Long Short Term Memory Autoencoder-aided Evolutionary Algorithm to Solve an Energy-Minimized Task Scheduling Problem

This paper addresses a task scheduling problem with deadline constraints in a human-cyber-physical system, which contains three subproblems, i.e., allocating processer, and determining tasks' sequence and frequency. To efficiently find its energy-efficient solutions in a short time, an autoencoder-aided evolutionary algorithm is proposed. The main optimizer chosen for it is genetic programming. To extract the implicit relationship among three strongly-coupled subproblems, a novel long short term memory autoencoder is built. In it, a group of long short term memory units are used to learn major features of decision variables and generate a low-dimensional hidden representation of a solution. After that, some network-aided mutation operators are designed to generate offsprings in the resulting low-dimensional space with informative features. Numerical experiments comparing the proposed method with several competitive methods verify the effectiveness of the proposed method in finding high-quality schedules in a reasonable time.