Scholarlink

<p>Control-theoretic adaptation provides a well-established foundation for managing self-adaptive software systems. Traditional model-based approaches incur costly system identification and exhibit reduced robustness under frequent runtime changes, limiting responsiveness and scalability in dynamic environments. Recent hybrid approaches combine machine learning with control theory to tune controller parameters, but many still lack formal assurances for requirement satisfaction and mechanisms for managing runtime uncertainty. This paper introduces a hybrid adaptation framework that integrates model-free adaptive control with a multi-objective differential evolutionary algorithm. The proposed method eliminates explicit system modeling and predefined objective weights while providing a formal stability analysis. The adaptation process maintains predictable computational behavior as optimization complexity increases, supporting computation-aware runtime decision making in complex systems. The framework addresses multiple sources of runtime uncertainty, including internal system behavior, environmental conditions, and evolving requirements, using data-driven mechanisms such as autoencoder-based fault detection, random forest–based fault localization, and change-point detection. The effectiveness of the method is demonstrated through case studies across different application domains and benchmarking against two multi-objective control-theoretic adaptation methods. Experimental results demonstrate consistent satisfaction of system requirements, effective uncertainty handling, and scalable planning overhead, indicating applicability to adaptive systems operating under computational and runtime constraints.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid adaptation method integrating model-free adaptive control and evolutionary algorithm for self-adaptive systems

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Mahnaz Mohammadkhanpour Yamchi,
  • Eslam Nazemi,
  • Mitra Mirzarezaee

摘要

Control-theoretic adaptation provides a well-established foundation for managing self-adaptive software systems. Traditional model-based approaches incur costly system identification and exhibit reduced robustness under frequent runtime changes, limiting responsiveness and scalability in dynamic environments. Recent hybrid approaches combine machine learning with control theory to tune controller parameters, but many still lack formal assurances for requirement satisfaction and mechanisms for managing runtime uncertainty. This paper introduces a hybrid adaptation framework that integrates model-free adaptive control with a multi-objective differential evolutionary algorithm. The proposed method eliminates explicit system modeling and predefined objective weights while providing a formal stability analysis. The adaptation process maintains predictable computational behavior as optimization complexity increases, supporting computation-aware runtime decision making in complex systems. The framework addresses multiple sources of runtime uncertainty, including internal system behavior, environmental conditions, and evolving requirements, using data-driven mechanisms such as autoencoder-based fault detection, random forest–based fault localization, and change-point detection. The effectiveness of the method is demonstrated through case studies across different application domains and benchmarking against two multi-objective control-theoretic adaptation methods. Experimental results demonstrate consistent satisfaction of system requirements, effective uncertainty handling, and scalable planning overhead, indicating applicability to adaptive systems operating under computational and runtime constraints.