Quantum-inspired neuro coevolution model applied to coordination problems

Abstract

Abstract In many real-world problems, some coordination between agents is necessary to enable the task to be optimally performed. However, obtaining this coordination can be challenging due to the quantity and characteristics of the agents, the dynamics of the environment and/or the complexity of the task, requiring much computation time. Furthermore, some problems require different types of agent specialization. In this case, it is very difficult for the programmers to define learning strategies and their parameters. Optimization of these parameters using standard evolutionary algorithms is also inadequate due to the high computational cost in these real multi-agent situations. The main objective of this study is therefore to propose a new neuroevolution model to be applied to agent coordination problems, termed the Quantum-Inspired Neuro Coevolution (QNCo) Model. QNCo makes use of paradigms from quantum physics and biological coevolution to evolve sub-populations of quantum individuals aiming convergence gains. The model has the capacity to autonomously obtain the best neural network topology of each agent, eliminating the need for the programmer to set this configuration. New quantum crossover and mutation operators were proposed and compared during function optimization of different dimensions. The proposed model was tested in two simulation problems, prey-predator and multi-rover tasks, and one real problem of mobile telephony coverage. The QNCo model yielded promising results compared to similar algorithms, with good solutions in terms of learning strategies and a great reduction in convergence time.