Autonomic Forecasting Method Selection: Examination and Ways Ahead

Abstract

Proactive adaptation improves the system performance of Autonomic Computing systems as it recognizes adaptation concerns in advance and adapts or prepares adaptation accordingly. To support this, forecasting methods use historical data to predict future system states. According to the "No-Free-Lunch-Theorem", there is no general forecasting method that performs best in all scenarios. Usually at design time, expert knowledge is required to decide on the forecasting method based on the anticipated characteristics of the resulting time series data. The uncertainty that results from the gap between design time and runtime for adaptive systems, as well as the environmental uncertainty at runtime, decreases the possibility that a forecasting method chosen at design time can cope with runtime demands. A common approach to tackle this problem is to use recommendation systems that automatically choose the forecasting methods. In this paper, we introduce a novel approach for forecasting method selection and a recommendation-based ensemble forecasting approach. We compare our approaches with one of the most widely used recommendation approaches for time series forecasting. Whereas the reference system uses static recommendation rules, we contrast a modified version which supports dynamic rule learning. The results of the evaluation show that our approaches outperform the original approach with static rule learning.