An Intelligence-Based Optimization Model of Passenger ... - IEEE Xplore

1290

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 14, NO. 3, SEPTEMBER 2013

An Intelligence-Based Optimization Model of Passenger Flow in a Transportation Station J. K. K. Yuen, E. W. M. Lee, S. M. Lo, and R. K. K. Yuen

Abstract—This paper proposes an intelligence-based approach to predict passengers’ route choice behavior, which is crucial to the effective utilization of transportation stations and affects passenger comfort and safety. The actual route choice decisions of passengers are extremely difficult to mimic as they involve human behavior. A comprehensive methodology for capturing route choice behavior is still lacking because extensive labor and time resources are required to collect passenger movement data from different stations. In this paper, a four-month site survey was carried out to collect actual route choice behavior information in nine transportation stations in Hong Kong during peak hours. We developed an intelligent model to capture passengers’ route choice decision-making that achieved prediction accuracy of 86%. The applicability of this intelligent route choice model is demonstrated by optimizing the number of gates in a transportation station to inform the spatial design of the station. Index Terms—Artificial neural network (ANN), human factors, neural network applications, route choice, transportation.

I. I NTRODUCTION

T

HE STUDY of passenger flow in transportation stations is particularly important as there is typically a high passenger volume, a short train headway, and limited capacity in stations, which affects both the comfort and safety of passengers. In reality, the passenger flow inside a transportation station is extremely complicated, and passengers must make route choice decisions as soon as they enter the station. An intelligent route choice model is thus important to analyze passenger flows in transportation stations. Several dynamic evacuation and pedestrian models have been developed during the past few decades to model crowd movements. These models include social force models [1], cellular automata (CA) models [2], lattice gas models [3], fluid-dynamic model [4], agent-based models [5], and SGEM models [6], [7]. Many of the existing pedestrian flow models for simulating the dynamic movement of pedestrians are based on simple route choice models that consider either the shortest distance or the walking time between the origin and the destination [8]. Unlike the pedestrian flow models, drivers can obtain traffic conditions or information easily from the intelligent transportation systems [35]–[37], and there are some innovated dynamic models [38], [39] that are ready to be used to Manuscript received July 11, 2012; revised October 3, 2012 and February 6, 2013; accepted April 8, 2013. Date of publication May 6, 2013; date of current version August 28, 2013. This work was supported by Mass Transit Railway Corporation. The Associate Editor for this paper was H. A. Rakha. The authors are with the Department of Civil and Architectural Engineering, City University of Hong Kong, Kowloon, Hong Kong (e-mail: kakuiyuen2-c@ my.cityu.edu.hk; [email protected]; [email protected]; bckkyuen@ cityu.edu.hk). Digital Object Identifier 10.1109/TITS.2013.2259482

provide real-time traffic information. All these factors motivate us to develop an intelligent optimization model to facilitate the spatial design of the stations. Pedestrian route choice is influenced by many factors, such as personal experience, building geometry, interactions between occupants, and environmental factors [9]–[11]. Normally, passengers choose their desired path according to the shortest travel time, travel distance, or a combination of both [12]. However, Sime [13] and Proulx [14] found that evacuees choose familiar routes rather than the shortest path to the exits because they feel that unknown paths increase the threat. This pioneering work depicts the complexity of pedestrian route choice in human movement. To mimic reasonably the movement of a crowd in a station, actual pedestrian route choice behavior must be captured, rather than solely relying on mathematical equations or assumptions to make predictions. In recent simulations, Gwynne et al. [10], [15] has proposed an exit selection behavior model based on “queuing and familiarity behavior,” and Lo et al. [7] has proposed a game-theory-based exit selection model for evacuation. Both pedestrian route choice models use mathematics to simulate human decision-making. Hoogendoorn and Bovy [16] proposed a new theory of pedestrian route choice behavior under uncertainty based on the concept of “utility maximization” in continuous space. Zarita and Lim [17] introduced an intelligent approach that incorporated the probabilistic neural network (PNN) into the CA model to determine the decision-making ability of evacuees. Unfortunately, the PNN can only provide a single choice and not a distributed result. This paper thus proposes the alternative approach of applying a model of an artificial neural network (ANN) to predict passengers’ route choice behavior inside stations. According to Zhang et al. [22], ANN models are capable of capturing the fine details of functional relationships via a “learning” process from historical data. This learning feature is particularly useful for pedestrian route choice models as the relationships between the input parameters are less well known than in highly structured expert systems or equation-based approaches [18], and highly specialized human expertise or assumptions are not required. Once passengers enter a transportation station, they must make several decisions, such as how they will travel from the concourse to the platform and which escalators or stairs should be used to reduce the journey time. Generally, passengers travel from the entrance to the train by the following steps. 1) Enter the station through the entrance. 2) Move from the unpaid area to the paid area by choosing an automatic fare collection (AFC) gate group.

1524-9050 © 2013 IEEE

YUEN et al.: INTELLIGENCE-BASED OPTIMIZATION MODEL OF PASSENGER FLOW

Fig. 1.

Passenger flow inside a transportation station.

1291

Fig. 2. Route choice behavior of passengers.

3) Move from the concourse to the platform by choosing the escalators or stairs. 4) Take the train. To study the complex route choice behavior of passengers inside the transportation station, we divided the station into three main zones: Zone 1, which is the entrance to the ticket control; Zone 2, which is the area between the ticket control and the escalators or stairs; and Zone 3, which is the area between the escalators or stairs and the train (see Fig. 1). A site survey was separately carried out in each zone and the passenger flow in each zone was further studied by using an ANN and genetic algorithms (GAs) to mimic route choice behavior in crowds and optimize the usage of the station facilities. This paper focuses on Zone 1 (the passenger flow between the entrance and the ticket control) for ease of demonstration of the proposed approach. Fig. 3. Typical architecture of the MLP model.

II. D EVELOPMENT OF THE A RTIFICIAL N EURAL N ETWORK M ODEL As previously mentioned in Section I, knowledge of pedestrian route choice behavior is critical when analyzing passenger flow in a transportation station, and a suitable pedestrian route choice model can improve the performance of simulation tools. To analyze the passenger flow in Zone 1 (entrance to the ticket control), we adopted a multilayered perceptron (MLP) model [19] to predict the probability of passengers choosing the three nearest AFC gate groups (even for the situation with more than three AFC gate groups, more than 99% of the passengers choose the three nearest AFC gate groups according to our site observation), as shown in Fig. 2. The MLP is one of the most widely used ANN models for forecasting due to its simple and flexible nature [20]. It has been mathematically proven that the MLP with a single hidden layer is a universal function approximator [21], subject to the provision of a sufficient number of hidden neurons, and

the model has been successfully applied to various types of forecasting [22], [33], [34], [40]. The MLP architectural model consists of several layers: the input layer, the hidden layer(s), and the output layer, as shown in Fig. 3. The neurons of each layer are interconnected with the neurons of the adjacent layers. A three-layered (i.e., an input layer, a hidden layer and an output layer) MLP model was adopted in this paper. A. Input Layer The number of neurons in the input layer corresponds to the number of input parameters in the model. In this paper, the following input parameters were selected to predict the probability of passengers choosing the three nearest gate groups. 1) Input Parameters 1–3—Shortest Distance Between the Entrance and the Gate Group: Passengers usually choose the shortest route: this is their primary strategy for pedestrian route choice. For the same reason, passengers may choose gate

1292


Fig. 4. Shortest distances between Entrance A and the AFC gate groups EG1–EG3.

groups that are the shortest distance from their origin to reduce their journey time. Fig. 4 shows that the passengers in this paper can travel from Entrance A to the paid area by passing through gate groups EG1, EG2, or EG3 (the three nearest AFC gates groups); thus, the shortest distance between Entrance A and the gate groups, denoted by D1, D2 and D3, are recorded and used as input parameters of the ANN model. 2) Input Parameters 4–6—Number of Gates in the Gate Group: In addition to travel distance, the number of gates in the gate group also affects the route choice decision of passengers. As the capacity of the gates in the gate group is fixed, the more gates in the gate group, the higher the handling capacity of the group and vice versa. A high handling capacity of a gate group attracts more passengers to use it as they can reduce their queuing time around the gates. In this case, travel distance may become a less important factor. 3) Input Parameters 7–9—Percentage of Wide Gates in the Gate Group: There are two types of AFC gates in Hong Kong: “normal” gates and “wide” gates. Passengers are usually more eager to use normal gates than wide gates because the wide gate is designed for passengers carrying bulky luggage. In addition, passengers can travel between the paid area and the unpaid area in both directions using the wide gate, whereas passengers using normal gates can travel in a single direction only. Passengers using wide gates may thus need to wait for other passengers to move from the paid area to the unpaid area, which increases their journey time. Usually, gate groups are a combination of normal gates and wide gates. The percentage of wide gates in the gate groups (i.e., the number of wide gates in a specific gate group/the total number of gates in that group) is included as a model parameter.

Fig. 5.

Architecture of the MLP model. TABLE I I NPUT AND O UTPUT PARAMETERS OF THE MLP M ODEL

The rule of thumb adopted here is the rule developed by Ward Systems [23] described in the following: Nh = (Nin + Nout )/2 +

Ns

(1)

where Nh , Nin , and Nout are the number of neurons in the hidden layer, the input layer, and the output layer, and Ns is the number of training samples, respectively.

B. Hidden Layer The numbers of neurons in the input and output layers equal the number of input parameters and output parameters, respectively, in the model. This is defined by the system itself. The required number of hidden neurons is crucial to the performance of the model. There is currently no analytical approach to determine the number of hidden neurons, but various rules of thumb are available to provide heuristic hints to the user.

C. Output Layer The output of the MLP model is the probability of passengers choosing the three nearest gate groups (EG1–EG3). A site survey was required to collect this distribution, the details of which are introduced in the following. The input and output parameters of the MLP model and the architecture of the MLP model are shown in Fig. 5 and Table I, respectively.


1293

TABLE II A RRANGEMENT OF THE S AMPLES

Fig. 6. Route selection by passengers between the entrance and the gate groups.

III. DATA C OLLECTION M ETHOD Questionnaires and video and on-site observation are the data collection methods most commonly used in pedestrian studies [24]–[26]. Video observation has the advantage that data can be reviewed at any moment, but the camera coverage in the station was insufficient to capture the specific area that we wished to focus on as the purpose of station cameras is crowd and security control. In addition, this paper was carried during rush hours; passengers would be unlikely to stay in the station long enough to complete a questionnaire. On-site observation was thus chosen for its flexibility, which allowed us to take data from the whole station or any specific area. A four-month site survey was carried out on nine transportation stations in Hong Kong from June to September 2011 to capture passengers’ nonlinear route choice behavior. There are more than 70 transportation stations in Hong Kong, but only nine were chosen for the data collection as the average daily volume of these nine stations is similar and reasonably high (a mean average daily volume of 202 478 passengers). These nine stations also cover almost all of the critical route choice situations in Hong Kong, and such historical data are very important for the ANN model development process. The survey was conducted during the morning peak (07:30– 10:30, GMT+08:00 and the afternoon peak (17:30–20:30, GMT+08:00) on weekdays, when the hourly volume reached a maximum. The average daily volume during weekdays (202 478 passengers) is higher than that at weekends (168 262 passengers), indicating that passenger flow is more critical during weekdays than at the weekend. Twenty-five surveyors conducted the survey to ensure that the sample size was sufficiently large to capture the passenger characteristics, and the survey was completed on time. Fig. 6 shows the concourse of one of the transportation stations. In Zone 1, the distribution of passengers choosing gate groups EG1–EG3 from the entrance was recorded and con-

sidered as the output variable of the ANN model. The data collection procedure was as follows. 1) Surveyor stays near the target area (e.g., Entrance A). 2) Surveyor randomly picks one passenger as he/she enters the station. 3) Surveyor follows the passenger’s path to trace his/her route from the entrance (e.g., Entrance A) to the chosen AFC gate group (e.g., EG1–EG3). 4) Once the passenger has entered the paid area through the AFC gate group, the route of the passenger is recorded. 5) Surveyor returns to the target area and repeats step 1. 6) Surveyor determines the probability of choosing paths from the target area to the three nearest AFC gate groups every 50 records. 7) The procedure is stopped when the probability of passengers choosing the three nearest AFC gate groups in the current 50 records has been calculated, and there is a difference of not more than 5% difference in the accumulated records. In this paper, only those passengers who directly walked from the entrance to one of the three gate groups were recorded for ANN model development; all the intermediate actions (i.e., purchasing at the kiosks, getting services from the customer services center, etc.) and passengers’ bypass (i.e., from one entrance to another entrance) are not included. Eventually, 103 samples were collected (each sample is based on approximately 170 tracking records), and a total of 170 × 103 ≈ 17 500 tracking records were collected in Zone 1 for the model training over the four months of the site survey. IV. M ODEL T RAINING In the site survey, we defined the three gate groups to be EG1, EG2, and EG3, where EG1 is the nearest gate group to an entrance, and EG3 is the most remote gate group. Their corresponding probabilities are P1, P2, and P3. Indeed, the route choice behavior should be the same in different permutation of the gate groups in reality. Therefore, the process of permutation is required for developing a more general ANN model. For the permutation process, the input and output variables with its corresponding gate group are swept and recombined in different orders. Based on this idea, the collected 103 samples should be expanded to 618 samples by considering different permutation of the gate groups, as shown in Table II. Backpropagation [27] (BP) is the traditional training algorithm used for the MLP model. It feeds back the prediction errors from the output layer to the input layer and adjusts the

1294


Fig. 8.

Distribution of the three nearest AFC gate groups.

Fig. 7. Early-stop validation approach.

weights of the links between the neurons. Upon completion of the weight adjustments, a new prediction is carried out to evaluate the new prediction error for the next epoch of weight adjustments. These procedures are repeated numerous times until a satisfactory prediction result is achieved. In this paper, the early-stop validation approach was adopted to monitor and stop the BP training. Fig. 7 shows the concept of the earlystop training approach. To prevent “overfitted training,” the intermediate-state trained model in every training epoch was applied to the validation set to evaluate the prediction error (the validation error). The network training process was stopped when the validation error reached the minimum value. As we had no prior knowledge of the trend of the validation error, “early-stop” training was adopted, which continuously records the status of the model in the course of the training. When there was no reduction in the validation error over a predefined number of epochs (i.e., 150 in this paper), the model state with the minimum validation error was taken as the trained model. Upon completion of the network training, the trained MLP was applied to test samples to evaluate the performance indexes by comparing the target values T of the testing sample and the values P predicted by the trained model. The performance index used was the prediction error in the form of the RMS deviation e calculated by n 2 2 2 2 i,j,k=1 [(Ti − Pi ) + (Tj − Pj ) + (Tk − Pk ) ] e= n (2) where n is the number of performance evaluations and the subscripts i, j, and k refer to the three nearest gate groups EG1–EG3; and the specific accuracy a, calculated by e a= 1− × 100% (3) emax where emax represents the maximum possible error. A prediction error of zero implies that all of the testing samples are correctly predicted. The distribution of the three nearest gate groups is well bounded between the coordination x, y, and z, and the value of emax is geometrically the longest distance between any two points lying on the shaded plane shown in

Fig. 9. Example of the statistical results of the performance evaluation of the ANN model with a 95% confidence limit.

Fig. 8, which is, in fact, √ the distance between any two vertices of the triangle (i.e., 2). It should be noted that a random process is normally involved in the sample extraction process of the MLP model, particularly when the available samples are divided into training and validation sets. It is thus possible for the random process used to result in ¡◦ fortuitous¡± samples that show the evaluated performance indexes to be good. Rather than reporting only the best simulation result, a less prejudiced statistical approach is adopted to minimize the effect of randomization. We carried out 2000 trials of model training and testing (the results converged at 2000 trials). In each trial, the training, validation, and testing samples were randomly grouped. The model was trained and tested using the samples, and the performance index of that trial was obtained. As the performance indexes (the specific accuracy) were well bounded between 0 and 1, a beta distribution was used to represent the probability distribution of the 2000 values of specific accuracy of the model, as shown in the example in Fig. 9. The limit of the one-sided 95% confidence level from the right of the graph (i.e., x95 ) represents the specific accuracy with a 95% confidence level. This statistical approach reduces the effect of randomness in the model performance. V. R ESULTS AND D ISCUSSION Here, we compare the performances between the following models: 1) MLP with the leave-one-out approach; 2) MLP with the ensemble approach; 3) general linear model; and 4) prediction by the shortest distance. A. MLP With the Leave-One-Out Approach As only 618 samples were collected for network training and testing, we used the leave-one-out approach [28], [29] and


1295

Fig. 10. Iteration process of the leave-one-out approach.

the ensemble approach [30] to improve the performance of the MLP model. In the leave-one-out approach, one sample is reserved as the test sample to evaluate the performance of the model being trained by the other 617 samples in each trial. Fig. 10 shows the mechanism of the leave-one-out approach. In general, it is a two-step process. In the first step, the order of the samples is randomly shuffled. This process ensures that the samples used for training, validation, and testing are different for each trial. In the second step, assuming that there are N available samples, one sample is randomly taken out from the pool and kept back to evaluate the performance of the trained model, whereas the other samples (N − 1) are used as the training and validation samples for the model training. Here, 432 (i.e., 617 × 70% = 432) of the samples were used for the network training, 185 (i.e., 617 × 30% = 185) were hidden during the network training phase, and one sample served as the test sample to evaluate the performance of the trained network. The training set was used to train the model with the BP algorithm, and the validation set was used to monitor and stop the BP training using the early-stop validation approach. The test sample did not play a role in the training of the MLP model. Upon completion of the training, the trained MLP model was applied to the taken-out sample to evaluate its performance. Before evaluating the performance of the prediction model, a study of the sensitivity of the performance of the model to the number of hidden neurons was carried out to justify the number of hidden neurons selected, as determined by the rule of thumb [23] in (1), in which the number of training samples Ns was taken as 70% of the available samples initially (i.e., 0.7 × 617 = 432). According to Table I, the number of input Nin and output parameters Nout are 9 and 3, respectively. According to (4), 27 hidden neurons Nh are needed, i.e., √ Nh = (9 + 3)/2 + 432 = 26.78 ≈ 27. (4) The number of hidden neurons investigated ranged from 22 to 32 (i.e., 27 ± 5). For each number of neurons, 2000 trials of model training and performance evaluation, as described in Section IV, were carried out to obtain the 95% confidence intervals of the models with different numbers of hidden neurons. The results are displayed in Fig. 11, which shows that the 95% confidence intervals overlap, indicating that

Fig. 11. Performance of the models with different numbers of hidden neurons. TABLE III S UMMARY OF I NDEPENDENT VARIABLE R ELEVANCE

the performance of the MLP with the leave-one-out approach with different numbers of hidden neurons is comparable. This result demonstrates that the model performance is insensitive to the number of hidden neurons tested. Thus, the number of hidden neurons determined by the rule of thumb (i.e., 27) was adopted. In addition to the investigation of the number of hidden neurons, the importance of the input variables was investigated by the input variable relevance proposed in [41]. Relevance is a way to compare the contribution of each set of input variables in the model. The relevance of an individual input variable is the sum of square of weights for that input variable divided by the sum of square of weights for all input variables. In this paper, the leave-one-out approach with 2000 trials was carried out, and the summary of independent variable relevance is shown in Table III. Table III shows that the contribution of the shortest distance between the entrance and the three nearest gate groups plays the most important role (i.e., 56.53%) among the three set of the input variables. It agrees that the choice of the shortest route is the primary strategy in route selection, whereas the number of gates and the percentage of wide gates in gate groups perform similar contribution in our model (i.e., 22.95% and 20.52%). For the MLP with the leave-one-out approach, 2000 trials were run to ensure that all of the samples were thoroughly tested in each iteration. In each trial, the predicted output was compared with the target output of the test sample to determine the prediction error and the specific accuracy of the model. Over the 2000 trials, the prediction error and specific accuracy of the

1296


Fig. 12. Distribution of the specific accuracy of the model (MLP with the leave-one-out approach) approximated by the beta distribution. Fig. 13.

MLP with the leave-one-out approach were 0.2231 and 84.23%, respectively. These 2000 trials are presented in histogram form in Fig. 12. As the specific accuracy was well bounded between 0 and 1, a beta distribution was used to describe the distribution of the specific accuracy of the model, i.e., 1 xα−1 (1 − x)β−1 /uα−1 (1 − u)β−1 du.

f (x|α, β) =

(5)

0

Applying the beta distribution to describe the distribution of the 2000 trials, the parameters of the distribution were estimated to be α = 6.9811 and β = 1.3281, and the distribution is shown in Fig. 12. The profile of the beta distribution matches reasonably well with the profile of the histogram. The onesided 95% confidence limit of the beta distribution is 60.33%, which is the minimum specific accuracy obtained by the trained MLP model with a 95% confidence level. Based on this result, we conclude that the performance of the MLP model is acceptable. B. MLP With the Ensemble Approach The ensemble approach [30] was also applied to the traditional MLP model to further enhance the performance of the ANN. The general concept of the ensemble approach is shown in Fig. 13. Assuming a total of N samples for the evaluation of the performance, a total of N tests were carried out. In the first test, the first sample is drawn out to be the test sample. In the second test, the second sample is drawn out to be the test sample, and so on. The ensemble approach is a two-step process. In the first step, the jth sample is drawn out as a test sample from the total N samples and is denoted as T j . In the first trial, the remaining N − 1 samples are shuffled and grouped into training samples and validation samples for the ANN model training. Upon completion of the training, the trained ANN

Iteration process for the ensemble approach.

model is applied to predict the output of the test sample. The predicted output is denoted as P1j , where the superscript “(j)” refers to the test with the jth sample, which is taken as the test sample, and the subscript “1” refers to the first trial of this test. In the second trial, the N − 1 samples are shuffled and grouped again. The grouped training and validation samples are used to train another ANN model and applied to the test sample to generate the next predicted output (i.e., P2f ). The process is repeated M times (i.e., 50 times). Eventually, the number of j ) is obtained, and the final predicted outputs (i.e., P1j , P2j , . . . Pm j prediction result of the jth P is the mean of the M results, as shown in the following: P¯ (j) =

M

(j)

Pi /M.

(6)

i=1

The MLP with the ensemble approach gave a prediction error and specific accuracy of 0.1955 and 86.17%, respectively. Using the beta distribution to describe the distribution of the 103 samples, the parameters of the distribution were estimated to be α = 15.9735 and β = 2.56263. The distribution is shown in Fig. 14. The profile of the beta distribution reasonably matches the profile of the histogram. Fig. 14 shows that the one-sided 95% confidence limit of the beta distribution is 71.40%, which is the minimum fraction of correct predictions obtained by the trained MLP model with the ensemble approach. Based on this result, we can conclude that the performance of the MLP model with the ensemble approach is reasonably good. C. General Linear Model To verify the nonlinear nature of the human route choice behavior, a general linear model was developed to compare the performance with our trained MLP models. The general


1297

Fig. 15. Existing layout (in which the location of the AFC gates groups is fixed) of route selection by passengers.

Fig. 14. Distribution of the specific accuracy of the model (MLP with the ensemble approach) approximated by the beta distribution. TABLE IV S UMMARY OF THE R ESULTS OF THE T RAINED MLP M ODEL

than the general linear model and the prediction by the shortest path. Although the mean specific accuracy of the prediction by shortest path is reasonably well (i.e., 80.65%), the limit of the one-sided 95% interval of it is extremely low (i.e., only 1.59%). It represents that the prediction is sometimes very precise but sometimes poor. It is not a good prediction model in practical application. In contrast, the mean specific accuracy of the MLP model with the ensemble approach is 86.17%, whereas the limit of the one-sided 95% interval of it is 71.40%. It shows that the performance of the developed ANN model mimics nonlinear human route selection behavior reasonably well. VI. A PPLICATION OF THE ROUTE C HOICE A RTIFICIAL N EURAL N ETWORK M ODEL TO O PTIMIZE THE S PATIAL D ESIGN OF S TATIONS

linear model is based on our MLP model with a leave-one-out approach but switching the activation function from “sigmoid” to “linear” with 2000 iterations. The mean of the specific accuracy and the one-sided 95% confidence level of the general linear model are 69.07% and 48.26%, respectively. The performance of the general linear model is worse than the MLP model with a leave-one-out or ensemble approach, as shown in Table IV. It proved that the route choice behavior is nonlinear rather than linear in this paper. D. Prediction by the Shortest Path We further compare our trained MLP models with the prediction by the shortest path to determine the effectiveness of the model. For the prediction by the shortest path, we simply assume that all the passengers will choose the shortest gate group as their desired path. Table IV compares the performances of 1) MLP with the leave-one-out approach, 2) MLP with the ensemble approach, 3) general linear model, and 4) prediction by the shortest distance. In general, the performances of the MLP model with the leave-one-out approach and the ensemble approach are better

Here, we demonstrate the application of the developed MLP model with GAs [31] for spatial design. The advantage of GAs over conventional searching algorithms is that GAs can deal with a wide range of areas and are able to find “acceptable” solutions in a reasonable time [32]. They have been also shown to be superior to conventional optimization techniques, particularly for discontinuous and noisy functions [33]. Fig. 15 shows a real situation in which passengers can travel from Entrance A in the unpaid area to the paid area by passing through either gate groups EG1, EG2, or EG3. To fully utilize the gate groups, the designer is required to determine the number of gates in each gate groups, such that the total number of the gates remains unchanged but the expected number of gates to be used by passengers is maximized. The following is the fitness function to be maximized in the GA optimization to achieve this design objective: f (x) = 1/

3

P i Ni

(7)

i=1

where Pi and Ni represent the probability of a passenger using the ith gate group and the number of gates in the ith gate group, respectively. Therefore, the result is converged if the fitness value reaches the minimum. Currently there are a total of 13 gates (EG1 = 6, EG2 = 1, and EG3 = 6) in the existing layout. According to the survey result, the probability of passengers choosing EG1 to EG3 are 99.66%, 0.00%, and 0.34%, respectively, and the fitness value is 0.1667. To utilize fully these 13 gates, (7) is employed, together

1298


TABLE V S UMMARY OF I NPUT AND O UTPUT PARAMETERS

Fig. 16. Optimized layout (in which the location of the AFC gates groups is fixed) of route selection by passengers.

with the input and output parameters, as shown in Table V for the optimization. Furthermore, some constraints are listed as follows to provide a bounded condition for the optimization. 1) Linear equalities are set to ensure that the total number of gates fit the current situation (i.e., EG1 + EG2 + EG3 = 16). 2) The lower boundary is added to ensure that at least one gate serves the specific area (i.e., EG1, EG2, and EG3 1). 3) The upper boundary is added to ensure that the optimized number of gates can suit the available space (i.e., EG1 8, EG2 8, and EG3 8). The optimization process involves both the trained MLP model and GA model. The main purpose of the trained MLP model is to determine the probabilities in route choice [i.e., Pi in (7)], and that of the GA model is to search the number of gates of each gate groups to minimize the fitness function, as shown in (7). After the optimization process, the best fitness value is 0.1449, which is lower than the existing situation (i.e., 0.1667). The optimized number of gates in gate groups EG1 (N1 )–EG3 (N1 ) is 8, 3, and 2, and the corresponding probabilities of P1 –P3 are 0.7953, 0.1273 and 0.0774, respectively. The optimized layout between Entrance A and AFC gate groups EG1–EG3 and the process of optimization are shown in Figs. 16 and 17, respectively. VII. C ONCLUSION In this paper, we have proposed an ANN model to predict the route choice behavior in a transportation station. A fourmonth survey was conducted in nine transportation stations in Hong Kong during peak hours to determine the distribution of passengers between the entrance and the three nearest gate groups by following passengers and recording their chosen route. Based on these historical data, an MLP model was established both with the leave-one-out approach and the ensemble approach in which the shortest distance between the entrance and the gate groups, the number of gates, and the percentage of wide gates in the gate groups were considered as input parameters. The model trained by the data from the

Fig. 17.

Fitness value versus generation during the optimization process.

stations captures the general behavior of the stations; we do not expect that the intelligent model can provide any accurate flow pattern predictions. However, it can serve as an intelligent route choice engine and be implemented in any pedestrian flow model for detail movement simulation, which will account for the interaction with kiosk, ticket counters, and barrier locations. Both the MLP model with the leave-one-out approach and the MLP model with the ensemble approach successfully mimicked passenger decision-making on route choice. The mean of the specific accuracy of the ensemble approach reached 86.17%. In addition, the percentage of correct predictions was statistically justified, and the specific accuracy with a onesided 95% confidence level of the leave-one-out and ensemble approaches is 60.33% and 71.40%, respectively. The applicability of this trained ANN model has been successfully demonstrated through the optimization of the number of gates in each gate group to maximize the use of the gate groups. This model will be further developed in the next stage of our work to maximize the use of gate groups by considering not only the unpaid area but also the area connecting the gates and the platforms. R EFERENCES [1] D. Helbing and P. Molnar, “Social force model for pedestrian dynamics,” Phys. Rev. E, vol. 51, no. 5, pp. 4282–4286, May 1995. [2] S. Wolfram, “Statistical mechanics of cellular automata,” Rev. Mod. Phys., vol. 55, no. 3, pp. 601–644, Jul.–Sep. 1983.


[3] Y. Tajima and T. Nagatani, “Scaling behaviour of crowd flow outside a hall,” Phys. A, Stat. Mech. Appl., vol. 292, no. 1–4, pp. 545–554, Mar. 2001. [4] G. E. Bradley, “A proposed mathematical model for computer prediction of crowd movements and their associated risks,” in Proc. Eng. Crowd Safety, 1993, pp. 303–311. [5] R. L. Goldstone and M. A. Janssen, “Computational models of collective behaviour,” Trends Cogn. Sci., vol. 9, no. 9, pp. 424–430, Sep. 2005. [6] S. M. Lo, Z. Fang, P. Lin, and G. S. Zhi, “An evacuation model: The SGEM package,” Fire Safety J., vol. 39, no. 3, pp. 169–190, Apr. 2004. [7] S. M. Lo, H. C. Huang, P. Wang, and K. K. Yuen, “A game theory based exit selection model for evacuation,” Fire Safety J., vol. 41, no. 5, pp. 364– 369, 2006. [8] W. Daamen, P. H. L. Bovy, and S. P. Hoogendoorn, “Influence of changes in level on passenger route choice in railway stations,” J. Transp. Res. Board, vol. 1930, pp. 12–20, 2005. [9] P. A. Thompson and E. W. Marchant, “A computer model for the evacuation of large building populations,” Fire Safety J., vol. 24, no. 2, pp. 131– 148, 1995. [10] S. Gwynne, E. R. Galea, P. J. Lawrence, and L. Filippidis, “Modelling occupant interaction with fire conditions using the building EXODUS evacuation model,” Fire Safety J., vol. 36, no. 4, pp. 327–357, Jun. 2001. [11] H. Entamo, S. Heliövaara, S. Hostikka, and T. Korhonen, “Modeling evacuees’ exit selection with best response dynamics,” in Pedestrian and Evacuation Dynamics 2008. Berlin, Germany: Springer-Verlag, 2010, pp. 309–319. [12] J. de Dios Ortuzar and L. G. Willumsen, Modelling Transport. New York, NY, USA: Wiley, 1994. [13] J. D. Sime, “Movement toward the familiar person and place affiliation in a fire entrapment setting,” Environ. Behav., vol. 17, no. 6, pp. 697–724, Nov. 1985. [14] G. Proulx, “A stress model for people facing a fire,” J. Environ. Psychol., vol. 13, no. 2, pp. 137–147, Jun. 1993. [15] S. Gwynne, E. R. Galea, M. Owen, and L. Filippidis, “Adaptive decisionmaking in response to crowd formations in building EXODUS,” J. Appl. Fire Sci., vol. 8, no. 4, pp. 301–325, 1999. [16] S. P. Hoogendoorn and P. H. H. Bovy, “Pedestrian route-choice and activity scheduling theory and models,” Transp. Res. B, Methodol., vol. 38, no. 2, pp. 169–190, 2004. [17] Z. Zarita and E. A. Lim, “Intelligent exit-selection behaviours during a room evacuation,” Chin. Phys. Lett., vol. 29, no. 1, pp. 018901-1– 018901-4, Jan. 2012. [18] G. D. Garson, Neural Complex Transportation Systems: An Introductory Guide for Social Scientists. Thousand Oaks, CA, USA: Sage, 1998. [19] F. Rosenblatt, Principles of Neurodynamics. Washington, DC, USA: Spartan, 1962. [20] M. W. Gardner and S. R. Dorling, “Artificial neural networks (the multilayer perceptron)––A review of applications in the atmospheric sciences,” Atmos. Environ., vol. 32, no. 14/15, pp. 2627–2636, Aug. 1998. [21] K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward network are universal approximators,” Neural Netw., vol. 2, no. 5, pp. 359– 366, Jan. 1989. [22] G. Q. Zhang, B. Patuwo, and M. Y. Hu, “Forecasting with artificial neural networks: The state of art,” Int. J. Forecast., vol. 14, no. 1, pp. 35–62, Mar. 1998. [23] Neuroshell 2 Manual, Ward Syst. Group Inc., Frederick, MA, USA, 1996, Ward. [24] M. R. Hill, “Spatial structure and decision-making of pedestrian route selection through an urban environment,” Ph.D. dissertation, ETD Collection, Univ. Nebraska, Lincoln, NE, USA, Jan. 1, 1982, paper AAI8306484. [Online]. Available: http://digitalcommons.unl.edu/ dissertations/AAI8306484 [25] C. Y. Cheung and H. K. Lam, “Pedestrian routes choices between escalator and stairway in MTR stations,” J. Transp. Eng., vol. 124, no. 3, pp. 277–285, May/Jun. 1998. [26] J. J. Fruin, Pedestrian Planning and Design. New York, NY, USA: Metropolitan Association of Urban Designers and Environmental Planners, 1971. [27] P. J. Werbos, “Beyond regression: New tools for prediction and analysis in the behavioural sciences,” Ph.D. dissertation, Committee Appl. Math., Harvard Univ., Cambridge, MA, USA, 1974. [28] L. Breiman, “Bagging predictors,” Mach. Learn., vol. 24, no. 2, pp. 123– 140, Aug. 1996. [29] B. Efron, “Estimating the error rate of a prediction rule: Improvement on cross validation,” J. Amer. Stat. Assoc., vol. 78, no. 382, pp. 316–331, 1983.

1299

[30] A. K. Jain, R. C. Dube, and C. C. Chen, “Bootstrap techniques for error estimation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-9, no. 5, pp. 628–633, Sep. 1987. [31] J. H. Holland, Adaptation in Natural and Artificial Systems. Cambridge, MA, USA: MIT Press, 1975. [32] D. Beasley, D. R. Bull, and R. R. Martin, “An overview of genetic algorithms: Part 1, fundamentals,” Univ. Comput., vol. 15, no. 2, pp. 58–69, 1993. [33] R. K. K. Yuen, E. W. M. Lee, S. M. Lo, and G. H. Yeoh, “Prediction of temperature and velocity profiles in single compartment fire by the intelligent system,” Fire Safety J., vol. 41, no. 6, pp. 478–485, Sep. 2006. [34] R. K. K. Yuen, E. W. M. Lee, C. P. Lim, and G. W. Y. Cheng, “Fusion of GRNN and FA for online noisy data regression,” Neural Process. Lett., vol. 19, no. 3, pp. 227–241, Jun. 2004. [35] V. Punzo and B. Ciuffo, “Integration of driving and traffic simulation: Issues and first solutions,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 2, pp. 354–363, Jun. 2011. [36] Q. Song and X. Wang, “Efficient routing on large road networks using hierarchical communities,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 1, pp. 132–140, Mar. 2011. [37] L. Figueiredo, I. Jesus, J. A. T. Machado, J. R. Ferreira, and J. L. M. de Carvalho, “Towards the development of intelligent transportation systems,” in Proc. IEEE Intell. Transp. Syst. Conf., 2001, pp. 1206–1211. [38] I. Chabibi and S. Lan, “Adaptations of the A ∗ algorithm for the computation of fastest paths in deterministic discrete-time dynamic networks,” IEEE Trans. Intell. Transp. Syst., vol. 3, no. 1, pp. 60–74, Mar. 2002. [39] S. Kim, E. Lewis, and C. C. White, “State space reduction for nonstationary stochatstic shortest path problems with real-time traffic information,” IEEE Trans. Intell. Transp. Syst., vol. 6, no. 3, pp. 273–284, Sep. 2005. [40] J. G. Juang and K. C. Cheng, “Application of neural networks to disturbances encountered landing control,” IEEE Trans. Intell. Transp. Syst., vol. 7, no. 4, pp. 582–588, Dec. 2006. [41] S. L. Özesmi and U. Özesmi, “An artificial neural network approach to spatial habitat modelling with interspecific interaction,” Ecol. Model., vol. 116, no. 1, pp. 15–31, Mar. 1999.

J. K. K. Yuen received the Bachelor’s degree in building service engineering from City University of Hong Kong, Kowloon, Hong Kong, in 2008. He is currently working toward the Ph.D. degree with the Department of Civil and Architectural Engineering, City University of Hong Kong. His research interests include artificial intelligence, pedestrian behaviors, crowd movement, and evacuation analysis in transportation stations.

E. W. M. Lee received the B.Eng. (Hons.) degree in building services engineering and the Ph.D. degree in fire engineering from the Department of Civil and Architectural Engineering, City University of Hong Kong, Hong Kong, in 2000 and 2004, respectively. Before pursuing his Ph.D. studies, he was an Executive Engineer in one of the leading consultant firms in Hong Kong. He is currently an Assistant Professor with the Department of Civil and Architectural Engineering, City University of Hong Kong, Kowloon, Hong Kong. His research interests include the application of soft computing techniques in building engineering. Mr. Lee is a Registered Professional Engineer in Hong Kong.

1300


S. M. Lo received the M.Sc. degree in construction management from the City Polytechnic of Hong Kong, Hong Kong, in 1993 and the Ph.D. degree in architecture from the University of Hong Kong in 1997. Before pursuing his Ph.D. studies, he was in the building industry for many years. He is currently a Professor with the Department of Civil and Architectural Engineering, City University of Hong Kong, Kowloon, Hong Kong. He is also an Authorized Person registered under the Hong Kong Buildings Ordinance. His research interests include spatial planning for pedestrian flow traffic and evacuation modeling, real estate and building development, project management, building defect diagnosis, sustainable development, and fire safety engineering.

R. K. K. Yuen received the B.Sc. (Hons.) and M.Sc. degrees in mechanical engineering from the University of Hong Kong, Hong Kong, in 1981 and 1991, respectively, and the Ph.D. degree in mechanical engineering from the University of New South Wales, Australia, in 1998. He is currently an Associate Professor with the Department of Civil and Architectural Engineering, City University of Hong Kong, Kowloon, Hong Kong. His research interests include building energy conservation, building fire modeling, combustion, pyrolysis, computation fluid dynamics applications in building design, and heating, ventilation, and air conditioning system simulation. Mr. Yuen is a Registered Professional Engineer in Hong Kong.