Machine Learning and Deep Learning Models for Trac Flow Prediction: A Survey

Trac congestion is one of the problems for cities around the world due to the rapid increasing of vehicles in urbanization. Trac ow prediction is of a great importance for Intelligent Transport System (ITS) which helps to optimize the trac regulation of a transportation in the city. Nowadays, several researches have been studied so far on trac ow prediction, accurate prediction has not yet been exploited by most of existing studies due to the impact of inability to effectively deal with spatial temporal features of the times series data. Trac information in transportation system will also be affected by different factors. In this research we intended to study various models for Trac ow prediction on the basis machine learning and deep learning approaches. Factors affecting the performance of trac ow prediction intensity are studied as well. Benchmark performance evaluation metrics are also reviewed. Generally, this manuscript covers relevant methods and approaches, review the state-of-art works with respect to different trac ow prediction technique help researchers in exploring future directions so as to realize robust trac ow prediction. bench mark performance Various machine learning and deep learning models for been

reality in the near future, as industries, universities, and governments around the world devote signi cant efforts and resources to the development of safer vehicles and infrastructure for road transport.These investments can be veri ed through many national and international initiatives dedicated to vehicular networks [4] [5] [6].

Therefore, intelligent transport system is part of vehicular networks which can be applied in the smart city domain for tra c ow prediction.As suc , we can say that Intelligent transport systems (ITS) are applications of smart city based on the internet of things (IOT) and tra c ow prediction is applicable in the intelligent transport system eld.

Tra c problems have seriously affecting the urban societies quality of life as well as development of the city, and prediction of tra c congestion as a great importance for both individuals and the governments as well.However, understanding and modeling of the tra c ow conditions can be extremely challenging [7].

People are getting increasingly concerned about tra c congestion, which has seriously affected their life quality and urban development.To monitor r al-time tra c conditions, cities around the world have deployed embedding sensors, like inductive-loop detectors and video image processors in road networks [8].

The increasing availability of data and services has created opportunities to predict tra c conditions like predicting travel speed and tra c volume n the entire city [9].We are experiencing urbanization shift and it is predicted that more than 60% of the world's population will live in urban areas by 2050 [10].

Therefore, Tra c congestion is one of the major challenges to enable good urbanization and the deployment of Intelligent Transportation Systems (ITS) n urban areas brings the opportunities to prevent or reduce tra c congestion.

Tra c ow prediction serves as the key component of ITS to forecast and prevent tra c congestion, control and manage tra c e ciently, and plan the best traveling route.Machine learning (ML) based approaches such as k-nearest neighbor (KNN) algorithm [11], Markov process-based scheme [12] and Arti cial Neural Network (ANN) schemes [13] [14] are tra c ow predictions.When we compare those ML approaches, deep learning (DL) models have the advantages in simplifying data preprocessing procedure and outperforming other ML methods in terms of accuracy.Therefore, DL schemes have received extensive attention recently in tra c ow prediction [15] [16] [17] [18] [19] [20].

Furthermore, when we compare machine learning algorithms and deep learning models the later have an advantage in scaling with data availability which is eep neural networks usually make better use of massive amount of data by learning customized feature representations.In particular, convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent neural networks (RNNs) have demonstrated their peculiar advantages in modeling and predicting spatiotemporal data.Recently, such models have received extensive attention in the eld of tra c ow prediction [21] [22].Moreover, tra c ow prediction enables those travelers to optimize resources like time, gas oil, and power consumption and it is one of the most important aspect in smart city.


Motivation

Urban road tra c can be a problem on societies to environmental effects and health problem.As we know, Transportation is one of the most import

t way that
rban society in the world always uses daily to move from place to place.The quality of daily life for the society can be improved through smart city with e cient travel services.Intelligent transport system is one of the solutions to provide this type of service.Tra c ow prediction is very important for better performance of intelligent transport system in smart city domain.However, the prediction model which have been done so far by different researches are not adequate enough to provide better tra c ow prediction services.

There are various factors that have impacts on the performance of tra c ow prediction.As such, Study and analyze those factors helps to model better tra c ow pre iction scheme.Therefore, the motivation for this research is to study on various state-of-the art tra c ow prediction approaches so as to nd out approaches how to design accurate tra c ow prediction.This helps to improve the quality of tra c ow prediction service quality in intelligent transport system as long as the way how accurate tra c ow prediction is properly studied.Producing robust tra c ow prediction technique has a great importance for urban road users and government.Therefore, urban societies daily transportation lives will be improved.


Statement of the Problem

Urban development is becoming important for the modern society and the situation of urban tra c congestion is becoming very serious due to

mber of vehicles usage is
increasing in the city [23].One of the most important solution for this problem is the application of intelligent transport system (ITS) which is applied to predicts congestion of tra c ow.One of the most important part in smart city development is tra c ow prediction management system in most of the cities around the world.Therefore, providing scienti c prediction model of tra c congestion ensures the safety of tra c environment and good to prevent tra c congestion which helps to reduce tra c accidents.

The vehicles are increased in urban areas which lead to problems in the transportation system.Among problems consumption of fuel, air pollution, congestion and accident.
rom these issues, tra c congestion is the big problem in urban transportation.Better tra c ow prediction is needed to resolve the aforementioned issues.Therefore, tra c ow prediction is very important to handle tra c congestion problem and the quality of urban citizens daily life can be affected by tra c ow problems.

There are various tra c ow prediction researches that have been tried so far by various researchers to predict tra c ow.But there is no accurate prediction model yet.It is u able to effectively deal with various features of times series data results di cult to accurately represent tra cs ow.Therefore, extra investigation needs to have better prediction model to predict tra c ow better.In this research, we are intended to review various tra c ow prediction techniques to suggest researchers to design accurate tra c ow prediction.Accurate tra c ow prediction can be provided by studying and analyzing the most important factors that affect the tra c ow prediction.Therefore, we study on the factors affecting the performance of tra c ow prediction.As a result, transportation quality can be improved in urban areas to successfully deploy intelligent transport system.


Objectives


General Objective

The objective of this research is to review tra c ow prediction techniques to produce robust tra c ow prediction scheme.


Speci c Objective

• Det

and study on the s
ate-of-the art research


Methodology

With the aim of achieving the aforementioned research objectives the following met

dologies will be a
plied


Literature Review

Exhaustive study and explorations wi

be made on
he areas related to deep learning specially in the eld of intelligent transport system which is used in smart city.This will be accomplished by reading books, journals or conference papers which have been done so far with different approaches, so as to have su cient understanding on the problem.Techniques and approaches appropriate for development of various algorithms for modeling tra c ow prediction in intelligent transport system will also be investigated as well.After deep analysis on the existing models we suggest a new approach to achieve better tra c ow prediction in intelligent transport system.


Application Results

This research will signi cantly contribute to the smart city domain speci cally for the intelligent transport system for effective urban tra c ow prediction.Be

use, as long as the
ra c ow prediction become accurate the urban road users will be satis ed with the services like reduced tra c accidents, tra c congestion, travel time, and air pollution which decreases important impact on urban population health.This work will bene t the services of intelligent transport system applications and the most important application areas that will bene t from this work are the entire citizens who use transportation service in the city and the government who controls tra c ow in the city.

Chapter 2: Literature Review


Overview

Nowadays, deep learning is becoming popular and essential approach in arti cial intelligence discipline applied in several application areas.R searchers throughout the world are applying deep learning approach to nd out solutions to problems in various application areas.Among those several application areas, modeling tra c ow prediction is the most important application that can be produced with deep learning.This Chapter covers overview of intelligent transport systems and tra c ow prediction models in intelligent transport systems which have been done so far using different approaches.Number of tra c ow prediction models have been proposed because of its signi cant importance of tra c ow prediction.


Intelligent Transport System

Intelligent Transport System (ITS) is a technology or an application of intelligence in any form like management strategies, logistics, statistics, using pre

ctive human behavior, non-ele
tronic and electronic forms of dispensing information, advanced sensors, computers and communication technologies that improve the quality of transportation.ITS is a vast eld that encompasses driver assistance, vehicle tracking, license plate recognition, inter vehicle communication, air tra c management, road sign prediction, modeling and simulation, and intelligent tra c management and so on [24].ITS provides to the multidimensional needs and overlap of the transport eld and others, for example license plate detection for better policing, air tra c management for safe ight operations, tra c ow and tra c management for e cient use of existing infrastructure.The commonly used approach to increase vehicular population is upgrading the existing road infrastructure.However, not all problems are due to insu ciency of the road infrastructure but due to poor management of tra c ow and congestion control.Therefore, due to lack of proper management of road tra c, existing roads cannot be su cient enough specially in developing countries.

Intelligent Transportation Systems (ITSs) provides e cient services related to different modes of transport, making the transport networks smarter [25].ITSs are mainly applied in road transport but are also designed to offer interfaces with other modes of transport [26].Some of them involve surveillance of the roadways [27], others may have an important role in the context of urban development [28].ITSs are based on different technologies, which could vary from car navigation systems or tra c signal control systems to advanced applications integrating live data from other sources, such as parking guidance systems.Predictive techniques are designed to allow advanced modeling and comparison with historical data [29].

Intelligent Transportation Systems (ITS) is a technology that has just recently developed to overcome tra c congestion in several developed countries.ITS is used for computing systems and communica ion technology for various purposes, such as tra c management, routing planning, vehicle and road safety and emergency services [30].ITS uses various kinds of sensing and communication to help transport authorities and vehicle drivers in making informative decisions as well as comfort and safety in driving [31].The utilization of ITS increases road and vehicle security systems become more e cient and environmentally friendly [32].The use of wireless communication technology enables ITS to open opportunities for various types of road and user's safety applications.The ITS application utilizes data collected from vehicles to increase driver safety and rationalize public infrastructure use [33].


Tra c Flow Prediction

Tra c ow prediction has been regarded as a key functional component in ITSs [34].For the past few decades, a number of tra c ow prediction models have been developed to assist tr

c management and contr
l to improve the e ciency of transportation ranging from route guidance and vehicle routing.The objective of tra c ow prediction is to provide tra c ow information.

Tra c ow prediction has gained more attention with the rapid development and deployment of intelligent transportation systems (ITSs).It is regarded as a critical element for the successful deployment of I S subsystems, particularly advanced traveler information systems, advanced tra c management systems, advanced public transportation systems, and commercial vehicle operations.

From the researches which have been done so far, the autoregressive integrated moving average (ARIMA) model is used to predict short-term freeway tra c ow [35].Then, extensive variety of models for tra c o prediction have been proposed by researchers from different areas, such as transportation engineering, statistics, machine learning, control engineering, and economics.Tra c ow prediction approaches have been extensively researched in literature and generally can be grouped into three basic categories, such as parametric, non-parametric and simulation based models.Parametri models include time-series models [36], Kalman ltering models [37], etc. Non-parametric models include Support Vector Regression (SVR) methods [38], Arti cial Neural Networks (ANNs) [39], etc. Simulation based approaches use tra c simulation tools to predict tra c ow [40] Levin and Tsao applied Box-Jenkins time-series analyses to predict expressway tra c ow and found that the ARIMA (0, 1, 1) model was the most statistically signi cant for all forecasting [41].[42].Many variants of ARIMA were proposed to improve prediction accuracy, such as Kohonen ARIMA (KARIMA) [43], subset ARIMA [44], ARIMA with explanatory variables (ARIMAX) [45], vector autoregressive moving aver ge (ARMA) and space-time ARIMA [46], and seasonal ARIMA (SARIMA) [47].Except for the ARIMAlike time-series models, other types of timeseries models were also used for tra c ow prediction [48].


M. Hamed et al. applied an ARIMA model for tra c volume prediction in urban arterial roads

In the area of multi-interval tra c volume prediction Chang et al. proposed a k-nearest neighbor nonparametric regression (kNN-NPR) model [49].Even if time-series data values vary abruptly or show uctuations, th presented model shows effective accuracy.Although k-NN based methods cannot perform spatial and temporal modeling simultaneously [50].E. Faouzi developed a kernel smoother for the autoregression function to do short-term tra c ow prediction, in which functional estimation techniques were applied [51].Sun et al. used a local linear regression model for short-term tra c forecasting [52].A Bayesian network approach was proposed for tra c ow forecasting [53].An online learning weighted support vector regression (SVR) was presented for short-term tra c ow predictions [54].Various ANN models were developed for predicting tra c ow [55] [56] [57] [58] [59].

Although the deep architecture of NNs can learn more powerful models than shallow networks, existing NN-based methods for tra c ow prediction usually only have one hidden layer.It is hard to train a deeplayered hierarc ical NN with a gradient-based training algorithm.Recent advances in deep learning have made training the deep architecture feasible since the breakthrough of Hinton et al. [60], and these show that deep learning models have superior or comparable performance with state-of-the-art methods in some areas.

Generally, many tra c ow prediction algorithms have been developed because of the need for tra c ow prediction in ITS with different approaches in different application areas.The proposed models are designed with a small separate speci c tra c data, and the accuracy of tra c ow prediction approaches are dependent on the of features of tra c ow embedded in the collected spatiotemporal tra c data.But it is di cult to determine that one approach is clearly better than other methods in a certain state.Moreover, literature shows when using Neural Networks produces robust prediction but need to investigate more for accurate prediction model.The most important determinant factors should be studied which can be signi cantly produce robust tra c ow prediction model.

Finally, the challenge for tra c ow prediction is the presence of unusual factors such as accidents, weather, and planned events.For example, if the rainfall intensity increases, then both speed and ow decreases [61].The cha ces of error also exist due to the uncertainties in complicated factors such as holidays and availability of alternative routes etc. Consideration of such unusual factors while designing the tra c ow prediction architectures can signi cantly improve prediction accuracy.


Chapter 3: Related Work

In this Chapter, we discuss researches which have been done particularly related to tra c ow prediction for intelligent transport systems.Due to the number of vehicles increased in the city, urban peo

e demand for better tra
ow prediction is also increasing.There are research works which have been done on tra c ow prediction using various approaches.Identifying urban tra c ow is one of the most important facets in ITS.Furthermore, existing research works are discussed based on machine learning and deep learning techniques.Finally, the factors that affect the ow of the tra c will be discussed.Urban tra c congestion estimation and prediction based on oating car trajectory data [64].Tra c ow prediction is an important precondition to alleviate tra c congestion in large-scale urban areas.In this paper, the researchers proposed a novel approach to estimate and predict urban tra c congestion using oating car trajectory data e ciently.In this method, oating cars are regarded as mobile sensors, which can probe a large scale of urban tra c ows in real time.In order to estimate the tra c congestion, they use a new fuzzy comprehensive evaluation method in which the weights of multi-indexes are assigned according to tra c ows.To predict the tra c congestion, an innovative tra c ow prediction method using particle swarm optimization algorithm is responsible to calculate tra c ow parameters.Then, a congestion state fuzzy division module is applied to convert the predicted ow parameters to citizens' cognitive congestion state.Experimental results show that the proposed method has advantage in terms of accuracy, instantaneity and stability.But weather data is not considered as long as it has impact on the prediction e ciency.


Machine Learning Based Tra c Flow Prediction


X. Ling et al. proposed Short-term Tra c Flow Prediction with Optimized Multi-kernel Support

Vector Machine [65].Accurate prediction of tra c state can help solve the problem of urban tra c congestion, providing guiding advices for users travel and tra c regulation.In this paper, they proposed a novel short-term tra c ow prediction algorithm, which is based on Multi-kernel Support Vector Machine (MSVM) and Adaptive Particle Swarm Optimization (APSO).Firstly, both the nonlinear and randomness characteristic of tra c ow, and hybridize Gaussian kernel and polynomial kernel to constitute the MSVM are explored.Secondly, optimize the parameters of MSVM with a novel APSO algorithm by considering both the historical and real-time tra c data.The algorithm is evaluated by doing thorough experiment on a large real dataset.The results show that the proposed algorithm can do a timely and adaptive prediction even in the rush hour

en the tra c conditions rapidly changed.The proposed method took the in uence of historical a
d real-time data to the ow of tra c in future moment into account, thus provides more accurate prediction result.Analysis Support Vector Machine theory [66].Based on the previous literature review, this research builds a short-term tra c speed forecasting model using Support Vector Machine (SVM) regression theory referred as SVM model.Besides the advantages of the SVM model, it also has some limitations.Perhaps the biggest one lies in the choice of appropriate kernel function for the practical problem.How to optimize the parameters e ciently and effectively presents another problem.

Unfortunately, these limitations are still research topics in current literature.This research puts an effort to investigate these limitations.In order to nd the effective way to choose the appropriate and suitable kernel function.This research constructs a new kernel function using a wavelet function to capture non-stationary characteristics of short-term tra c speed data.In order to nd the e cient way to identify the model structure parameters, the Phase Space Reconstruction theory is used to identify the input space dimension.To take the advantage of these components, the paper proposes a short-term tra c speed forecasting hybrid model Chaos-Wavelet Analysis-Support Vector Machine model, referred to as C-WSVM model.The real tra c speed data is applied to evaluate the performance and practicality of the model and the results are encouraging.The theoretical advantage and better performance from the study indicate that the CWSVM model has good potential to be developed and is feasible for short-te m tra c speed forecasting study.But needs further studies to apply the model to other tra c variable data sets such as tra c volume, travel time and average occupancy.This study chooses the tra c speed as the demonstration.Other limitations, such as the choice of the appropriate loss function for shortterm tra c variables forecasting model and parameters determination deserve further investigation.2. S. V. Kumar et al. proposed a prediction scheme by using the Kalman Filtering Technique (KFT) was proposed and evaluated [67].The proposed system requires only limited input data.The Kalman lter [68] allows a uni ed approach for prediction of all processes that can be given a state space representation.Tra c movement prediction using both signi cant and real-time data on the day of interest was also attempted.Promising results were obtained with mean absolute percentage error (MAPE) of 10 between observed and predicted ows and this indicates the suitability of the proposed prediction scheme for tra c ow estimation in ITS applications.[69].Even if timeseries data values vary abruptly or show uctuations, the presented model shows effective accuracy.Although k-NN based methods cannot perform spatial and temporal modeling simultaneously [70].


Chang et al. proposed a k-nearest neighbor non-parametric regression (kNN-NPR) model

Support Vector Regression (SVR) falls under supervised ML algorithms, i.e., it is trained to learn a function to map input feature to output and is mostly use

for classi cation and regression.The purpose of SVR is to map given data to a high dimensional featur
space followed by performing linear regression with the same space.Here, rst, each item in the dataset is plotted as a point in n-dimensional feature space.Then, classi cation is performed by locating hyperplane that divides the given input into classes.Compared to NN, SVR involves principles of Structural Risk Minimization (SRM).Also, it guarantees the localization of global minima.Some literature work has used SVM instead of SVR.However, both of these methods are almost the same, but the difference came in the type of value they provide as output (SVR outputs a real number whereas SVM outputs either 0 or 1).Authors in [71] incorporated an online version of Support Vector Machine (SVR) named incremental SVR and concluded that proposal is better against ANN approach.However, this study represents a drawback in the experimental setup.The authors, however did not present su cient evaluation of performance on their work.


J.Y. Ahn et al. proposed Predicting Spatiotemporal T

c Flow based on Support Vector Regression
and

Bayesian Classi er [72].To satisfy the demand of tra c ow estimation, this research studies the method of real-time tra c ow prediction based on Bayesian classi er and support vector regression (SVR).First the tra c ow is modeled and its relations on the roads using 3D Markov random eld in spatiotemporal domain.Based on their relations, the researchers de ne cliques as combination of current cone-zone and its neighbors.The dependencies on the de ned cliques are estimated by using multiple linear regression and SVR.Finally, the tra c ow at next time stamp is predicted by nding the speed level with decreasing the energy function.To evaluate the performance of the proposed method, it has been tested on tra c data obtained from Gyeongbu expressway.The experimental results showed that the approach using SVR-based estimation showed superior accuracy than linear-based regression.But extra investigation using deep learning is necessary for accurate prediction of tra c ow.


Deep Learning Based Tra c Flow Prediction

A Supervised Deep Learning Based Tra c Flow Prediction (SDLTFP) was proposed which is a type of fully-connected deep neural network (FC-DNN) [73].Timely prediction is also a major issue in guaranteeing reliable tra c ow prediction.However, training deep network ould be time-consuming, and over tting is might be happening, especially when feeding small data into the deep architecture.The network is learned perfectly during the training, but in testing with the new data, it could fail to generalize the model.The Batch Normalization (BN) and Dropout techniques are adopted to help the network training.SGD and momentum are carried out to update the weight.Then take advantage of open data as historical tra c data which are then used to predict future tra c ow with the proposed method and model.Experiments show that the Mean Absolute Percentage Error (MAPE) for the tra c ow prediction is within 5% using sample data and between 15-20% using out of the sample data.Training a deep network faster with BN and Dropout reduces the over tting.However, the spatio-temporal data relationship is not considered by adding each road segment and need to examine the prediction with network perspective.PCNN: Deep Convolutional Networks for Short-Term Tra c Congestion Prediction [74].

Understanding and modeling the c conditions extremely be di cult, and the observations from real tra c data reveal two properties.First similar tra c congestion patterns exist in the neighboring time slots and on consecutive workdays.Second the levels of tra c congestion have clear multiscale properties.To capture these characteristics, a novel PCNN method is proposed, which is based on a deep convolutional neural network, modeling periodic tra c data for short-term tra c congestion prediction.PCNN has two pivotal procedures such as time series folding and multi-grained learning.It rst temporally folds time series and constructs a 2-D matrix as the network input, such that both real-time tra c conditions and past tra c patterns are well considered.Then, with a series of convolutions over the input matrix, it is able to model the local temporal dependency and multiscale tra c patterns.In particular, the global trend of congestion can be addressed at the macroscale, whereas more details and variations of the congestion can be captured at the microscale.Experimental results on a real-world urban tra c data set con rm that folding time series data into a 2-D matrix is effective and PCNN outperforms the baselines signi cantly for the task of short-term congestion prediction.However, another way of deep learning model

needed
o further investigate accurate tra c ow prediction.[75].This work addresses non-stationary tra c ow prediction by implementing an intelligent update scheme to deep neural networks.The intelligent update scheme works by monitoring the frequency domain features extracted from the tra c ow time series.The features at present are compared with the previous ones through a distance function.The resulting similarity is then fed to the exponentially weighted moving average to detect whether a notable change in the tra c ow is present or not.The proposed method is evaluated using experimental analysis and it can be able to handle non-stationarity and produce acceptable tra c ow prediction.Apart from this, the result shows that using limited training data, the predictor is able to maintain good quality prediction in the 13/24 relatively far future.Moreover, when compared with the fully stochastic gradient descent update scheme, it as been shown that the proposed method is able to save time and computational resources up to 13% without losing the performance of prediction.But this work needs to scale up by including more features and increasing the number of freeways included in the model as w ll as the duration of the tra c ow data to have accurate tra c ow pre iction.


Summary

Tra c ow prediction is to predict the tra c ow rates based on he number of vehicles within a certain minute on the lane in the tra c netw rk.Tra c ow prediction can be done based on the historic l and current tra c ow data, trajectory data, weather data, and events etc.This is a t pical big data driven state forecasting problem for large dynamic systems, and is a fundamental problem in transportation system scheduling and optimization.Due to these reasons, Deep Neural Network based tra c ow prediction has attracted great research attentions [76]  We believe that, the aforementioned questions are very important determinant factors for effective prediction of tra c ow.Therefore, we are intended to suggest to consider the aforementioned questions whenever we want to a new tra c ow prediction to contribute robust tra c ow prediction which is actually be a tributary for to state of the art models.There are many deep learning models that can be utilized to learn from the tra c data collected through a smart city's infrastructure which predict the tra c ow in the city.Moreover, most of the researches which have been studied so far for tra c ow prediction uses RMSE and MAE for measuring the performance of proposed models.However, these two metrics fail when input datasets are entirely different from each other [80].Hence, we will de ne benchmark metrics for performance evaluation of our tra c ow prediction model.Tra c ow prediction is an important aspect in intelligent transportation system.The primary purpose of this proposed work is to

ra c ow f
r urban citizens in time will be designed, so as to realize the intelligent transport system in smart city domain.As such, our proposed solution offers the way how stable tra c ow prediction system.

Chapter 4: Tools For Experimental Analysis


Overview

This Chapter deals with the implementation tools described.The performance evaluation metrics and analytical tools are

iscussed.Various p
rformance metrics such as mean absolute error (MAE), root mean-square error (RMSE), and mean relative error (MRE) help to evaluate the effectiveness of the models of tra c ow system to know how the objective achieved are discussed.


Development Tools

DL platform provides an interface to design deep learning architectures easily by using pre-built and optimized libraries or components.Optimized performance, easy to code, parallelization, reduced computations, automatic gradient computations are some key characteristics of a good deep learning platform.Leading companies such as Google, Microsoft, Nvidia, Amazon are investing heavy money in developing raphic Processing Unit (GPU) accelerated deep learning platforms

r implement
tion of fast and large computation.From those all the existing platforms, TensorFlow is widely used and most popular among the users that is why we use in this research.

In this section deep learning platforms are reviewed as follows.


TensorFlow

This platform was introduced by Google brain team in late 2015.Its support languages such as Python, C++, R, and Java which make this tool popular.Moreover, it allows to work with one or more CPUs and GPUs with high data scalability.Hence, an individual person with a tablet or largescale distributed system can rely on TensorFlow.However, scholars suggested to use TensorFlow with server grade multi-thread implementation [81].It takes any model as directed acyclic graph (DAG) where nodes of the graph present mathematical operations whereas edges present tensors (multi-dimensional array) between them.Video analysis, visualizati n of distribution, sound recognition, analysis, and object detection are some uses of TensorFlow.Furthermore, TensorFlow supports distributed training, provides low latency for mobile users, and

sy to integrate
with SQL tables.

TensorFlow is more suited for most of deep learning models due to its features like that it supports an extensive built-in support for deep learning and Mathematical function for neural network.


Deeplearning4J

Deep Learning for Java (DL4J) is a robust, open-source distributed deep learning framework for the JVM created by Skymind [82], which has been contributed to the Eclipse Foundation and their Java ecosystem.DL4J is designed to be commercial-grade as well as open source, supporting Java and Scala APIs, operating in distributed environments, such as integrating with Apache Hadoop and Spark, and can import models from other deep learning frameworks (TensorFlow, Caffe, Theano) [83].It also includes impl

entatio
s of restricted Boltzmann machines, deep belief networks, deep stacked autoencoders, recursive neural networks, and more, which would need to be built from the ground up or through example code in many other platforms.


Theano

Theano is a highly popular deep learning platform designed primarily by academics which, unfortunately, is no longer supported after release 1.0.0 (November, 2017).Initiated in 2007, Theano is a Python library designed for performing mathematical operations on multi-dimensional arrays and to optimize code compilation [84], primarily for scienti c research applications.More speci cally, Theano was designed to surpass other Python lib

ries,
ike NumPy, in execution speed and stability optimizations, and computing symbolic graphs.Theano supports tensor operations, GPU computation, runs on Python 2 and 3, and supports parallelism via BLAS and SIMD support.


Torch

Torch is also a scienti c computing framework; however, its focus is primarily on GPU accelerated computation.It is implemented in C and provides its own scripting language, LuaJIT, based on Lua.In addition, Torch is mainly supported on Mac OS X and Ubuntu 12C, wh

e Windows impleme
tations are not o cially supported [85].Nonetheless, implementations have been developed for iOS and Android mobile platforms.Much of the Torch documentation and implementations of various algorithms are community driven and hosted on GitHub.Despite the GPU-centric implementation, a recent benchmarking study [86] demonstrated that Torch does not surpass the competition (CNTK, MXNet, Caffe) in single-or multi-GPU computation in any meaningful way, but is still ideal for certain types of networks.


Caffe2

Caffe was by Berkeley AI Research (BAIR) and the Berkeley Vision and Learning Center (BVLC) at UC Berkeley to provide expressive architecture and GPU support for deep learning and primarily image classi_cation, originating in 2014

7] [88].Caf
e is a pure C + + and CUDA library, which can also be operated in command line, Python, and interfaces.It runs on bare CUDA devices and mobile platforms, and has additionally been extended for use in the Apache Hadoop ecosystem with Spark, among others.Caffe2, as part of Facebook Research and Facebook Open Source, builds upon the original Caffe project, implementing an additional Python API, supports Mac OS X, Windows, Linux, iOS, Android, and other build platforms [89].4.2.6 Keras Though not a deep learning framework on its own, Keras

ovides a high-level API that integr
tes with TensorFlow, Theano, and CNTK.The strength of Keras is the ability to rapidly prototype a deep learning design with a user-friendly, modular, and extensible interface.Keras operates on CPUs and GPUs, supports CNNs and RNNs, is developer-friendly, and can integrate other common machine learning packages, such as scikit-learn for Python [90].In addition, it has been widely adopted by researchers and industry groups over the last year.


MXNET

Apache MXNet supports Python, R, Scala, Julia, C++, and Perl APIs, as well as the new Gluon API, and supports both imperative and symbolic programming.The project began around Mid2015, with version 1.0.0 released in December of 2017.MXNet was intended to be scalable, and was designed from a systems perspective to reduce data loading and I/O complexity [91].It has proven to be highly e cient primarily in single-and multi-GPU implementations, while CPU implementations are typically lacking

92].


Microsoft Cognitive Tool
it (CNTK)

The Microsoft Cognitive Toolkit, otherwise known as CNTK, began development in Mid-2015.It can be included as a library in Python, C#, and C + + programs, or be used as a standalone with its own scripting language, BrainScript.It can also run evaluation functions of models from Java code, and utilizes ONNX, an open-source neural network model format that allows transfer between other deep learning frameworks (Caffe2, PyTorch, MXNet) [93].Conceptually, CNTK is designed to be easy-to-use and production-ready for use on large production scale data, and is supported on Linux and Windows.In CNTK, neural networks are considered as a series of computational steps via directed graphs, and both neural network building blocks and deeper libraries are provided.CNTK has emerged as a computationally powerful tool for machine learning with performance similar to other platforms that have seen longer development and more widespread use [92].


Performance Evaluation Metrics

To evaluate the effectiveness of our proposed solution, we use the mean absolute error (MAE), root-meansquare error (RMSE), and mean relative error (MRE) as the performance evaluation metrics.Root me n square error (RMSE) and mean absolut

percentage
rror (MAPE) are the most commonly used performance measures.RMSE variants like Normalized RMSE(NRMSE) and RMSE with cost (RMSEC) have also been used.Performance measures like Mean Absolute Relative Error(MARE), Coe cient(EC), R square, Mean Square Error(MSE), Mean Relative Error(MRE), Accuracy and Variance Absolute Percentage Error(VAPE) are adopted in signi cant numbers but not nearly as much as the ones mentioned above Performance measures that are unconventional to the eld of tra c forecasting like precision, recall, F1, FP rate, sensitivity and various others as well as custom performance measures proposed by authors, all together make up a signi cant amount of the total.This makes it harder to compare models between papers.Zilu Liang and Yasushi Wakahara suggested that that Symmetric MAPE be used instead of the widely adopted MAPE as MAPE yields a biased evaluation when real value is close to zero [94].The p imary objective of this work is to study on various models of tra c ow prediction so as to realize robust tra c ow prediction scheme in intelligent transport system by identifying various important factors that affect the performance of modeling tra c ow prediction.

From this study, applying deep learning model can effectively predicts the tra c ow in the city.To design tra c ow prediction, need to consider spatial and temporal features, which can capture more tra c data features.Hybrid deep learning network structures a e also better to capture spatial temporal tra c information.Existing technologies, tools and approaches are studied.This study is used to know the appropriate approach to design robust tra c ow prediction architecture so as to characterize the dynamics of tra c.

Deep learning models are characterized by handling time series data.Deep learning models can learn periodicity of tra c ow data and ability of memorizing long-term dependencies.Spatial and temporal characteristics should be considered to produce accurate prediction model.Furthermore, to identify how to provide more accurate prediction, we analyze critical factors that affect tra c ow.Therefore, urban transport quality can be improved to effectively deploy intelligent transport system.This survey helps researchers to derive the essential characteristics of tr

c ow predicti
n and identi es various techniques which are successful for tra c ow prediction.Generally, th manuscript covers relevant state-of-the-art machine learning and deep learning methods, which would help researchers to explore future directions.
Contribution

This work studies relevant methods and approaches how to produce tra c ow prediction.

State-of-art works are studied with respect to different tra c ow predicti n techniques which would help researchers to explore future directions.

The factors that hav

impact on th
performance of tra c ow prediction, development tools and bench mark performance evaluation metrics for the tra c w prediction are studied.

Various machine learning and deep learning models for tra c ow prediction have been studied.


Future Work

Design a deep learning model for tra c ow prediction considering spatial temporal features from times series data.

There is no accurate tra c ow prediction yet which lacks accurate representation of tra cs due to inability to effectively deal with spatial temporal features of times series data and needs extra investigation to have better prediction model which can predict tra c ow better.Therefore, as a solution, an approach is required to e ciently detect and identify the ow of tra c.



1. J. Wang et al. proposed Short-term tra c speed forecasting hybrid model based on Chaos-Wavelet




1. Z. Zheng et al. proposed a Deep and Embedded Learning Approach for Tra c Flow Prediction in Urban Informatics [63].Most previous studies on tra c ow prediction fail to capture ne-grained tra c information like link-level tra c and ignore the impacts from other factors, such as route structure and weather conditions.This research proposed a deep and embedding learning approach (DELA) that can help to explicitly learn from ne-grained tra c information, route structure, and