**1. Introduction**

Advanced traveler information systems (ATISs), which are an integral component of Intelligent Transportation Systems (ITSs), are designed to provide real-time information that enables drivers to choose rationally from among alternative routes. The effectiveness of ATIS is dependent on the drivers response to received information. Incorporating information into the modeling practice may enhance the accuracy of route choice models by adding realistic behavioral mechanisms and thus improve the effectiveness of ITSs. Accordingly, it is essential to capture the behavioral generalization of informed drivers in order to enhance ATIS design.

From a modeling perspective, traditional transportation research attempts to replicate driver route choice behavior assuming that individuals are capable of accurately perceiving route performance and attempt to maximize their expected utility. Mathematically, such assumptions make it cost-effective and technically simpler to model traveler behavior. Most attempts at route choice modeling are discrete choice models that are econometrically derived from random utility theory. Since the 1970s, transportation researchers have studied the decisions associated with route choice modeling. In the past forty years, innovations in discrete choice models have progressed in three stages, namely: Multinomial Logit Modeling [1], Nested Logit Modeling [2] and Mixed Logit Modeling (Ben-Akiva et al. unpublished manuscript, 1996). Each enhancement attempted to direct the logit model towards more flexible model structures. Despite previous achievements, the ability of these models to capture realistic route choice behavior has been increasingly challenged due to insights from the psychology field. Human choice behavior often produces *imperfect* dynamical systems that can be controlled using techniques described in [3]. Through a range of empirical research, drivers were detected to be not omniscient, as expected in traditional models, in precisely perceiving the actual route performance. Bounded rationality was initially introduced by Simon [4] to explicitly account for the fact that human beings are incapable of identifying the best route among multiple alternatives due to limitations in knowledge, cognition and information acquisition. Tawfik and Rakha [5] verified Simon's theory using a real world experiment, demonstrating that drivers generally only had a 50% accuracy in perceiving route information (e.g., travel time, travel distance, speed). Even though travelers occasionally have correct perception of route performance, they may not be willing to switch to the perceived better route; rather, they stick to the habitual choice until its performance is not satisfying. In other words, travelers are not necessarily utility maximizers [6]. Satisfying psychology triggers individuals' behavioral mechanisms in seeking a satisfactory solution instead of the optimal one. Irrational behaviors deviate travelers from the best route and are not easily predictable by traditional models due to limitations in model assumptions.

Route information provides an explicit description of the actual performance of the choice sets, which has the potential to improve travelers knowledge and direct them towards the objectively optimal decision. Accordingly, route information is expected to facilitate travelers to make more logical choices (choose faster routes in this study). The accuracy of traditional discrete choice models may probably be improved by integrating information effects into the modeling practice. An explicit generalization of the effect of information on route choice behavior is thus studied.

The proposed research attempts to provide valuable insights in addressing a number of important questions, namely: does route information enable drivers to behave more rationally? How does the information affect behavioral mechanisms from person to person as well as from trip to trip? What is the difference in behavioral effects between information types? This study is a follow-up experiment of [5]. The results of the two experiments are compared and provide significant implications to the behavioral effect of real-time information. The major contribution of this study is to design a real world route choice experiment and study realistic route choice behavior of informed drivers and their day-to-day behavioral variations. This study differs from most of the studies in the literature that have investigated driver route choice behavior in a hypothetical environment such as simulation and questionnaire that is unable to completely reflect reality. The paper also comprehensively presents the heterogeneity of drivers' responses to the provided route information, considering the diversities in driver's age, gender, and personal traits, trip characteristics, and temporal variation. The findings of this study are critical and insightful to the modeling of route choice behavior and personalized ATIS design.

## **2. Literature Review**

Empirical research found that the factors considered by travelers in making route choice decisions were not unitary [6]. Numerous attributes were found to be important considerations, including travel time, trip distance, average speed, and the number of traffic signals along the route. Nonetheless, previous attempts at identifying the attributions of route choice identify travel time as the most important factor even though travelers may also consider other factors. In accordance with Tawfik and Rakha's study, 70% of drivers' route choices was successfully explained by travel time followed by average speed and distance traveled [5,7]. Consequently, travel time information is provided to the test participants in this study.

As captured in the "hot stove" effect [8], individuals were not inclined to select options associated with high variability, although these might actually provide larger benefits. Considering uncertainty, people do not have perfect knowledge of the gains that could be accrued and the loss associated with risking changing habitual choices. Prospect Theory [9] explicitly and thoroughly describes this

psychological behavior that risk-seeking behavior would likely exhibit in the loss domain rather than in the gain domain. In relation to route choice, Katsikopoulos et al. [10] verified the results of Prospect Theory through a simulated experiment in which participants were provided with the information of travel time variability, indicating that risk aversion emerged in the gain domain (alternative route is faster but riskier) while risk seeking emerged in the loss domain (alternative route is slower but riskier). Accordingly, drivers repeatedly make illogical choices due to the risk aversion in the gain domain. Information is expected to reduce the uncertainty and enhance rational behavior partially by leading travelers to risk seeking in the gain domain. Katsikopoulos et al. [10,11] revealed that the provided information supported choice rationality and reduced inertia.

The effect of travel time information on route choice behavior has been incrementally studied both from a theoretical and practical standpoint. Early studies, such as Lida et al. [12] and Yang et al. [13], pioneered the investigation of the information effects on drivers route choice behavior, both of which conducted studies in the simulation environment, with the number of participants 40 and 20, respectively. Ben et al. [14] thoroughly investigated the combined effects of information and driving experience on route choice behavior using a simulated experiment in which a total of 49 participants were recruited. The results provided evidence to sugges<sup>t</sup> that the expected benefit of information is achieved only if drivers lacked long-term experience. Based on this study, a discrete choice model with Mixed Logit specifications was developed to accurately describe the respondents' learning process under the provision of real-time information [15]. Further, Ben et al. [15] also demonstrated that information provided on average travel time resulted in different responses compared to information on travel time variability, which remains to be verified. Using a simulation- and a stated preference-based approach, numerous attempts were made to econometrically address the various behavioral mechanisms of drivers' route choice with real-time information. The studied behavioral mechanisms involved logical choice [14,15], inertia choice [11,16], switching behavior [17–19], habit and learning [20,21], and others [22–27]. Specifically, Karthik et al. [16] designed an inertia behavior simulation study and demonstrated that user experiences decreased inertia behavior in day-to-day variation. The travel time information was demonstrated by many studies to effectively move route choice towards rationality ([14,15,17,19,25,26,28–32]), however, the effect of information strongly depends on other factors, such as personal traits, trip characteristics, and other decision considerations. From the personal trait perspective, Jou et al. [17] concluded that elderly travelers would be less likely to switch due to the habitual and risk-aversive effects, and male travelers would be more likely to switch to the best route. Also, trip characteristics and traveler preferences were proved by Polydoropoulou et al. [18] to significantly affect route switching and compliance with information. In summary, to the authors' best of knowledge, existing studies have typically lacked realism (either based on simulation or stated preferences approaches) and have not characterized the effect of information details of trip characteristics, such as directness of the route, number of intersections, conflicts with non-motorized traffic on driver route choice behavior. This study attempts to address this void.

Although previous attempts provided econometric and empirical generalizations, most were based on simulation and stated preference approaches. In the simulator surroundings, however, respondents make decisions in a digital and virtual environment. Stated preference is an investigative approach in which respondents are given questionnaires to make choices hypothetically. Both approaches are performed under fictitious conditions and may not accurately capture actual choice behavior. Consequently, an in-field case study is needed to address the driver route choice behavior. To the author's best of knowledge, this study, as a follow-up test of Tawfik and Rakha's experiment (in which information was not available), is the first attempt at addressing this need using dynamic travel time information, which differs from the previous real-world experiments (e.g., [33,34]) that conducted experiments for a short time period (e.g., several days) and did not capture the day-to-day variation of route choice behavior using the learning mechanism that accounts for information effects. For example, Papinski et al. [33] had 31 participants involved in a real-world

route choice study with GPS as the data collection tool. The study provided valuable insights into the use of GPS trajectory data for route choice analysis, ye<sup>t</sup> was only conducted for two days, thus not capturing day-to-day variations and dynamics of learning behavior.

Drivers' responses to information may differ based on personal characteristics, demographics, preferences and choice situations [35,36]. Nonetheless, few studies so far have attempted to quantitatively investigate such discrepancy. Tawfik et al. [21] developed a latent class choice model by classifying personal traits and choice situations into four behavioral groups as illustrated in Table 1. The results demonstrated that the model outperformed traditional hierarchical models in predicting realistic behavior. However, Tawfik et al.'s study did not incorporate the effect of information in the modeling practice. Accordingly, this study attempts to investigate the information effect considering different participants and choice situation characteristics in order to capture preliminary insights for modeling in the future horizon.

In general, given the incomplete picture of the behavioral aspects of route-choice decision making, more attempts are justified. The proposed research is thus initiated by a real world case study to provide a better understanding of underlying behavioral effects of travel time information on route choice decisions.

**Table 1.** Four identified behavioral driver types [21].

## **3. Experimental Design**

As aforementioned, Tawfik et al. identified four route choice patterns observed in a real world experiment. This experiment attempts to quantify the influence of route information on traveler route choice behavior by comparing the choice patterns between Tawfik et al.'s experiment and the experiment conducted in this study. Occasionally, drivers prefer a route they frequently choose instead of switching to the actually faster route; or may deviate from the habitual route to the alternative route, which is on average worse, only because the performance of the usually-taken route becomes bad on a random day. These irrational behaviors may probably be caused by a lack of precise information. The study attempts to address a number of questions: Will travel time information make drivers behave more rationally? Will the effect of information be different among individuals? What type of information will be most effective?

A total of 20 participants were recruited within two age groups including 18-33 and 55-75. These two age groups were selected because the authors wanted to investigate the impact of drivers' age on the information effectiveness in changing choice behavior. The big difference of drivers' age in the two age groups may more easily distinguish the difference of information effect attributed to drivers' age. Each of them was required to accomplish three sectors of the experiment: a pre-run questionnaire, on-road test and a post-run questionnaire. The pre-run questionnaire was conducted before the beginning of the on-road test, which gathered the participants' demographics, driving experiences, preferences, habits, information usage and the perception of route performance. Noticeably, each participant was demonstrated to have little knowledge of the route performance according to the results of pre-run questionnaire. The on-road test was conducted around the areas in Blacksburg and Christiansburg, VA for the morning, noon and evening peak from October 2013 to April 2014. The participants were asked to drive as if (When drivers were doing the test, they were asked to drive from one predefined origin to the destination during every trip, and they actually did not commute during the test. However, the researchers wanted to emulate the trip as a commute trip on which travel time may probably be the first consideration by the drivers. So "as if" here means that drivers were asked to behave like in a commute.) they were commuting in order to ensure that travel time was an important consideration when they were to make choices. Each participant was asked to drive 11 trials, 5 of which provided participants with strict information (average travel time) and 5 provided with range information (travel time variability). The last one trial was not provided with any information, aiming to see how well information impacted drivers. It should be noted that the information was provided one time with average travel time and one time with travel time variability in order to eliminate the bias on each of the information types. The average travel time information provided to each trial was estimated by averaging the experienced travel time of three previous trials (The experienced travel time was recorded by GPS during the testing; three previous trials were selected to be averaged because the trials before has little impact on the decision based on the literature [5]; information used for the first trial was obtained from the experiment in [5].) and travel time variability was estimated using the average value and standard deviation (*average travel time* ± 2 ∗ *standard deviation*), so that the information could be dynamically updated each day to enhance the reliability estimate. It is worth noting that the provided travel time information was collected using GPS during the real-world experiment rather than from on-road or in-vehicle sensors, but the experimental design methodology of this study is also applicable to sensor-based information. For each trial, there were five O-D trips, each of which had two alternative routes, one route was on average faster in travel time than the other. The characteristics of each route were specified in Table 2. The participants' task was to repeatedly make choices between the two alternatives on each trip. Statistically, 55 choice observations were collected for each participant, 100 observations by each trial and 220 on each trip. Upon the completion of 11 trials of the on-road test, the post-run questionnaire was thereafter conducted, whereby the participants were asked whether the provided information was beneficial. The accuracy of travel time perception would be compared between the two questionnaires in order to have a knowledge of whether the participants' perception was improved as a result of providing them with information.

The logical choice rate—the proportion of times in which the faster route is chosen as a function of time (trial number), participant and trip, respectively—was selected as the indicator of the positive role of information in facilitating rational behavior. The inertial choice rate—the proportion of participants remaining on their habitual but slower route—served to evaluate whether the information contributed to enhancing participant attitudes of risk seeking in the gain domain. The on-road data collected by Tawfik was applied to estimate the choice rates specified as "without information" group. Tawfik's experiment was conducted on the same trips in Blacksburg and Christiansburg in 2012, which was also a day-to-day commuting test in which participants were asked to repeatedly make choices between the two alternative routes on each trip. The difference between the two experiments was that the proposed study provided participants with travel time information. For more details of Tawfik's study, see [5].

