Apprentissage par renforcement flou pour l’auto-adaptation des systèmes conteneurisés en périphérie de réseau

Abstract

Among the desired aspects of a self-adaptive containerized systems architecture is the need to be able to permanently survey the operational environment, catch and observe anomalies, and provide a good policy for self adaptation, recovery, and configuration of processing resources to dynamically respond to an abrupt change in its operational environment. The behavior of a microservices architecture is constantly changing over time, making it difficult to use a static model to plan the appropriate runtime action when an anomaly is detected. To achieve the excellent levels of self-adaptability desired, this research allows us to realize a microservices architecture model following the MAPE-K model. Our solution proposes a combination of Q-learning, fuzzy logic and neural networks to dynamically select the adaptation action that brings the highest reward. The presence of fuzzy Q-learning (FQL) allows fuzzy adaptation rules to be learned and modified at runtime without the need for prior knowledge. The implementation of Fuzzy Deep Q Network (FDQN) in the context of the adaptation process improves the efficiency of adaptation and reduces the risks associated with adaptation, including resource fluctuation. The experimentation demonstrates the feasibility and utility of our approaches. Moreover, the results show that the performance of FDQN is better than that of FQL.

Date	11 Aug 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Abdelouahed Gherbi (Supervisor)