Evolutionary approaches to signal decomposition in an application service management system

The increased demand for autonomous control in enterprise information systems has generated interest on efficient global search methods for multivariate datasets in order to search for original elements in time–series patterns, and build causal models of systems interactions, utilization dependencies, and performance characteristics. In this context, activity signals deconvolution is a necessary step to achieve effective adaptive control in Application Service Management. The paper investigates the potential of population–based metaheuristic algorithms, particularly variants of particle swarm, genetic algorithms and differential evolution methods, for activity signals deconvolution when the application performance model is unknown a–priori. In our approach, the Application Service Management System is treated as a black– or grey– box, and the activity signals deconvolution is formulated as a search problem, decomposing time–series that outline relations between action signals and utilization–execution time of resources. Experiments are conducted using a queue– based computing system model as a test–bed under different load conditions and search configurations. Special attention was put on high–dimensional scenarios, testing effectiveness for large–scale multivariate data Tomasz D. Sikora · George D. Magoulas Department of Computer Science and Information Systems Birkbeck, University of London Malet Street, London WC1E 7HX Tomasz D. Sikora Phone: +44 20-3287-8256, +48 787-622-787 E-mail: [email protected] George D. Magoulas Phone: +44 20-7631-6717 E-mail: [email protected] analyses that can obtain a near–optimal signal decomposition solution in a short time. The experimental results reveal benefits, qualities and drawbacks of the various metaheuristic strategies selected for a given signal deconvolution problem, and confirm the potential of evolutionary–type search to effectively explore the search space even in high–dimensional cases. The approach and the algorithms investigated can be useful in support of human administrators, or in enhancing the effectiveness of feature extraction schemes that feed decision blocks of autonomous controllers.