International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Kuniaki Sasaki, Akane Sawada

Last modified: 28 March 2017


The travel behavior analysts have focused on the activities of decision maker as the essence of demand, which is called activity-based model (ABM). ABM basically focuses on the activity on a day and the trips are treated as the derived demand of activities. Most of ABM considers the activities of one day, thus they consider the interaction and schedule of activities and trips. Those models generally use disaggregate choice models to analyze the behavioral characteristics such as destination, travel mode, and time of the day. As ABM models the individual behavior of activity, ABM is with strong affinity for the micro-simulation system. However, there are still some challenges at the validity of activity forecasting when applying the micro-simulation based on the ABM. One of the big challenges is the location choice of activities. There are a number of alternatives on the choice set of destination choice, so that the fit of the model would be worse. To avoid such condition, a lot of studies introduced some assumptions and certain constraints to bring the ABM to the real choice. One of such trials is the time-space constraints which take into account the limits of available time. Assigning the alternative specific constants to all alternatives is one of the ways to improve the fit of the destination choice model. However, the number of alternative specific constants would be equal to the number of zones in the metropolitan area, which is generally large number. Besides, the stability of the constants in the future is not promising, because the meaning of the constants is regarded as the expectation of unobserved factors.

On the other hand, ICT has developed rapidly and enables us to collect various kinds of logs of the state. These data are called "big data" and a lot of researches started to apply "big data" to the travel behavior analysis. It is necessary to develop the method to apply not only conventional data, but such "big data" to the travel demand analysis. Data assimilation is one of the methodologies to fuse such two types of analysis, which are the numerical simulation of a state of traffic and the observation of that state. That is the process by which observations of the actual state of a system are incorporated into the state of a numerical model of that system. ABM performs activities of an individual of a day, it can produce the various types of travel index as the state of traffic, so that various type of data can be used as the observation of the state of traffic. If we can improve the model performance of ABM using big data, the model analysis will be useful for decision making of transport policy.

In this study, we are going to simulate individual behavior using ABM and then assimilate the output with the current observation of the state of traffic. By this scheme, the model's forecasting power will be improved, even if the data to estimate ABM is collected once at decade such as person trip survey of Japan. Especially, we focused on the destination choice modeling because that is inevitable to improve ABM. As the observation of the state of traffic, we used the number of persons staying at a zone estimated by the cellular phone access to the base-station. We used "mobile space statistics data" offered by NTT-Docomo which occupies more than 60% of share of cellular phone in Japan. Here, we have integrated the whole system in the scheme of the state-space model as we defined the state as the number of persons on each zone. In the proposing system, the ABM-based simulator is the structure equation and the observation data of the state is the "Mobile space statistics".

We estimate the state by the system model first and estimated traffic state again using the observation, the latter process is so called filtering. In our empirical study, we adopted the typical ABM proposed by "Bowman and Ben-Akiva (2000)". We use Tokyo person trip survey in 2008 for model parameter estimation and then forecast the travel behavior using the LOS variable in 2015. We assume time-of day choice, destination choice, tour type and purposes choice in ABM, then applied this model system for forecasting the distribution of the residents of Yokohama city on each time of day staying at the center of Tokyo (14 zones). The process of forecast is applying repeated micro-simulation based on the estimated ABM and then filtered the forecasted state using "Mobile space statistics". The difference between the estimated distribution and the observed distribution after filtering was reduced to 6 % of the difference of non-filtered distribution. This is the matter of course because the target distribution of filtering is the observed distribution. However, the proposing model is not sequential decision model but simultaneous decision making model of activity of a day. Thus, filtering the distribution of all of the time of day yielded the adjustment bias. The adjustment bias means that the filtering on a time of day affects other time of day behavior because the output of model is the entire activity of a day.

This trial improved the forecasting power of choice mode based micro-simulation using distribution of the number of persons in zones from big data. However, there are many limitations in this study. One is the computational burden. Particle filtering needs more computational power. Another is the utilization of big data. We can utilize observation almost day by day, although it is hard for the models to consider day by day distribution.

But we are sure that this trial is one of the direction of utilizing big data into the disaggregate choice models.

Conference registration is required in order to view papers.