Driving licensing renewal policy using neural network-based probabilistic decision support system

Int. J. Computer Applications in Technology, Vol. 51, No. 3, 2015

155

Driving licensing renewal policy using neural network-based probabilistic decision support system Wa’el H. Awad Faculty of Engineering Technology, Department of Civil Engineering, Al Balqa’ Applied University, P.O. Box 15008, Amman 11134, Jordan Email: [email protected]

Randa Herzallah* Non-linearity and Complexity Research Group, Aston University, Birmingham B4 7ET, UK Email: [email protected] *Corresponding author Abstract: This paper investigates neural network-based probabilistic decision support system to assess drivers’ knowledge for the objective of developing a renewal policy of driving licences. The probabilistic model correlates drivers’ demographic data to their results in a simulated written driving exam (SWDE). The probabilistic decision support system classifies drivers’ into two groups of passing and failing a SWDE. Knowledge assessment of drivers within a probabilistic framework allows quantifying and incorporating uncertainty information into the decision-making system. The results obtained in a Jordanian case study indicate that the performance of the probabilistic decision support systems is more reliable than conventional deterministic decision support systems. Implications of the proposed probabilistic decision support systems on the renewing of the driving licences decision and the possibility of including extra assessment methods are discussed. Keywords: driving knowledge; licensing renewal; probabilistic decision support system; uncertainty. Reference to this paper should be made as follows: Awad, W.H. and Herzallah, R. (2015) ‘Driving licensing renewal policy using neural network-based probabilistic decision support system’, Int. J. Computer Applications in Technology, Vol. 51, No. 3, pp.155–163. Biographical notes: Wa’el H. Awad received his Bachelor degree in 1986 from University of Jordan in Civil Engineering; Master and PhD from University of Colorado at Denver, 1991 and 1997, respectively. Served as the Dean of Engineering at Al-Ahliyya Amman University, on Sabbatical leave from Al Balqa’ Applied University. He is a senior traffic and transportation expert with more than 25 years of experience in traffic and traffic safety. Areas of expertise in addition to traffic and traffic safety, including traffic simulation, human behaviour, artificial intelligence applications in transportation, feasibility studies, traffic impact studies, engineering economy and applied statistics. Randa Herzallah received her BSc in Industrial Engineering and MSc in Industrial Engineering/Manufacturing and Design at Jordan University, Jordan, and later her PhD from Aston University, Birmingham, UK. She was a Part-time Lecturer in the Applied Science University, Jordan and a Full-time Lecturer in Al-Balqa’ Applied University, Jordan in 1998 and 1999, respectively. She rejoined Al-Balqa’ Applied University, Jordan in 2003. She is currently a Lecturer in the Faculty of Engineering and Applied Science at Aston University, Birmingham, UK. Her current research interests are non-linear control theory, intelligent control, stochastic control, adaptive neural control and optimal control.

Copyright © 2015 Inderscience Enterprises Ltd.

156

1

W.H. Awad and R. Herzallah

Introduction

Traffic safety is a global concern owing to the complexity of the relationship among road, vehicle and driver. Countermeasures have been taken at all levels to reduce crash frequency and severity; drivers are taking most of the burden when it comes to paying the price of a crash whether this price is direct or indirect. Driver training and education before and after becoming licenced plays a big role in improving traffic safety (Siegrist, 1999; McKenna, 2010). McKenna distinguished between training (which is concerned with skill acquisition) and education (which is concerned with knowledge acquisition) in the driving context. Driver’s knowledge consists of a wide range of stored information, including recognition, skilled performance, rules and principles. The knowledge builds up continuously as driver receives instruction, training and experience in the driving process. Aging among other factors contributes to the deterioration of driving capabilities of most drivers. The driving capabilities are either driving skills or knowledge of driving rules and regulations. Qualified drivers become licenced after passing two important exams (after proving physical fitness); theoretical written driving test and practical driving test on the road. Driver’s knowledge influences other driver qualities and reflects on driving behaviour. This research is concerned with drivers’ knowledge in Jordan. The objective here is to develop a reliable decision support system to assist in the decision of renewing the driving licences in Jordan. To achieve this, we propose a novel probabilistic formulation for constructing a probabilistic decision support system. In this formulation, the drivers are grouped into two groups of passing and failing a SWDE. The probabilistic decision support system is then constructed such that it has the ability to discriminate between these two groups of drivers and classifies them into either of the two groups. This nonstandard formulation of the proposed probabilistic decision support system allows quantifying and incorporating uncertainty information into the decision-making process. Compared with its deterministic version, the probabilistic decision support system provides a theoretical foothold for constructing more reliable life-critical decision support tools. International experiences regarding driver licensing procedure are almost common where applicants must pass knowledge test and driving test before being authorised to drive. A Swedish study (Wolming and Wiberg, 2004) concluded that the Swedish drivers’ licensing examination is suitable for a two-stage testing model (knowledge test and a driving test). Considering the age of applicants, graduate driving licensing (GDL) programs were initially introduced in the USA (Florida in 1996) to supervise novice drivers (Preusser and Tison, 2007; Hedlund, 2007). Major concerns towards elderly drivers’ physical and mental capabilities while driving constitute applying restrictions (Hanson and Hildebrand, 2011).

Licensing renewal provisions in the USA (www.iihs. org/laws/olderdrivers.aspx) varies among states according to two aspects: length of time between renewals and additional requirements that might be imposed on older drivers (typically 65 or 70). The report shows that the length of regular renewal cycle vary from a minimum of four years (e.g., Alabama, Arkansas, Idaho, Illinois and many others) to a maximum of 10 years (e.g., Colorado and South Carolina). The effectiveness of written driver knowledge test was examined more than 30 years ago by Virginia Highway and Transportation Research Council (Stoke, 1979) through conducting an experimental evaluation to randomly assigned members of the renewal population of drivers. The outcome of this study based on studying the performance of different groups of drivers at four subsequent time intervals showed few statistically significant differences, none of which suggest the practicality of knowledge testing as an effective highway safety countermeasure. Driver relicensing is essential in all cases where driving jeopardises the safety of the driver and/or his surroundings. Related literature investigated cases of elderly drivers; violators; and drivers involved in traffic crashes. In a case study (Stamatiadis et al., 2003), authors suggested systematic review of current practices in the USA regarding licence renewal and retesting and identified potential changes in these procedures. A more frequent renewal period was proposed for older people. Road crashes are considered the major public health problem facing the world, According to the WHO annual report (WHO, 2009); over 1.2 million people lose life each year on the world’s roads and more than 20 million suffer injuries. Over 90% of deaths occur in low income and middle income countries, the report indicates that about two-thirds of the countries do not have a national road safety strategy to address the traffic safety problem. Linking drivers’ knowledge to road crashes was investigated by many researches. Martinez and Porter (2004) examined pedestrian crashes in Virginia from 1990 to 1999 to investigate the significance of variables believed to predict these crashes. Phone interviews were conducted to assess drivers’ knowledge towards certain pedestrian safety issues. Another study concluded that road crashes are as likely to be related to driver personality variables as they are to the knowledge of vehicle operation and rules and regulations (Arthur and Doverspike, 2001). Norwegian drivers’ knowledge concerning their perceptions of enforcement of speed limits (Jørgensen and Pedersen, 2005) was examined against different driver characteristics concluding that older drivers had less knowledge about the threshold level of serious speeding, but more knowledge about the detection rate than younger drivers do. Jordanians acknowledge the shocking situation of traffic safety in the country. Authorities are holding road users accountable for most of traffic crashes. Although it is unproven, authorities are not taking any significant steps on the ground to elevate the problem and improve traffic

Driving licensing renewal policy using neural network-based probabilistic decision support system safety. Local statistics (Jordan Traffic Institute, 2011) shows the continuous increase in crash numbers since the early 1980s. The rate of increase accelerated since year 2003 up to this date (Figure 1). Figure 1

Traffic accidents in Jordan (see online version for colours)

Source: Jordan Traffic Institute (2011)

The number of registered vehicles in Jordan is increasing steadily from year to year; passing the 1 million registered vehicles before year 2010 (Figure 2). The number of licenced drivers in Jordan (Jordan Traffic Institute, 2011) was 1,744,852 drivers (28% of Jordan’s population). This ratio is much less than many international ratios, for example, the FHWA reporting an average of 68.5% (FHWA, 2008). As of 2010, the ratio of registered vehicles in Jordan to inhabitants (motorisation rate) was 17.6%. Globally, Jordan ranked 81, while USA ranked on the top (765 vehicles to 1000 population). The licenced drivers to registered vehicles ratio in Jordan is 62% compared with 89% in the USA (http://www.nationmaster.com/graph/ tra_mot_veh-transportation-motor-vehicles). Figure 2

Vehicles in Jordan (see online version for colours)

Source: Jordan Traffic Institute (2011)

There is no clear national policy in Jordan towards controlling the dreadful situation of traffic safety. In view of the steady increase in population, registered vehicles, and licenced drivers, it would be unfair to blame drivers as the sole contributing factor to the safety problem in Jordan. The quality of imported and registered vehicles that operate on Jordan’s roads is a question of concern that has never been

157

investigated. In this paper the issue of concern that will be investigated is the drivers’ competencies in terms of traffic knowledge. A previous research question of interest was raised on whether the drivers in Jordan possess adequate driving knowledge or not? Two previous research papers in Jordan investigated this issue (Awad and Al-Kharabsheh, 2001; Awad et al., 2004a). The first paper conducted a pilot study to investigate drivers’ traffic knowledge in Jordan through simulating a written driving licence exam similar to the official one. More than 96% of subjects who were administered to the exam failed to pass. The study recommended imposing higher standard of knowledge for traffic rules on applicants for new driver licence in Jordan, and a nationwide program to assess the relationship between driver knowledge; driver behaviour; and crash and fatality rates. The second paper administered a modified version of the written driving licence exam to 953 drivers holding valid Jordanian driving licence. The subjects were randomly selected from all over Jordan to examine the relationship between their scores in the exam and a set of factors grouped into three categories (human, exposure and safety). Despite of the past experience of the drivers administered to the exam; more than 90% were unable to score over 90% (the passing score required in official driving test). The overall mean of scores was 63.9 (SD = 13.3). The scores showed significant differences with gender, education, using seat belt, crash severity, years of driving, average daily driving distance, and daily trip length. On the other hand, this research failed to find any significant differences between SWDE scores and each of age, occupation, type of driving licence, driving record or crash involvement. The authors believed that a great deal of the traffic safety problem in Jordan can be explained by the diversity between drivers, including cultural background, skills and qualifications, motivation and crash risk. The application of Artificial Intelligence Techniques in traffic engineering and traffic safety is not new. A previous study (Awad and Janson, 1998) explored previous attempts to employ AI techniques in predicting and forecasting traffic and modelling traffic problems. In that study, authors employed different techniques (conventional regression and AI) to predict truck accidents at freeway ramps in Washington State. Later on, artificial intelligent techniques have been used to provide rich, powerful and robust alternative tools to model traffic engineering and traffic safety applications. Using the artificial neural network (ANN) approach, Ross et al. (1998) developed ANN and regression models to estimate the traffic emission rates of CO, HC and NOx. Song and Gao (2013) developed a visual method for estimating traffic visibility based on homogenous area extraction. Ivan and Sethi (1998) used ANN with the backpropagation algorithm for traffic incident detection. Kalyoncuoglu and Tigdemir (2004) estimated the number of drivers involved in traffic accidents using driver

158

W.H. Awad and R. Herzallah

characteristics (i.e., age, gender, education, driving experience and driving time per day). Basu and Maitra (2006) modelled stream speed in the heterogeneous traffic environment using ANN-lessons learnt. Sazi (2006) developed an ANN transport energy demand model for Turkey. Zhao (2012) developed a prediction model based on interval type-2 fuzzy neural network and self-organising learning algorithm to predict short term traffic flow. Genetic algorithm technique has also been adapted for a large number of applications in road safety engineering. The GA approach was employed to analyse driver’s behaviour in collision avoidance (Nagai et al., 1997). Another research (Chapman et al., 2002) used visual search patterns technique to report a training intervention that informs novice drivers about their typical patterns of visual search. Awad (2004) used neural networks technique to estimate traffic capacity on weaving segments, also an adaptive neuro-fuzzy inference system for traffic cycle optimisation was developed by Awad et al. (2004b). Another recent study used fuzzy signal detection theory to investigate drivers’ hazard perception ability (Wallis and Horswill, 2007). Fuzzy signal theory was employed to determine why experienced and trained drivers respond faster than novice drivers in a hazard perception test. Two models were developed to discriminate between novice drivers and experienced drivers, and trained and untrained drivers. The authors concluded that novice drivers are poorer at discriminating more hazardous from less hazardous situations than experienced drivers. And novice drivers require a higher threshold of danger to be present before they notice. Blinova (2007) used neural networks to forecast passenger traffic demand and proved the adequacy of the Neural Network model to forecast for the short-term, the Russian air transport network. Akgüngör and Doğan (2009) used artificial intelligence and genetic algorithm techniques to estimate future accidents and fatalities in Ankara, Turkey. The validity of driver licence in Jordan is 10 years regardless of the driver’s age, number of violations or involvement in traffic crashes. Those drivers can renew their driver licence after passing visual adequacy test. No rules of thump, or current regulation can revoke a valid driving licence in Jordan based on aging or other demographic characteristics including involvement in traffic crashes. This research paper attempts to develop a decisionmaking tool (classifier) to help identifying whether or not a Jordanian driver with a valid driver licence is holding enough driving knowledge to entitle him or her for driving and help in taking the renewal decision of his or her driving licence. The assessment of driving knowledge is based on a simulated written driving exam (SWDE) similar to the one administered by Jordanian authorities. None of the previous studies have considered such classifiers as predictors to recommend that the driver should go through another driving test after some period of time to be qualified for driving again. Some of previous local studies tried to correlate, using simple statistics methods,

between drivers’ knowledge and specific demographic characteristics of the driver (Awad and Al-kharabsheh, 2001; Awad et al., 2004a). In this paper, authors use neural network-based classifiers that group drivers into two groups of passing and failing a SWDE. Furthermore, using recent development of neural network the objective then is to assess the reliability of those classifiers. To achieve the above objectives, a sample of Jordanian drivers’ data compiled through the Jordan Traffic Institute (JTI) was used to construct classifiers capable of predicting the probability distribution of the exam result rather than a single estimate. This means that the classifier will provide a measure of uncertainty about predicting exam results from different samples. Multilayer perceptron neural network will be used to construct those classifiers. This study attempts to identify drivers who are expected to fail a SWDE as not having adequate driving knowledge to hold a driving licence and at the same time to assess the reliability of those classifiers by measuring the uncertainty of their predicted outputs.

2

Method

2.1 Participants Four hundred and forty-eight Jordanian drivers were administered to a SWDE. The sample consisted of 30 females and 418 males, with an average age of 34.6 years (ranging between 19 and 71 years of age). All participants possessed a valid Jordanian driving licence, and on average had been driving for 15.2 years (ranging between 3 and 50 years). The participants’ education varies from high school to holding higher degrees and working in a wide spectrum of professions. About 25% of participants were holding commercial driving licence. About 34% of participants have never been ticketed for violations and ~74% have never been involved in a crash.

2.2 Procedure A simulated driving written exam was administered to participants while they were visiting the Motor Vehicle Department offices in nine of the major provinces in Jordan. The exam comprises 25 multiple choice questions that cover driving rules, skills, and ethics. In addition to the exam, participants were asked to complete a demographic questionnaire. Participants completed the questionnaire within 20 minutes. Drivers’ demographic data were reported such as age, gender, education, occupation, driving record and crash history and date of acquiring driving licence, travelling pattern (daily route, total hours of daily travel, average daily travelling distance, number of daily trips, average trip length, purpose of trip and time of trips); driving pattern (ability to use other modes of travel such as public transportation, vehicle type, using seat belt, respecting speed limit and control system). The maximum possible score in the written driving exam is 100 for 25 correct answers. The

Driving licensing renewal policy using neural network-based probabilistic decision support system mean score of participants was 65 (ranging between 24 and 92). The histogram of the scores of the various drivers is shown in Figure 3. Figure 3

Histogram of the drivers scores in the simulated written driving exam (see online version for colours)

Lowe (2008), where neural network models are used to provide a prediction for the conditional expectation of the drivers’ results in the exam and the uncertainty of its residual error. For the classification problem of drivers’ result in the exam, a classifier is constructed using a three layer multi-layer perceptron neural network. The classifier is constructed with nine inputs which are found to be highly correlated with the drivers’ knowledge. These inputs are: age, gender, occupation, education, licence type, daily distance, seat belt, number of violations and accidents. The output of the classifier is a two-component vector with desired target values of y = {y1, y2}, where y1 = [1, 0] and y2 = [0, 1] for passing and failing the SWDE, respectively. The architecture of this classifier is shown in Figure 4. The number of hidden neurons of the classifier is optimised using the k-fold validation method as will be discussed in the next section. Figure 4

It is well known that a crucial property for constructing a reliable estimator is the persistence of excitation of the available data. Persistence of excitation guarantees convergence of the estimator parameters. However, Figure 3 shows that 80% of the drivers’ scores are clustered in the interval of [50–80]. There is almost no data available in the ranges between [0–30] and [90–100], 12% of the data are in the interval [30–50] and 7% of the data are in the interval [80–90]. This means that any constructed estimator using these data will be biased and unreliable. In this paper, since the renewing decision of the driving licence depends only on whether a driver passes a written driving exam or not, drivers’ scores were grouped into two groups (passing and failing) using the mean of the scores values as the cut point. Moreover, grouping drivers into two groups of failing and passing a written driving exam helps also in minimising the persistence of excitation problem. The demographic data was ranked according to their correlation with exam scores, and the nine mostcorrelated data were used to construct a classifier discriminating between drivers passing and failing a written driving exam.

2.3 Methodology 2.3.1 Probabilistic classifiers To assess the reliability of the drivers’ classifier the probability distribution of a driver’s result in the SWDE rather than a single estimate of the mean value should be constructed. This means that a network model should be identified such that it predicts the mean value of the result in a SWDE and at the same time provides a prediction for the variance around the predicted mean value. This would provide us with a measure of confidence in a network prediction. The method adopted for this purpose is based on the method proposed by Herzallah (2007) and Herzallah and

159

Classifier multi-layer perceptron network (see online version for colours)

The parameters of the classifier network are then adjusted using an appropriate gradient-based method to optimise a performance function based on the error between the desired target values and the output of the classifier network. After optimisation, the predicted output of this classifier network is simply the conditional expectation of the desired output of the classifier. Therefore, denoting expectation using angle bracket notation, the following stochastic model can be constructed. yx = yx | x + ε ,

where x is the drivers’ input for the classification problem, y x | x is the conditional expectation of the desired output as predicted from the classifier conditioned on a specific

160

W.H. Awad and R. Herzallah

input value, ε represents the residual error of the classifier output which is assumed to be Gaussian random noise of zero mean and Σ covariance. This term, in other words, represents uncertainty of the classifier. Once the first multi-layer perceptron network (classifier) is optimised on predicting the result of a driver in a SWDE, the error of each driver using that model becomes available. Hence, these input dependent residual errors from the classifier are used as target values for another neural network model which has the same structure and same inputs as the classifier network. The parameters of this second network model are then adjusted using an appropriate gradient-based method to optimise a performance function based on the error between its output and the calculated input dependent residual error values from the classifier network. Therefore, training a second network on this set of error values gives a model with outputs that optimally attempts to reproduce the expected value of the target conditional on the input. In this case this conditional expectation is an estimate of the first network variance. For a given driver, the second network produces an estimator to the following conditional variance expectation: Σ = || yx − yx | x ||2 | x .

In this paper, the network estimator of this quantity is referred to as the predictive error bars. After optimised, this second network model can then be used for predicting the error of an arbitrary driver not part of the training set. Following this approach the probability distribution of the driver’s result in a SWDE will be Gaussian distribution with predicted means provided by the classifier network and an input dependent variance given by the predicted residual error values from the second network.

2.3.2 Model validation and selection The k-fold cross validation method was used for determining the performance of the classifiers and choosing the number of hidden neurons in the hidden layer. The available data are divided into two sets: the validation and training set containing 400 drivers’ data and the test set containing 48 drivers’ data. The training and validation dataset is then broken into five partitions and the parameters of the classifiers were optimised based on data from 320 drivers and validated on the remaining 80 drivers left out from the training process. The experiment was repeated for each validation set in turn and the mean validation error is calculated in terms of the number of misclassified data. On the basis of the results of the average number of miss-classifications in the validation set, the best optimal structure of the neural network is found to consist of 23 hidden neurons in the hidden layer. It is worth mentioning that a linear classifier with zero hidden neurons and linear mapping between its inputs and outputs was also included in the validation process where it is found

to give the highest number of miss-classifications in the validation set.

2.3.3 Classifier results On the basis of the obtained validation results, the expected value of the driver pass or fail a SWDE is predicted from a multi-layer perceptron neural network with 23 hidden neurons, which is found to be the optimal structure. As discussed in Section 2.3.1 another multi-layer perceptron neural network with 23 hidden neurons is trained on the input dependent residual error values to provide a prediction for uncertainty of drivers’ in the test set. Figure 5 provides a plot for the predicted values of passing the SWDE from the classifier network for each driver in the test set. The figure also simultaneously plots the driver-specific predictive error bars as obtained from the second multi-layer perceptron network. The figure orders the drivers from smallest to largest predicted output value. The output of the classifier is labelled with open circles for class one (pass a SWDE) and with a square for class two (fail a SWDE) with vertical lines showing the driver specific predictive variance. This figure gives a classification rate of 62.5% accuracy with 13 drivers from class one and 5 drivers from class two classified incorrectly. Table 1 shows the confusion matrix which summarises the classifier performance. Table 1

Confusion matrix showing the classification performance of the drivers exam result

Predicted classes True classes

Class 1

Class 2

Class 1

15

13

Class 2

5

15

The predictive error bars of the drivers’ result in the exam shown here are very high. It can be seen that shifts in the predicted exam result within the predictive error bars could have a major impact on classification rates, simply based on how the model was sampled, giving a different model within the variance bars. These additional confidence value estimates indicate the ability of the classifier to correctly predict the drivers’ result in the SWDE. Only a small number of drivers can be predicted to pass or fail the SWDE with confidence. The remaining drivers who have large predictive error bars are intrinsically unclassifiable since any attempt to provide an exam result must be effectively random. In short, although an exam result estimator can be obtained on a driverspecific basis as the mean of a distribution, the variance of that distribution is so large as to make using the mean meaningless as a representative of the distribution. Only 11 drivers in Figure 5 have low error bars, the remaining drivers have large uncertainties which would undermine the classification results if another random

Driving licensing renewal policy using neural network-based probabilistic decision support system sample from the model space was selected inside the error bar. Figure 6 shows the results where the drivers are ordered based on the magnitude of the variance values from left to right with drivers with smallest predictive error bars being on the left in the figure. Figure 5

Classifier results of drivers passing a simulated written driving exam, ordered by the value of passing the exam output. Open circles represent pass the exam drivers and squares represent fail the exam drivers. Predictive error bars are shown as vertical lines (see online version for colours)

161

driver misclassified into fail the exam group. In this figure, only about fifth of these drivers can be confidently classified into their exam result groups. These drivers instead of being classified into specific groups should therefore be labelled as unclassifiable and further assessment methods are required to help make a decision. Table 2 shows the confusion matrix of only the drivers with high confidence values attached. It can be seen that the classification rate has increased significantly to about 73%. Hence after, using only high confidence drivers in predicting the outcome will give more reliable results. Alternatively, forcing unclassifiable drivers into groupings will deteriorate the outcome of the process and harm the drivers and the community by making the wrong decision. Table 2

Confusion matrix showing the classification performance of the drivers’ exam result using only those drivers with high confidence

Predicted classes

Figure 6

Classifier results of drivers passing a simulated written driving exam, ordered by the value of the variance. Open circles represent pass the exam drivers and squares represent fail the exam drivers. Predictive error bars are shown as vertical lines (see online version for colours)

It can be seen that drivers on the left can confidently be separated between the two groups of passing and failing the SWDE. Drivers with large error bars cannot be confidently classified into any group. Nevertheless, confidence levels of drivers can be very useful in augmenting the classification result. There will usually be some outliers which perform differently. For example, Figure 6 depicts two highly confident but fail the exam drivers misclassified into pass the exam group and one highly confident but pass the exam

True classes

Class 1

Class 2

Class 1

3

2

Class 2

1

5

3

Discussion

The supervised classifier proposed in this paper shows that the driver-specific predicted result can be unreliable owing to the persistence of excitation problem and because of the limited drivers data, which does not span the entire approximation space. The method of estimating confidence intervals by calculating the residual error from the predicted output of the classifier is proved viable to provide any reliable prediction. Using this additional information about uncertainty measures helps prediction in decision support systems to become more reliable. Figures 5 and 6 show that only a small number of drivers can be classified into either of the groups based on their predicted variance value. The set of drivers with large uncertainties constitute ~80% of all available drivers. These drivers should not be classified into specific groups based simply on the predicted mean from the classifier. They should instead be labelled as unclassifiable and further assessment methods should be claimed to help make a decision. Furthermore, placing unclassifiable drivers into two class groupings may harm either the driver or the community by making the wrong decision.

4

Conclusion

This paper investigated a probabilistic decision-making classifier in the renewing policy of the driving licence. The probabilistic classifier considered in this paper permits

162

W.H. Awad and R. Herzallah

taking into account knowledge of uncertainty in assessing drivers’ knowledge. A case study of Jordanian drivers is used to construct a probabilistic classifier that discriminates between drivers passing and failing a SWDE. From the obtained results, it is clear that the classifier can be used to take a decision about whether a driver can automatically renew his driving licence or must undergo another training phase before he or she can renew his or her driving licence. However, the error bars of predicting drivers’ results in the exam were very large indicating that the constructed classifier, especially with the persistence of excitation problem in the available data, yields unreliable results. The renewal decision of the driving licence should not be determined by only considering the mean value of the classifier and extra assessment methods may be required for the final decision. Theoretical knowledge of the drivers can help in taking the renewal decision but it is not the only assessment methods that could contribute to the final decision. Other assessment methods such as the practical knowledge of the driver should also be included to augment the decision support system. However, the predictive uncertainty should also be attached to the results. In addition, those drivers with large error bars in their predicted outcomes should be considered unclassifiable rather than misclassified into groups. Trying to classify the unclassifiable drivers and claiming them to be a success of the method is unrepresentative. In this study, a driver’s result in a SWDE only is used to construct the classifier and help in taking the renewal decision of drivers’ licences. In future work, we will demonstrate how we can exploit the same machinery to augment the proposed classifier with extra assessment methods.

References Akgüngör, A.P. and Doğan, E. (2009) ‘An artificial intelligent approach to traffic accident estimation: Model development and application’, Transport, Vol. 24, No. 2, pp.135–142. Arthur Jr., W. and Doverspike, D. (2001) ‘Predicting motor vehicle crash involvement from a personality measure and a driving knowledge test’, Journal of Prevention & Intervention in the Community, Vol. 22, No. 1, pp.35–42. Awad, W., Al-Agtash, S. and Nsour, A. (2004b) ‘Adaptive neuro-fuzzy inference system for traffic cycle optimization’, International journal of modelling & simulation, Vol. 24, No. 2, pp.80–84. Awad, W.H. (2004) ‘Estimating traffic capacity for weaving segments using neural networks technique’, Applied Soft Computing, Vol. 4, No. 4, pp.395–404. Awad, W.H. and Al-kharabsheh, M.R.S. (2001) ‘Investigating drivers’ traffic knowledge in Jordan’, Driving Assessment 2001: The First International Driving Symposium on Human Factors in Driver Assessment, Training and Vehicle Design, Aspen, Colorado.

Awad, W.H. and Janson, B.N. (1998) ‘Prediction models for truck accidents at freeway ramps in Washington State using regression and artificial intelligence techniques’, Transportation Research Record: Journal of the Transportation Research Board, Vol. 1635, No. 1, pp.30–36. Awad, W.H., Shbeeb, L.E. and Suliman, M.R. (2004a) ‘Assessment of Licensed Drivers in Jordan’, Conference proceedings, Jordan International Conference on Sustainable Development of Transportation Systems, Amman, Jordan. Basu, D. and Maitra, B. (2006) ‘Modeling stream speed in heterogeneous traffic environment using ANN-lessons learnt’, Transport, Vol. 21, No. 4, pp.269–273. Blinova, T.O. (2007) ‘Analysis of possibility of using neural network to forecast passenger traffic flows in Russia’, Aviation, Vol. 11, No. 1, pp.28–34. Chapman, P., Underwood, G. and Roberts, K. (2002) ‘Visual search patterns in trained and untrained novice drivers’, Transportation Research Part F: Traffic Psychology and Behaviour, Vol. 5, No. 2, pp.157–167. Federal Highway Administration (FHWA) (2008) Highway Statistics, http://www.fhwa.dot.gov/policyinformation/ statistics/2008/dl1c.cfm Hanson, T.R. and Hildebrand, E.D. (2011) ‘Are age-based licensing restrictions a meaningful way to enhance rural older driver safety? The need for exposure considerations in policy development’, Traffic Injury Prevention, Vol. 12, No. 1, pp.24–30. Hedlund, J. (2007) ‘Novice teen driving: GDL and beyond’, Journal of Safety Research, Vol. 38, No. 2, p.259. Herzallah, R. (2007) ‘Adaptive critic methods for stochastic systems with input-dependent noise’, Automatica, Vol. 43, No. 8, pp.1355–1362. Herzallah, R. and Lowe, D. (2008) ‘A Bayesian perspective on stochastic neuro control’, IEEE Transactions on Neural Networks, Vol. 19, No. 5,pp.914–924. Ivan, J.N. and Sethi, V. (1998) ‘Data fusion of fixed detector and probe vehicle data for incident detection’, Computer-Aided Civil and Infrastructure Engineering, Vol. 13, No. 5, pp.329–33. Jordan Traffic Institute (2011) Jordan, website: http://www.jti. psd.gov.jo/ Jørgensen, F. and Pedersen, H. (2005) ‘Enforcement of speed limits – actual policy and drivers’ knowledge’, Accident Analysis & Prevention, Vol. 37, No. 1, pp.53–62. Kalyoncuoglu, S.F. and Tigdemir, M. (2004) ‘An alternative approach for modelling and simulation of traffic data: artificial neural networks’, Simulation Modelling Practice and Theory, Vol. 12, No. 5, pp.351–362. Martinez, K.L.H. and Porter, B.E. (2004) ‘The likelihood of becoming a pedestrian fatality and drivers’ knowledge of pedestrian rights and responsibilities in the Commonwealth of Virginia’, Transportation Research Part F, Vol. 7, No. 1, pp.43–58. McKenna, F. (2010) Education in Road Safety: Are We Getting it Right? (No. 10/113). Nagai, M., Onda, M. and Katagiri, T. (1997) ‘Application of genetic algorithm to analysis of driver's behaviour in collision avoidance’, International Journal of Vehicle Design, Vol. 18, No. 6, pp.626–638.

Driving licensing renewal policy using neural network-based probabilistic decision support system Preusser and Tison (2007) ‘GDL then and now’, Journal of Safety Research, Vol. 38, No. 2. Ross, M., Goodwin, R., Watkins, R., Wenzel, T. and Wang, M.Q. (1998) ‘Real-world emissions from conventional passenger cars’, Journal of the Air & Waste Management Association, Vol. 48, No. 6, pp.502–515. Sazi, M.Y. (2006) ‘Comparison of fuzzy logic and artificial neural networks approaches in vehicle delay modeling’, Transportation Research Part C: Emerging Technologies, Vol. 14, No. 5, pp.316–334. Siegrist, S. (1999) Driver Training, Testing and Licensing: Towards Theory-based Management of Young Drivers’ Injury Risk in Road Traffic: Results of EU-project GADGET, Work Package 3 (No. 40). Schweizerische Beratungsstelle für Unfallverhütung, bfu. Song, H. and Gao, Y. 2(013) ‘Traffic meteorological visibility estimation based on homogenous area extraction’, International Journal of Computer Applications in Technology, Vol. 48, No. 1, pp.36–48. Stamatiadis, N., Agent, K.R. and Ridgeway, M. (2003) ‘Driver license renewal for the elderly: A case study’, Journal of Applied Gerontology, Vol. 22, No. 1, pp.42–56. Stoke, C.B. (1979) The Effectiveness of Written Driver Knowledge Tests, No. VHTRC 79-R41.

163

Wallis, T.S. and Horswill, M.S. (2007) ‘Using fuzzy signal detection theory to determine why experienced and trained drivers respond faster than novices in a hazard perception test’, Accident Analysis & Prevention, Vol. 39, No. 6, pp.1177–1185. WHO (2009) Global Status Report on Road Safety 2009, http://www.who.int/violence_injury_prevention/road_safety_ status/2009/en/ Wolming, S. and Wiberg, M. (2004) ‘The Swedish driver licensure examination: exploration of a two-stage model’, Journal of Safety Research, Vol. 35, No. 5, pp.491–495. Zhao, L. (2012) ‘Application of interval type-2 fuzzy neural networks to predict short-term traffic flow’, International Journal of Computer Applications in Technology, Vol. 43, No. 1, pp.67–75.

Website Motor Vehicles by Country, United Nations World Statistics Pocketbook and Statistical Yearbook, http://www. nationmaster.com/graph/tra_mot_veh-transportation-motorvehicles