Modelling Public Transport Preferences Using Artificial Neural Networks PROJECT By GADI VENKATA NAGA SAI CHANDANA 18CE02015 Under Supervision of Dr. Debasis Basu SCHOOL OF INFRASTRUCTURE INDIAN INSTITUTE OF TECHNOLOGY BHUBANESWAR BHUBANESWAR -7502050, ODISHA September 2022CONTENTS CHAPTER | Introduction... 1.1. General 1.2. Need of the Study 1.3. Objective caverns 14, Scope of the Work..... CHAPTER 2 Literature ReVieW josie 2.1. General 2.2. Binary Logit Model ..nnsnnnnnnnnnnnnnnnn 2.3. Logistic Regression 2.4, Artificial Neural Network Approach... CHAPTER 3 About Artificial Neural Network: 3.1, Introduction, ul 3.2. Basics of Neuron and Neural Network... ul 3.3. Feed Forward Neural Networks: 13 3.3.1. Single Layer Feed Forward NN (or Single Layer Perceptron):... 13 3.3.2. Multilayer Feed Forward Network. 13 3.3.3. Supervised ANN... 17 3.4, Neural Network Training ....snsenennansiennseennse . seo IT 3.4.1, Feedforward Propagation... 7 3.4.2 Activation Function... sesnennsen ssnenesseneseees IT 3.4.3, Weight Initialization, 18 3.4.4. Backpropagation Error. 18 3.4.5. Optimization/Learning Algorithm... a 18 Study Area and Mode Choice Model Development..... oe 2M 4.1. General a 4.2, Study Aten es oe 2d 4.3. Attributes for Neural Network Modelling 2 References oe BCHAPTER 1 Introduction 1.1. General The transportation system is a vital component of every growing economy. Forecasting travel demand is a crucial component of transportation planning. Travel demand can be defined as the number of people or vehicles per unit time that can be anticipated to travel on a specific segment of a transportation system, considering the land use pattern, socioeconomic conditions, and environmental factors. There is a need to forecast the travel demand to estimate the vehicular volume for the future, based on which various transportation system alternatives can be designed or modified. Due to substantial growth of the urban and suburban population the urban travel demand is, increasing. Each member in this process has individual decisions according to their personal preferences. This individual choice involves trip purposes, frequency of trip, timing, destination of trips as well as mode of travel. Travel is not only an end objective of the trip maker rather it is also associated with work, shopping, and recreation. ‘The fundamental cause of the escalating problems with traffic congestion is the public's steadily growing desire in personal automobiles. Private transportation has traditionally been the favored, dependable, comfortable, and time-effective option. It has had a significant impact on the pollution, accidents, and traffic congestion that are all getting worse. It was stated that younger and economically underprivileged people make up a disproportionate number of car accident victims. There needs to be a drastic solution (o this issue, one that encourages commuters to give up their automobiles and take the public transportation instead (car, bus, and vanpool). According to Belwala [4], the importance of public transportation in our everyday lives necessitates the pro} ion of highly effective public services. 1.2, Need of the Study A more accurate answer is produced by ANN, which is independent of any mathematical framework, and is an outmoded alternative to traditional route choice modelling techniques like linear regression and the MNL model, When used to find a model for the choice of transport mode and route, ANN perform well for statistical approaches. As ANN is non-linear, it has a better fit to the data because it doesn't make any assumptions about the data before estimating the output, which leads to better prediction or forecasting of the output. ‘The learning and adaptability of ANN enable it to adapt itself to the situation and update its internal architecture in response to the changing environment. 1.3. Objective The primary objective of the work is to develop a feed-forward artificial neural network (ANN) model for mode choice model. The ANN model will be of supervised nature. The best selected ANN model will be used to compare its performance with the outcomes of bi- nary logit-based choice model. 1.4, Scope of the Work ‘* Understanding ANN ‘* Develop a feed-forward artificial neural network (ANN) model for Mode choice model using MATLAB ‘* Calculate the probability of different mode choice. ‘* Compare the result between ANN and logit model and find out which is more efficient,CHAPTER 2 Literature Review 2.1, General Modal split is the third step of the conventional travel demand modelling, The step deals with the mode-choice analysis, Mode Choice problem has been approached by transportation planners in many ways. In a broad way all these approaches can be classified into two categories- discrete choice models and non-diserete choice models. Discrete choice models include probit model, multinomial logit model and nested logit model. Non-discrete choice models include regression approach, ctoss classification tables and diversion curves. In our research we shall concentrate on three models - Logit model, Linear regression approach, and the Artificial neural networks approach, All three methods model the probability of a passenger choosing a given mode of transportation, 2.2, Binary Logit Model It is the most common method to do travel mode model due to its ease of estimation and simple mathematical structure (CH. Wen & F.S. Koppelman,2001). It makes a simple assumption that alternative of utility function is independent and identically distributed (IID). Probability is calculated based on Gumbel distribute, The Logit model says, the probability that a certain mode choice will be taken is proportional to ¢ raised to the utility over the sum of ¢ raised to the utility Binomial logit model or binary logit model is a case of logit model which have only two alternatives whereas when a logit model have more than two alternatives it is called as Multinomial logit (MNL) model. The utility function used in logit model is linear in nature and for estimation choice function parameter maximum likelihood method is used.3. Logistic Regression Unlike logit model, logistic regression (LR) model don’t assume a linear relationship between different attribute and the alternative choice in a choice model. Also, LR model don’t required variable to be normal distributed and require less condition than linear regression model (D. Collett, 2003). Advantage of LR model is that not only it predicts the probability of a choice which depend on the attributes value (categorical or numerical), but it also estimates probability of choice of traveler. Binary outcome model is the most common regression model. Binary LR model takes two values for example, true/false, A or B etc. Whereas multinomial LR model is used for classification problem. Linear regression transformation is logistic regression which is achieve with help of sigmoid function. In equation yx is Bernoulli distributed. 1 70) = Trin he tog (==) = Bo + Bit t+ Badia Figure 2.3 Logit = In (a/ (1 ~ m)) vs probability (x) (J. N. Hussain,2008) Maximum Likelihood function is used to estimate unknown parameters (Bi)2.4, Artificial Neural Network Approach Over the last few years, artificial intelligence techniques like neural networks are being used in the place of traditional regression methods for developing various travel demand forecasting models, In addition to being able to learn from examples, generalise what they have learned, and apply it to new data, neural networks are also better able than other computational techniques to capture complicated relationships. Neural networks have been employed in the forecasting of intercity flows by Nijkamp et al. (1992), Teodorovi and Vukadinovic (1998), and Chin et al. (1992) among other writers (1996). In theory, feedforward neural networks can be thought of as "universal approximators” because they accurately estimate unknown functions. A multilayered feedforward neural network with one hidden layer can approximate any continuous function up to the appropriate level of accuracy provided it contains a sufficient number of nodes in the hidden layer, according to the theorem established by Homik et al. (1989) and Cybenko (1989). Hussain Dhafir Hussain, et.al, This study illustrates a comparison between artificial neural network ANN and multinomial logistic regression MNL to predict the behavioural transportation of mode choice. There are three parts of the dataset: the training set, the validation set, and the test set. In this study, 70% of the data for the training set and 15% for each of the validation and test sets was utilized accordingly. ‘An artificial neural network (ANN)-based method is built using MATLAB to examine transportation features with the goal of identifying the mode chosen for predictability between private vehicles and public transit, specifically the bus and vanpool. The findings from the Multinomial logit clearly showed that artificial neural networks outperformed them in terms of both predictability and validation. MNL. model predictability for the evaluation of auto, bus, and vanpool is 72.60%, while ANN model predictability is 82.6%. MNL model is 76.2% and ANN model is 80.9% for validating purposes.The results of the most accurate predictability tests between the artificial neural network (ANN) model and the multinomial logit model are presented in the conclusion, and it is shown that the ANN model is better able to forecast the mode of a decision than the multinomial logit model.CHAPTER 3 About Artificial Neural Networks 3.1. Introduction Artificial neural networks may be thought of as simplified models of the networks of neurons that occur naturally in the animal brain. A paradigm for information proc sing that draws inspiration from the brain is called an artificial neural network (ANN). ANNs learn by imitation much like people do. Artificial Neural Networks (ANN) are brain-inspired algorithms thatare used to foresee problems and model complex pattems. The idea of the biological neural networks in the human brain gave rise to the Artificial Neural Network (ANN), a deep leaming technique. An effort to simulate how the human brain functions led to the creation of ANN, Although they are not exactly the same, the operations of ANN and biological neural networks are very similar. Only structured and numeric data are accepted by the ANN algorithm. Through a leaning proces an ANN is tailored for a particular purpose, such as pattem recognition or data categorization 3.2, Basies of Neuron and Neural Network The neuron, also known as a node or unit, is the fundamental computational component of ‘a neural network. It computes an output after receiving input from various nodes or an extemal source. Each input has a corresponding weight (w), which is determined by how significant it is in relation to other inputs. The weighted total of the inputs to the node is subjected to a function. The concept is that synaptic strengths (the weights w) are learnable and govern the strength of effect and its direction: excitatory (positive weight) or inhibitory (negative weight) of one neuron on another. In the fundamental model, the signal is sent via the dendrites to the cell body, where it is added together. The neuron can fire and send a spike along its axon if the total is higher than a certain threshold. In the computational model, we assume that just the frequency of the firing conveys information and that the precise timings of the spikes are irrelevant. An activation function (e.x sigmoid function), which simulates the frequency of the spikes down the axon, is used to describe the firing rate of the neuron,Dendeites of ext neuron a=g(2)} Figure 3.1 Similarity between Biological and Artificial Neural Unit Neural Network Architecture is divided into layer of 3 types: 1, Input Layer: The training observations are fed through these neurons. No computations are done in this layer. 2. Hidden Layers: In Hidden layers is where intermediate processing or computation is done, they perform computations and then transfer the weights (signals or information) from the input layer to the following layer (another hidden layer or to the output layer). 3. Output Layer: Using a desired/well-performed activation function to map outputs E. , sigmoid activation function for classification) 4, Activation function: Given an input or group of inputs, a neuron's activation fanction determines the output of that node. A typical computer chip circuit may be thought of as a digital network of activation functions that can be "ON" (1) or "OFE" (0) depending on the input. This is comparable to how a linear perceptron functions in neural networks. ‘The nonlinear activation function, however, is what enables such networks to calculate nontrivial problems with just a few nodes. This function is also known as the transfer function in artificial neural networks. 5. Learning rule: A learning rule is an algorithm that adjusts the neural network's parameters such that a particular input will create a preferred output. The weights and thresholds are generally changed as a result of this learning process.3.3. Feed Forward Neural Networks: A feedforward neural network is a type of artificial neural network where there are no cycles in the connections between the units. In this network, information travels only in one direction—forward—from the input nodes to the output nodes, passing via any hidden nodes that may exist. The network doesn't contain any loops or cycles. Feed Forward Neural Networks are divided into two parts 3.3.1. Single Layer Feed Forward NN (or Single Layer Perceptron): ‘This is the simplest feedforward neural Network and does not contain any hidden layer, which means it only consists of a single layer of output nodes. This is said to be single because whenwe count the layers, we do not include the input layer, the reason for that is because at the input layer no computations is done, the inputs are fed directly to the outputs via a series of weights. a, wa a oN a(os am) SNe a; ) Wr ay Figure 3,2 Simple Perceptron network, BPNN (Backpropagation neural network), here error is minimized by updating the weights and used for pattem prediction, RBFN (Radial basis function network) which have radial basis function hidden layer. . Multilayer Feed Forward Network Feedforward neural networks is the simplest forms of ANN consists of three types of, layers based on type on node iden. In feedforward neural network input, output and data or the input signal travels in only forward direction. The data passes through the input layer then through hidden layer and it exits on the output layer. It has a front propagated wave and no backpropagation wave by using some classified activation function During each front propagation the weights and the basis value is calibrated depending on the different between out produce and the result and this process is repeatedto minimize the error. The number of neurons in the input layer is equal to total number of attributes of our problem set. Data from input layer transfer to hidden layer here the transformation of data take place and then out layer predict the feature as the result. Its application is found in computer vision and speech recognition where classifying the target classes is complicated. It also responsive to the noisy data and itis easy to maintain Hidden Layer Input Layer Out Layer Outputs Inputs i C) yl X2—e, v2 i Xn——w Ya Bias Figure 3.3 Feedforward Neural Network Node of Neural network: The processing unit of an artificial neural network is called neurons or node. Inthe figure 3.2 node is denote by a circle the and signal flow is denoted by arrow. Variable x1, x2, and x3 are representing input signals whereas variable w1, w2, and w3 represent weights for corresponding signals. Lastly, b represent bias, which is a factor associated with the storage of information, Figure 3.4 Diagram of structure of Nodexi are the input signal of nodes which comes from another node o from the outside it current node is input note then the signal get multiplied by wi (weight). When the weighted signals are goes at the neuron (node), then summation of these values are added to b (bias) which is calculated as follows. 3 v=) uxwtb From equation we can see that xi with a greater wi has a greater effect on result (v). Finally, weighted sum from node goes into the activation function and the resultant value left to node (y). The activation function also called transition function determines the behavior of the node. Y= flv) = fix + w+ b) {{) in this equation is the activation function. Activation Funetion: They are the function applied to the output of neuron which ‘modifies the output before transferring the data further. Activation functions are used in hidden and output layer. Some example of activation functions is Step Function. Linear Function, Sigmoid Function, Sign Function give in the figure. Figure 3.5 Activation Functions ‘Weight Matrix: Weight matrixes are used to represent weight to each layer of a neural network. So that weight can be easily assessable and can be updated for a particular layer.6. 6; oe o' ih Al, Al, ol, Figure 3.6 Weight representation in ANN Here input is a 1x4 matrix 10 and the weight for first hidden layer is a 4x3 matrix W1 and the matrix multiplication of input and weight matrix will give 1x3 input matrix Z1 for activation function of first hidden layer considering bias value is 0. 10 = [xl x2.x3 x4] War Win Was Wei Wao Was Way Wrz Waa War Wee Was, Z1=10W1 Z1=[z1 2223] i= Dxj x Wj Same step is repeated for each layer as earlier until the output layer3.3.3. Supervised ANN In this case both the input and output are available. The inputs undergo process and gives some outputs which are then compared with the desired output. The errors occurred in the system uses back propagation technique to adjust the weights given in the inputs. By this process the error gets eliminated and the system remains balanced. This process ‘occurs in the system frequently. The connections between the weights are continually refined, as the same set of data is processed several times during the training of a network, In case some of the important information are lacking in input then the system may not eam some specific operations (Sharma et al., 2001). In case a network is unable to process the input some of the factors needs to be reviewed which includes number of layers, the connections between the layers, the number of elements per layer, the input and outputs, transfer, training functions, the summation, and also the first weights. To adjust the weights one adaptive feedback system is required, Back propagation technique is the best adaptive feedback system. The ANN can also be converted into a hardware for the speed of the process. This stage arrives after there is no requirement of further learning and the weights can be declared as frozen (Anderson et al,, 1992), 3.4, Neural Network Training Backpropagation (BP) algorithm is used in the training process of ANN which include input training through feedforward, backpropagation to minimize loss, and calibration of weights, 3.4.1. Feedforward Propagation The input xi from each node for previous layer is feed to next layer node and their weighted summation value with a bias goes to an activation function and its out is forward to next layer node vo Sxixwit 3 3.4.2 Activation Function ctivation function play important part in backpropagation algorithm it should be continuous in nature and differentiable in the input set and it should be monotonicallynon-decreasing. During backpropagation the computation cost and time of its derivative should be less. 3.4.3, Weight Initialization Initial weight must be assigned to the network and the key point should be consider in the process is the weight be small (not to small) and it should not be same value and should have good variance. Initialization can be done through uniform distribution or Xavier initialization. In the equation given weigh initialization method through uniform distribution, Here fin is number of inputs in network. wij = U(a, b) 1Nfin, b=1/Nfi 3.4.4, Backpropagation Error ‘As ANN produce a target value correspond to set of input from which error and cost function is calculated. Mean square error (MSE) is calculated to check how close target value is to the actual value, error =(y-9)? cost function =¥(y-$)?/n Weight correction is done through backpropagation to minimize overall error. New weight can be calculated by the given equation where 1 is learning rate and L is error or loss function. 3.4.5. Optimization/Learning Algorithm Optimization algorithms are used to find the best fill line for given training data set. Some optimization algorithms are Gradient Descent, Stochastic Gradient Descent (SGD), mini batch Stochastic Gradient Descent. Gradient Descent Algorithm To find local minimum of error function gradient descent algorithm is used which is a first-order iterative (differentiable) optimization function, We calculate cost/loss function then calculate differentiable function of loss function w.r.t current weight value and put it to the below equation to find new weight and perform this process until the loss reache: global minimum,Wrew = Wold —n *(OL/ OWots) Gradient descent algorithm loss Initial loss — Gradient Global loss minimum Figure 3.7 Gradient descent algorithmCHAPTER 4 Study Area and Mode Choice Model Development 4.1. General ‘The steps in the mode choice modelling method begin with the extraction and pro essing of the data we are using because null points can produce inaccurate results. Next, the mode choice characteristic is defined. Afterward, a mode choice model is created using ANN and BNL models should both have predefined attributes. We are developing ANN ‘models based on the number of neurons in the hidden layer, number of hidden layers in the network, different optimization algorithm such as Gradient Descent or Gradient Descent with Momentum based on number of output neuron, Pee Matar aa cole oon Extraction of required Dataset Defining the attributes Deva KemeeN tan Artificial Neural Network approach Binary Logit Model Figure 4.1 Overall process of Mode choice model 4.2. Study Area Bhubaneswar, the capital city of Odisha, The cost of living is moderate in Bhubaneswar.4.3. Attributes for Neural Network Modelling For the analysis, two attributes are considered which were travel time and travel cost. The data have one or two dependent variable (output data point) which is choice of mode either SA or MB which in binary form (0/1) and 4 independent variable (input data point) which are TT_SA (travel time of share auto) in minute, TC_SA (travel cost of share auto) in rup TT_MB (travel time of Mo Bus) in minute and TC_MB (travel cost of Mo Bus) in rupees,References ‘Nijkamp, P., Reggiani, A. and Tritapepe, T., (1996) Modelling inter-urban transport flows in Italy: A comparison between Neural network analysis and logit analysis. Transportation Research C, 4, pp.323-338. Sharma S, Lingras P, Xu F, Kilburn P, (2001). Application of neural networks to estimate AADT on low-volume roads, Jounal of Transportation Engineering, ASCE, 426-432. Cybenko, G., (1989) “Approximation by Superposition of Sigmoidal Functions”, Mathematics of Control, Signals and System, 2, pp.193-204. Hussain Dhafir Hussain, Ahmed Madha Mohammed, Ali Dawod Salman, Riza Atiq bin ©. K. Rahmat and Muhamad Nazri Borhan., (2017) “Analysis, of Transportation Mode Choice using @ comparison of Artificial Neural Network and Multinomial Logit Models”, ARPN Journal of Engineering and Applied Sciences, VOL. 12, NO. 5.