Reinforcement Learning vs. Computational Intelligence: Comparing Service Management Approaches for the Cloud Continuum

TL;DR: This paper makes a comparison of different optimization algorithms and a first investigation of how they can perform in this kind of scenario and demonstrates how all approaches can solve the service management problem with similar performance—with a different sample efficiency—if a high number of samples can be evaluated for training and optimization.

Abstract: Modern computing environments, thanks to the advent of enabling technologies such as Multi-access Edge Computing (MEC), effectively represent a Cloud Continuum, a capillary network of computing resources that extend from the Edge of the network to the Cloud, which enables a dynamic and adaptive service fabric. Efficiently coordinating resource allocation, exploitation, and management in the Cloud Continuum represents quite a challenge, which has stimulated researchers to investigate innovative solutions based on smart techniques such as Reinforcement Learning and Computational Intelligence. In this paper, we make a comparison of different optimization algorithms and a first investigation of how they can perform in this kind of scenario. Specifically, this comparison included the Deep Q-Network, Proximal Policy Optimization, Genetic Algorithms, Particle Swarm Optimization, Quantum-inspired Particle Swarm Optimization, Multi-Swarm Particle Optimization, and the Grey-Wolf Optimizer. We demonstrate how all approaches can solve the service management problem with similar performance—with a different sample efficiency—if a high number of samples can be evaluated for training and optimization. Finally, we show that, if the scenario conditions change, Deep-Reinforcement-Learning-based approaches can exploit the experience built during training to adapt service allocation according to the modified conditions.