IEEE/ACM International Symposium on Quality of Service (IWQoS) 2021
Session Short Paper Session 2
Learning-based Approaches
Conference
5:40 PM
—
7:05 PM JST
Local
Jun 26 Sat, 1:40 AM
—
3:05 AM PDT
Can Online Learning Increase the Reliability of Extreme Mobility Management?
Yuanjie Li (Tsinghua University, China); Esha Datta (University of California, Davis, USA); Jiaxin Ding (Shanghai Jiao Tong University, China); Ness B. Shroff (The Ohio State University, USA); Xin Liu (University of California Davis, USA)
0
The demand for seamless Internet access under extreme user mobility, such as on high-speed trains and vehicles, has become the norm rather than an exception. However, state-of-the-art mobile networks, such as 4G LTE and 5G NR, cannot reliably satisfy this demand. Our empirical study over operational LTE traces shows that 5.5%-12.6% of LTE handovers fail on high-speed trains at 200-350 km/h, which results in repetitive user-perceived network service disruptions. A root cause is the exploration-exploitation tradeoff for QoS during extreme mobility: the 4G/5G mobility management has to balance the exploration of more measurements for satisfactory handover and the exploitation for timely handover before the fast-moving user leaves the serving base station's coverage.
In this paper, we formulate the exploration-exploitation tradeoff in extreme mobility as a composition of two online learning problems. Then we present BaTT, a multi-armed bandit-based online learning solution for both problems. BaTT uses ɛ-binary-search to optimize the threshold of a serving cell's signal strength to initiate the handover with a provable O(log J log T ) regret. We also devise an opportunistic Thompson sampling algorithm to optimize the sequence of target cells measured for reliable handovers. BaTT can be readily implemented using the recent Open Radio Access Network (O-RAN) framework in operational 4G LTE and 5G NR. Our analysis and empirical evaluations over a dataset from operational LTE networks on the Chinese high-speed rails show a 29.1% handover failure reduction at the speed of 200-350 km/h.
In this paper, we formulate the exploration-exploitation tradeoff in extreme mobility as a composition of two online learning problems. Then we present BaTT, a multi-armed bandit-based online learning solution for both problems. BaTT uses ɛ-binary-search to optimize the threshold of a serving cell's signal strength to initiate the handover with a provable O(log J log T ) regret. We also devise an opportunistic Thompson sampling algorithm to optimize the sequence of target cells measured for reliable handovers. BaTT can be readily implemented using the recent Open Radio Access Network (O-RAN) framework in operational 4G LTE and 5G NR. Our analysis and empirical evaluations over a dataset from operational LTE networks on the Chinese high-speed rails show a 29.1% handover failure reduction at the speed of 200-350 km/h.
SuperClass: A Deep Duo-Task Learning Approach to Improving QoS in Image-driven Smart Urban Sensing Applications
Yang Zhang, Ruohan Zong, Lanyu Shang, Md Tahmid Rashid and Dong Wang (University of Notre Dame, USA)
0
Image-driven smart urban sensing (ISUS) has emerged as a powerful sensing paradigm to capture abundant visual information about the urban environment for intelligent city monitoring, planning, and management. In this paper, we focus on a Classification and Super-resolution Coupling (CSC) problem in ISUS applications, where the goal is to explore the interdependence between two critical tasks (i.e., classification and super-resolution) to concurrently boost the Quality of Service (QoS) of both tasks. Two fundamental challenges exist in solving our problem: 1) it is challenging to obtain accurate classification results and generate high-quality reconstructed images without knowing either of them a priori; 2) the noise embedded in the image data could be amplified infinitely by the complex interdependence and coupling between the two tasks. To address these challenges, we develop SuperClass, a deep duo-task learning framework, to effectively integrate the classification and super-resolution tasks into a holistic network design that jointly optimizes the QoS of both tasks. The evaluation results on a real-world ISUS application show that SuperClass consistently outperforms state-of-the-art baselines by simultaneously achieving better land usage classification accuracy and higher reconstructed image quality under various application scenarios.
SeqAD: An Unsupervised and Sequential Autoencoder Ensembles based Anomaly Detection Framework for KPI
Na Zhao, Biao Han and Yang Cai (National University of Defense Technology, China); Jinshu Su (National University of Defence Technology, China)
0
Key Performance Indicator (KPI), a kind of time-series data, its anomalies are the most intuitive characteristics when failures occurred in IT systems. KPI anomaly detection is increasingly critical to provide reliable and stable services for IT systems. Unsupervised learning is a promising method because of lacking labels and the unbalance in KPI samples. However, existing unsupervised KPI anomaly detection methods suffer from high false alarm rates. They handle KPI sequence as non-sequential data and ignore the time information, which is an essential KPI character. To this end, in this paper, we propose an unsupervised and sequential autoencoder ensembles based anomaly detection framework called SeqAD. SeqAD inherits the advantages both from the sequence-to-sequence model and autoencoder ensembles. SeqAD reduces the KPI over-fitting problem effectively by introducing autoencoder ensembles. In order to better capture the time information of KPI, we propose a random step connection based recurrent neural network (RSC-RNN) to train the KPI sequence, which can provide random connections to construct autoencoders with different structures and retain time information to the most extent. Extensive experiments are conducted on two public KPI data-sets from real-world deployed systems to evaluate the efficiency and robustness of our proposed SeqAD framework. Results show that SeqAD is able to smoothly capture most of the characteristics in all KPI data-sets, as well as to achieve a high F1 score between 0.93 and 0.98, which is better than the state-of-art unsupervised KPI anomaly detection methods.
DeepDelivery: Leveraging Deep Reinforcement Learning for Adaptive IoT Service Delivery
Yan Li, Deke Guo and Xiaofeng Cao (National University of Defense Technology, China); Feng Lyu (Central South University, China); Honghui Chen (National University of Defense Technology, China)
0
To enable fast content delivery for delay-sensitive applications, large content providers build edge servers, Points of Presence (PoPs), and datacenters around the world. They are networked together as an integrated infrastructure via a private wide-area network (WAN), named content delivery network (CDN). To deliver quality services in the CDN, there are two critical decisions that should be properly made: 1) making assignments of PoP and datacenter for user requests, and 2) selecting routing paths from PoP to datacenter. However, with both the network variability and CDN environment complexity, it is challenging to achieve satisfying decisions. In this paper, we propose DeepDelivery, an adaptive deep reinforcement learning approach to intelligently make assignments and routing decisions in real time. Essentially, DeepDelivery adopts the Markov decision process (MDP) model to capture the dynamics of network variation, and the objective is to jointly maximize the infrastructure utilization of providers and minimize the total latency of end users. We conduct extensive trace-driven evaluations spanning various environment dynamics with both real-world and synthetic trace data. The result demonstrates that DeepDelivery can outperform the state-of-the-art scheme by 21.89% higher utilization and 11.27% lower end-to-end latency on average.
LCL: Light Contactless Low-delay Load Monitoring via Compressive Attentional Multi-label Learning
Xiaoyu Wang, Hao Zhou, Nikolaos M. Freris and Wangqiu Zhou (University of Science and Technology of China, China); Xing Guo (Anhui University, China); Zhi Liu (The University of Electro-Communications, Japan); Yusheng Ji (National Institute of Informatics, Japan); Xiang-Yang Li (University of Science and Technology of China, China)
0
Fine-grained energy consumption analysis has great potential value in applications of Smart Grids, renewable energy, and Artificial Intelligence of Things. Non-Intrusive Load Monitoring (NILM) is a single-sensor alternative to the conventional one-sensor-for-one-appliance solution due to its ability to deduce individual appliances states from mixed measurements from the main power interface. Despite its advantages of low cost and easy maintenance, a few drawbacks hinders its widespread adoption. To enhance the Quality of Service (QoS) of NILM, four objectives should be achieved by careful designing: high accuracy, user transparency, low response delay, and low data redundancy.
Inspired by observations of discriminative yet redundant current waveform and model sparsity, we propose LCL, a light-weight, contactless, plug-and-play solution for real-time load monitoring. The filtering module skips over unchanged input and compresses the measurements of interest using Compressed Sensing. The reconstruction-free inference module runs an attentional multi-label classification and returns all functioning appliance states directly from the compressed input. The compression module leverages model sparsity for real-time processing on edge devices. Evaluations based on our prototype deployed in real-life scenarios attest to the high QoS of LCL with a subset accuracy of 94.2% and a delay reduction of 52.2%. Our solution further filters out 96.8% of the redundant input and attains a Measurement Rate of 0.1 without noticeable impact on the performance.
Inspired by observations of discriminative yet redundant current waveform and model sparsity, we propose LCL, a light-weight, contactless, plug-and-play solution for real-time load monitoring. The filtering module skips over unchanged input and compresses the measurements of interest using Compressed Sensing. The reconstruction-free inference module runs an attentional multi-label classification and returns all functioning appliance states directly from the compressed input. The compression module leverages model sparsity for real-time processing on edge devices. Evaluations based on our prototype deployed in real-life scenarios attest to the high QoS of LCL with a subset accuracy of 94.2% and a delay reduction of 52.2%. Our solution further filters out 96.8% of the redundant input and attains a Measurement Rate of 0.1 without noticeable impact on the performance.
GreenTE.ai: Power-Aware Traffic Engineering via Deep Reinforcement Learning
Tian Pan (Beijing University of Posts and Telecommunications, China); Xiaoyu Peng (BUPT, China); Qianqian Shi (Beijing University of Posts and Telecommunications, China); Zizheng Bian (BUPT, China); Xingchen Lin and Enge Song (Beijing University of Posts and Telecommunications, China); Fuliang Li (Northeastern University, China); Yang Xu (Fudan University, China); Tao Huang (Beijing University of Posts and Telecommunications, China)
0
Power-aware traffic engineering via coordinated sleeping is usually formulated into Integer Programming (IP) problems, which are generally NP-hard thus the computation time is unbounded for large-scale networks. This results in delayed control decision making in highly dynamic environments. Motivated by advances in deep Reinforcement Learning (RL), we consider building intelligent systems that learn to adaptively change router/switch's power state according to varying network conditions. The forward propagation property of neural networks can greatly speed up power on/off decision making. Generally, conducting RL requires a learning agent to iteratively explore and perform the ``good'' actions based on the feedback from the environment. By coupling Software-Defined Networking (SDN) for performing centrally calculated actions to the environment and In-band Network Telemetry (INT) for collecting underlying environment feedback, we develop GreenTE.ai, a closed-loop control/training system to automate power-aware traffic engineering. Furthermore, we propose numerous novel techniques to enhance the learning ability and reduce the learning complexity. Considering both energy efficiency and traffic load balancing, GreenTE.ai generates near-optimal power saving actions within 276ms under a network testbed of 11 software P4 switches.
Session Chair
Yanjiao Chen, Zhejiang University, China
Made with in Toronto · Privacy Policy · IWQoS 2020 · © 2021 Duetone Corp.