Service Selection for Web Services with Probabilistic QoS

TLDR

Web services are defined by functional and non‑functional (QoS) properties, and while many services share the same function, they differ in QoS, which is a key criterion for selection, yet prior work has treated QoS as fixed values, ignoring its inherent randomness due to prediction uncertainty. The study aims to select atomic services for a composite service so that the probability of satisfying the composite’s QoS constraints is high while keeping execution time acceptable. To achieve this, the authors model each service’s QoS as a discrete random variable with a probability mass function, then iteratively refine an initial assignment using simulated annealing. Experimental results indicate that this probabilistic approach outperforms earlier integer‑programming and cost‑driven methods.

Abstract

Web services can be specified from two perspectives, namely functional and non-functional properties. Multiple services may possess the same function while vary in their non-functional properties, or called quality-of-service (QoS). QoS values are important criteria for service selection or recommendation. Most of the former works in web service selection and recommendation treat the QoS values as constants. However, QoS values of a service as perceived by a given user are intrinsically random variables because QoS value prediction can never be precise and there are always some unobserved random effects. In this work, we address the service selection problem by representing services' QoS values as discrete random variables with probability mass functions. The goal is to select a set of atomic services for composing a composite service such that the probability of satisfying constraints imposed on the composite service is high and the execution time is reasonable. Our proposed method starts with an initial web service assignment and incrementally adjusts it using simulated annealing. We conduct several experiments and the results show that our approach generally performs better than previous works, such as the integer programming method and the cost-driven method.