Modeling 5G Wireless Network Service Reliability Prediction with Bayesian Network Bilgehan Erman Bell Labs, Nokia 600 Mountain Avenue Murray Hill, New Jersey 07974 Email: bilgehan. erman@nokia.com Simon Yiu Bell Labs, Nokia 600 Mountain Avenue Murray Hill, New Jersey 07974 Email: simon. yiu@nokia.com Abstract—In 5G networks, prediction of service reliability is critical because of strict service performance requirements. In this work, we present a wireless service model for reliability computation, and we use Bayesian network (BN) to compactly represent the joint probability distribution. Furthermore, we use the model to predict network service reliability and infer hidden states of a network. Our approach provides a promising direction for modeling network service reliability and insight in designing next generation networks which comply with high service quality requirements. I. I NTRODUCTION With fifth generation (5G) wireless networks, the quantifi- able level of quality will be an integral part of a service. Therefore, not operating within the boundaries of specific KPIs (key performance indicators) will result in a service to be completely unusable. As of this writing, desired 5G services create unprecedented requirements for service scenarios with composite-performance-indicator bounds at highly elevated operational conditions. For example, ultra-high reliability and ultra-low latency scenario demands end-to-end latency under 10 ms, downlink/uplink bandwidth up to 10 Mbps, mobility up to 500 km/h, at 99.999% reliability. Some applications that this scenario would be applicable can be: automated traffic control, collaborative robots, eHealth: Extreme Life Critical, and remote object manipulation: remote Surgery [1]. A comprehensive coverage of the future of wireless access services can be found in [2] The key challenge with the strict 5G service requirements will be to predict the reliability of the service. In this study we address this challenge by first describing how the wireless network can be modeled for service reliability prediction and then by describing how Bayesian network (BN) representation can be applied to provide a feasible computational process. Our analysis only considers wireless services delivered by structured cellular networks, primarily the 5G cellular networks. Although the wireless service requirements are not as stringent, the analysis may be equally applied to the forth generation (4G) networks as well. In conjunction with the multi-RAT scenarios, the network model can also be extended with WiFi networks as a complementary factor that can increase the reliability of wireless services [3]. 978-1-4673-8626-5/16/$31.00 c 2016 IEEE In the subsequent sections, we first provide a background on reliability and BNs. Then the model description is given for wireless network services. Application of BN representation has two sections, one that provides an algorithm to populate large BN graphs, and the second one that describes a learning process that is used to populate BN graphs from wireless network data. II. BACKGROUND In this section, the concept of network service reliability is first introduced. We then review the various aspects of BNs followed by comments on the challenges of constructing BNs. A. Network Service Reliability Definitions of the reliability concepts may differ slightly based on the context. Here, instead of a system perspective, the context is based on the use of network services. From this point of view, we use the following definitions for the concepts of failure, failure rate, reliability, availability, and resiliency. Suppose T is the time to failure or failure time and is modeled as a random variable (r. v.) with a certain density function f T (t). Unreliability or simply failure is defined as the probability that the system will fail by time t and can be computed as follows: F T (t)  P (T ≤ t)=  t 0 f T (s)ds. (1) In other words, failure F T (t) is simply the cumulative distri- bution function (CDF) of the r. v. T . In the wireless service context, failure is the event that a user may not start or continue to use a started network service after time t for reasons related to network functions and configurations (excluding other operational and maintenance causes.) On the other hand, reliability R T (t) is defined as the probability that the system will not fail from time 0 to time t: R T (t)=1 - F T (t)=1 -  t 0 f T (s)ds =  ∞ t f T (s)ds. (2) Availability A(t) is the probability of a network service being ready for use from time 0 to time t. Repair time that is defined for equipment reliability does not apply to the network