Elsevier

Computer Networks

Volume 187, 14 March 2021, 107791
Computer Networks

Non-parametric models with optimized training strategy for vehicles traffic flow prediction

https://doi.org/10.1016/j.comnet.2020.107791Get rights and content

Abstract

With the rapid development of information technology at the beginning of the 21st century, the traditional transportation system is rapidly transforming to Intelligent Transportation System (ITS). Meanwhile, as a powerful optimization tool for the applications’ performance under the framework of ITS, the traffic prediction model has gained much attention. The traffic prediction model belongs to the traditional time series prediction model, widely used in industry for business analysis, abnormalities detection, etc. There are two main categories of traffic prediction models — the parametric model and the non-parametric model. The non-parametric model develops rapidly in recent years due to the machine learning theory’s maturing and increment in computing power. Compared to the parametric model, the non-parametric model is more accurate and requires less advanced analysis of traffic patterns’ correlation, and can better handle a large amount of historical data. However, the non-parametric model also requires more powerful devices for training and implementation. The traffic prediction system’s implementation environment, built by on-road resources and off-road infrastructures, is relatively limited compared to the traditional data center. This paper uses several well-known state-of-the-art non-parametric models and their Deep Learning structures for traffic prediction, and evaluates the models’ performance on both freeway and urban-road dataset. Focusing on the model’s accuracy and training cost while deploying prediction models in a large-scale traffic network, this paper provides a novel optimized training strategy called CTS to reduce the implementation cost of a complex inner structure model. This approach could be further used to reduce the deployment cost, especially the training time, for the big data intelligent system using machine learning.

Introduction

With the rapid development of information technology such as the 5G communication system and on-board computer, many investigations have focused on the Intelligent Transportation System (ITS) [1]. The ITS makes full use of new-generation information technologies such as the Internet of Things (IoT), cloud computing, mobile internet, etc., to optimize the resource allocation capabilities in the transportation system, improve the industry management capabilities and promote public service capabilities [2], [3].

As an essential part of ITS, the traffic prediction system also caused extensive concern from academia and industry. The traffic prediction system can help the user better organize the road resources for the applications under the ITS framework and improve the resource utilization rate [4]. For example, Vehicular Cloud, as one of the well-known ITS application, is built by using underutilized on-road resources, including vehicular computing and storage resources gathered from a group of on-road vehicles, as well as Vehicular to Vehicular (V2V) and Vehicular to Infrastructure (V2I) communication networks. VC can be used for drive routing, traffic light control, or provide on-road entertainment services. However, the on-road resources’ high mobility makes the task arrangement harder in a VC than the traditional center cloud. By using the traffic prediction system, the system can predict the number of potential vehicle candidates before building up a temporary VC, determine the cloud’s acceptable workload, and arrange the task assigned to the cloud [5].

The core of a traffic prediction system is an accurate and efficient traffic prediction model. The traffic prediction model belongs to the traditional time series prediction model, widely used in industry for business analysis, abnormalities detection, etc. Compared to other types of time-series data (such as sales figures), traffic data (such as traffic flow or traffic speed) shares some same macroscopic features. For example, they maintain a stationary seasonal vibration or a specific trend under long-term observation and have small uncertainty vibration in the short-term. On the other hand, the main difference between traditional time-series data and traffic data is that natural patterns and humans impact the traffic situation, bringing more nonlinear characteristics to the series and, consequently, more difficulties to the prediction. Meanwhile, different from sales figure prediction, an on-road traffic prediction system, serving applications such as VC, is required to make a short-term online prediction to engage with the implementation environment’s mobility. Moreover, the random factors and uncertain time-varying features among short-term traffic data will significantly affect the prediction model’s performance.

The proposed traffic prediction models can be broadly divided into two categories — parametric models and non-parametric models [6]. Parametric models, such as ARIMA and SARIMA, have pre-assumption on the data distribution and estimate the model’s parameters by analyzing the historical data. On the contrary, non-parametric models are distribution-free and data-driven algorithms.

Unlike the parametric model, which needs prior knowledge to build up mathematics relationship of the traffic patterns, the Non-parametric model fit the traffic variation tendency by directly learning from a large amount of historical data, and the model’s parameters are adjusted based on the differences between the models’ output and target data via the training process. The non-parametric model develops rapidly in recent years. The reasons are the maturing of the machine learning theory and increment in computing power, and the enrichment of data collection methods. Moreover, the non-parametric model is more fixable to fit the uncertain features in short-term traffic data, and the improving dataset makes sure the model can be well trained.

However, as we can see from the recent research, to get a higher accuracy of the model, the structure of the model becomes more and more complicated. For example, the Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) is more complicated than the traditional RNN module, which only has one gate structure. Thus, LSTM needs more time to converge in the training process. Meanwhile, introducing deep learning (DL) structure to an ML-based model can also improve the model’s performance, but it also raises the hardware requirement and implementation cost. It is costly to deploy high computing equipment to the roadside infrastructures to fit a complex prediction model’s needs. A traffic prediction model needs to balance its accuracy with its implementation cost, which can be seen from its training and predicting time.

This paper will build several well-known state-of-the-art non-parametric models and their Deep Learning structures for traffic prediction and evaluate their efficiency and scalability on both freeway and urban-road datasets. Furthermore, we try to balance the relationship between accuracy and efficiency by providing a novel optimized training strategy called CTS, reducing a prediction model’s implementation cost with a complex inner structure.

This paper is organized as follows: Section 2 will review several well-known non-parametric models. Section 3 introduces the deep-learning structures for traffic prediction tasks; Moreover, we will also provide a new optimization method to speed up the model’s training process, which is important for real-world implementation. Section 4 introduces the two datasets and metrics for evaluation. At the end of this paper, we will summarize and provide some ideas for future works.

Section snippets

Related works

This section will focus on several well-known non-parametric models and their application in traffic prediction task. Meanwhile, to give a comprehensive overview of the models used for traffic prediction, we will also give a short discussion on the parametric prediction models. At the end of the section, we will review the literature about the big data system’s development status.

Deep learning structures and useful add-ons

In the previous section, we have reviewed several well-known non-parametric models used for the traffic prediction task. To further improve those models’ accuracy, scientists introduced deep-learning (DL) structures to non-parametric models. By using the DL structure, the model can capture higher dimensional and more abstract features from the dataset. Meanwhile, some useful add-ons can also enhance the overall performance of the model. This section will review some DL structures and useful

Experiments

This section will give a comprehensive analysis of the non-parametric models’ performance and the deep-learning structures we have introduced in previous sections. The performance of a model includes the accuracy of the model and the efficiency of the model. Moreover, we will further discuss the cost of using deep-learning structures in traffic prediction tasks. In the end, we will give a short analysis of the Desensitization algorithm’s performance.

Conclusion and future work

This paper reviewed and tested several well-known non-parametric traffic prediction models and their deep-learning structures on traffic prediction tasks both on the freeway and urban areas. In our experiment, we compared the accuracy and the implementation time cost for the base models and the deep-learning structures. From the experiment results, we can see that ML-based models and their deep-learning structure have an advantage in prediction accuracy, especially while the model has

CRediT authorship contribution statement

Jiahao Wang: Conception and design of study, Acquisition of data, Analysis and/or interpretation of data, Writing - original draft, Writing - review & editing. Azzedine Boukerche: Conception and design of study, Acquisition of data, Analysis and/or interpretation of data, Writing - original draft, Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work is partially supported by NSERC-SPG, NSERC-DISCOVERY, Canada Research Chairs Program, NSERC-CREATE TRANSIT Funds . All authors approved the version of the manuscript to be published.

Jiahao Wang is currently a master student of computer science in Paradise Research Laboratory at the University of Ottawa. He received the B.Eng. degree in computer science from Sichuan University, Chengdu, Sichuan, China, in 2017. His current research interests includes traffic flow prediction, vehicular cloud.

References (49)

  • N. Aljeri, A. Boukerche, An adaptive traffic-flow based controller deployment scheme for software-defined vehicular...
  • T. Zhang, R.E.D. Grande, A. Boukerche, Vehicular cloud: Stochastic analysis of computing resources in a road segment,...
  • SunP. et al.

    Machine learning-based models for real-time traffic flow prediction in vehicular networks

    IEEE Netw.

    (2020)
  • ZhaoJ. et al.

    High-order Gaussian process dynamical models for traffic flow prediction

    IEEE Trans. Intell. Transp. Syst.

    (2016)
  • MirZ.H. et al.

    An adaptive Kalman filter based traffic prediction algorithm for urban road network

  • SunP. et al.

    A fast vehicular traffic flow prediction scheme based on fourier and wavelet analysis

  • VapnikV.

    The Nature of Statistical Learning Theory

    (2013)
  • A. Ding, X. Zhao, L. Jiao, Traffic flow time series prediction based on statistics learning theory, in: Proc. IEEE...
  • ZhangY. et al.

    Forecasting of short-term freeway volume with v-support vector machines

    Transp. Res. Rec.

    (2007)
  • DorigoM. et al.

    Ant system: Optimization by a colony of cooperating agents

    IEEE Trans. Syst. Man Cybern. B

    (1996)
  • MinghengZ. et al.

    Accurate multisteps traffic flow prediction based on SVM

    Math. Probl. Eng.

    (2013)
  • ÇetinerB.G. et al.

    A neural network based traffic-flow prediction model

    Math. Comput. Appl.

    (2010)
  • Y. Tian, L. Pan, Predicting short-term traffic flow by long short-term memory recurrent neural network, in: IEEE...
  • J. Guo, Z. Wang, H. Chen, On-line multi-step prediction of short term traffic flow based on GRU neural network, in:...
  • Cited by (4)

    • ANN-Based traffic volume prediction models in response to COVID-19 imposed measures

      2022, Sustainable Cities and Society
      Citation Excerpt :

      Recently, artificial intelligence models based on Artificial Neural Networks (ANN), Random Forest (RF), Support Vector Machines (SVM) have gained popularity in predicting traffic volumes due to their improved prediction accuracy (Boukerche and Wang, 2020, Abu-Lebdeh, et al., 2016) and suitability for non-recurrent conditions (Chikaraishi et al., 2020). They do not require advanced analysis of traffic patterns and are able to handle large amounts of historical data (Wang and Boukerche, 2021). Support vector regression was used to predict AADT (Castro-Neto et al., 2009) and traffic in case of poor weather (Bao et al., 2021).

    • Traffic Prediction with Deep Learning for Vehicular Communication Network

      2022, Proceedings of SPIE - The International Society for Optical Engineering

    Jiahao Wang is currently a master student of computer science in Paradise Research Laboratory at the University of Ottawa. He received the B.Eng. degree in computer science from Sichuan University, Chengdu, Sichuan, China, in 2017. His current research interests includes traffic flow prediction, vehicular cloud.

    Azzedine Boukerche (FIEEE, FEiC, FCAE, FAAAS) is a Distinguished University Professor and holds a Canada Research Chair Tier-1 position with the University of Ottawa. He is founding director of the PARADISE Research Laboratory and the DIVA Strategic Research Center, and NSERC-CREATE TRANSIT at University of Ottawa. He has received the C. Gotlieb Computer Medal Award, Ontario Distinguished Researcher Award, Premier of Ontario Research Excellence Award, G. S. Glinski Award for Excellence in Research, IEEE Computer Society Golden Core Award, IEEE CS-Meritorious Award, IEEE TCPP Leaderships Award, IEEE ComSoc ComSoft and IEEE ComSoc ASHN Leaderships and Contribution Award, and University of Ottawa Award for Excellence in Research. He serves as an Editor-in-Chief for ACM ICPS and Associate Editor for several IEEE transactions and ACM journals, and is also a Steering Committee Chair for several IEEE and ACM international conferences. His current research interests include sustainable sensor networks, autonomous and connected vehicles, wireless networking and mobile computing, wireless multimedia, QoS service provisioning, performance evaluation and modeling of large-scale distributed and mobile systems, and large scale distributed and parallel discrete event simulation. He has published extensively in these areas and received several best research paper awards for his work. He is a Fellow of IEEE, a Fellow of the Engineering Institute of Canada, the Canadian Academy of Engineering, and the American Association for the Advancement of Science.

    View full text