evolutionary Optimization 2020

“…In [10], de la Fraga analyzes the use of numbers with 16 bits in a conventional Differential Evolution (DE) algorithm. It is shown that the additional use of fixed point arithmetic can speed up the evaluation time of the objective function.…”