Article Preview
Top1. Introduction
Wireless Sensor Network (WSN) is conducted to achieve interoperability between Internet and physical world (Gubbi, 2013, Bellavista, 2012, Alex, 2016), which has been imported to adapt business processes and underlying software infrastructure to changes on things quickly and flexibly. Meanwhile, determining how to build high-reliability WSN Service Applications (WSAs) becomes a critical issue with wide-spread WSN services.
Software reliability is the probability of failure-free software operation for the specified period of time in a specified environment (Radatz, 1990). In traditional software reliability engineering, there are several approaches to building reliable software systems. Now, large-scale IOT (Internet of Things) applications lead to that building reliable system is exceedingly difficult. In case of IOT service, user-observed reliability not only relies on system, but also heavily depends on remote services and user characteristics (e.g., geographic locations, network conditions, operational profiles (Musa, 1993), etc.). Influenced by the unpredictable connections and many other environmental and operational factors, different service users may experience different reliability performance on the same WSN service. We can improve the reliability of service-oriented systems using fault-tolerate WSN services, which employ functionally equivalent yet independent services. To optimally select reliable WSN services (Wang, 2016), we must be able to predict the reliability rankings of WSN services. The ability to predict reliability rankings of invoked WSN services early at design phase can help to reduce maintenance cost and to produce more reliable service-oriented systems. Therefore, reliability ranking prediction of WSN services is crucial for the WSAs.
It is difficult for present reliability ranking prediction approaches to perform excellently, because we retrieve different rankings with different preference functions.
We propose a constrained learning approach to the prediction of reliability ranking for WSN services (CLPRR). The approach utilizes history logs of conducting real-world WSN service evaluation at client-side to predict reliability rankings of WSN services. Our approach serves to achieve higher prediction accuracy of reliability ranking prediction and improve the performance by pruning candidate services.
The contributions of this paper can be summarized as follows:
- 1.
We propose a constrained learning prediction of reliability ranking approach for WSN services, so that it can achieve higher prediction accuracy and improve the performance by pruning candidate services;
- 2.
We conduct comprehensive experiments on a real-world dataset, demonstrating the effectiveness of our approach.
The remainder of this paper is organized as follows: Section 2 describes a motivating Scenario. Section 3 presents our approach. Section 4 describes our experiments in detail. Section 5 discusses the related works, and Section 6 concludes the paper.
Top2. A Motivating Scenario
An approach to reliability ranking prediction of WSN services needs to record failure probability of invoked services at client-side, where the data can be recorded with little overhead and without compromising the functionality and usefulness of WSA. When an invocation happens, client-side should store records in logs. Then, the records are encoded and transferred to a proxy-server. This process can be made transparent to client and does not negatively influence system performance. Moreover, records logs convey only partial information of all services. Thus, user-service matrix of failure probability, which is constructed based on logs, is sparse, as shown in Figure 1.