Achieving Self-Management Capabilities in Autonomic Systems using Case-Based Reasoning

Abstract

Autonomic systems promise to inject self-managing capabilities in software systems. The major objectives of autonomic computing are to minimize human intervention and to enable a seamless self-adaptive behavior in software systems. To achieve self-managing behavior, various methods have been exploited in the past. Case- Based Reasoning (CBR) is a problem solving paradigm of artificial intelligence which exploits past experience, stored in the form of problem-solution pairs. Although CBR has been applied in the externalization architecture of self-healing systems at a limited scale, however it has not been fully exploited in autonomic systems in general. We have proposed and applied CBR to achieve autonomicity in software systems. The proposed approach has been described and evaluated on CBR implementation for externalization and internalization architectures of autonomic systems. The study highlights the effect of ten different similarity measures, the role of adaptation and the effect of changing nearest neighborhood cardinality for a CBR solution cycle in autonomic managers. The results show that the proposed CBR based autonomic systems exhibit 90 to 98% accuracy in diagnosing the problem and planning the solution. The learning process improves as more experience is added to the case-base. This results in a larger case-base. A larger case-base reduces the efficiency in terms of computational cost. To overcome this efficiency problem, this research work suggests to cluster the case-base, classify the reported problem in the appropriate cluster and devise the solution. This approach reduces the search complexity by confining a new case to a relevant cluster in the case-base. Clustering the case-base is a one-time process and does not need to be repeated regularly. The proposed approach has been outlined in the form of a new clustered CBR framework. The comparison of performance of the conventional CBR approach and clustered CBR approach has been presented in terms of their Accuracy, Recall and Precision (ARP) and computational efficiency. The proposed approach exhibits up to 90% accuracy. It indicates that the performance does not degrade using clustered CBR approach in terms of accuracy and at the same time, it improves the time complexity of the retrieval process. As the case-base grows in size, it is partitioned into different clusters in order to improve the retrieval efficiency. Deciding an appropriate number of clusters for a case-base is not a trivial problem. This research work proposes an approach to cluster the case-base into a random number of clusters. Two versions of the randomized approach have been presented. One of them guarantees success but its computational cost is a function of random variable. Other approach guarantees a deterministic computational cost but the success is not guaranteed. In order to ensure the retrieval time, a binary search based retrieval strategy has also been proposed. Randomized approach guarantees the same level of accuracy as in case of the clustered CBR approach and simplifies the clustering process by reducing its time complexity. The proposed approaches have been implemented on Rice University Bidding Sys- tem (RUBiS) and a simulation study of Autonomic Forest Fire Application (AFFA). Their theoretical and empirical results have been compared. The statistical analysis shows that the empirical and theoretical results are significantly similar.