Multi-indicators Multi-objective Evolutionary Algorithm with Q-Learning for Real-world Network Optimization

Abstract

Aiming at the deployment optimization of complex Internet of Things (IoT) systems, we propose a new multi-objective optimization algorithm using multiple indicators with reinforcement learning, called MIEA-RL. In MIEA-RL, a set of evaluation indicators are employed to guide the evolution of population, while a Q-learning method is designed to manage these indicators in an efficient way during the search. To be specific, each candidate indicator is determined by the performance improvement of the population selected by the current indicator. Moreover, the search biases of different indicators can be adaptively balanced according to a Q-learning table. Accordingly, the convergence and diversity can be maintained effectively while the algorithm complexity is not increased. Finally, the MIEA-RL is applied to resolve the real-world IoT optimization instances in the experiment. Results show the proposed algorithm is effective and efficient to handle with these problems.