Makoto Kasai

Refinement of Pattern-Matching Method for Travel Time Prediction

This paper reconsiders how to define the similarity between historical and current day travel time data in pattern matching for travel time prediction. The core idea of pattern matching is designing a measurement of similarity, and the similarity function is intuitively presumed to have a negative exponential distribution. Here, a gamma distribution that includes this exponential distribution is introduced as an alternative. Complimentary mechanisms are also tried. The results of application to data from an urban expressway are summarized as follows: although the prediction accuracy of the refined method is only slightly better than pattern matching based on previous pattern matching, marginally better fitness is found.

Application of hierarchical Bayesian estimation to calibrating a car-following model with time-varying parameters

The problem of congestion caused by capacity bottleneck phenomena in access-controlled road sections should be addressed. A description of the relation between car-following behavior and vertical gradient is expected to contribute to the development of effective measures, including accurate parameter tuning of adaptive cruise control systems. This paper develops a methodology for revealing this relation. First, a model with time-varying parameters allows the characteristics of the car-following behavior to be expressed depending on the vertical gradient. Second, to account for the gradual change in vertical gradient in considering car-following behavior, a hierarchical Bayesian model is applied to the description of gradual change. Third, Markov chain Monte Carlo method is implemented as a technique for finding a solution. An example of estimation is presented to demonstrate the procedure. Conclusions suggest future directions for extending this study to devising measures for mitigating congestion on expressways.

Exchange interaction in the time headway model in critical traffic flow states

Mitigation of traffic congestion is an important challenge for intelligent transportation systems (ITSs). However, congestion in basic sections on interurban expressways is not effectively considered in ITS design because spontaneous generation of congestion has not yet been modeled. This paper points out the difficulties of traditional approaches and proposes a new traffic flow model where only the exchange interaction of time headway is considered, which is different from standard car-following models. Through this approach, traffic flow in critical states can be approximated as connected particles, as in a magnetic model. The characteristics of traffic flow are limited to two dominant parameters expressing the robustness of time headway. The influence of the model parameters on traffic flow is exemplified through numerical simulations. This paper concludes with a discussion about the applicability of similar statistical physical models and their potential as a platform for associating the geometric design of roads with traffic capacity.

An experimental study on learning effects on mode choice behavior in dynamic park and ride

This study examines the learning process of users in mode choice behavior through an experiment on iterative trips simulating Dynamic Park and Ride (DP&R) as a new Travel Demand Management (TDM) scheme. The learning process is reviewed by three different analyses: 1) a cross-sectional analysis, 2) a longitudinal analysis using pooled data on experiences in the most recent trial, and 3) a longitudinal analysis that takes into account the cumulative effects of all preceding trials. Findings are summarized as follows: travel times expected by participants have an effect on mode choice, and past experiences as well as the most recent experience also have an effect. The paper concludes with a discussion of the importance of considering learning effects for widespread adoption of a new conceptual TDM scheme.

Extracting characteristics of traffic flow in bottlenecks with exchange interactions in time headway

Sag sections (sections that change grade) have recently been recognized as possible capacity bottlenecks. This study examines factors in capacity bottlenecks. First, it is shown that the relation between longitudinal alignment and traffic flow is important, whereas differences in driver characteristics, number of lanes, and overtaking have secondary effects since the bottleneck position is independent of day, time slot, and number of lanes. Second, to clarify the effect of longitudinal alignment on traffic flow, the accumulation of car-following method on a driving simulator is applied. Third, because traditional approaches that model traffic flow by car-following behavioral models and traffic flow simulation do not necessarily succeed in predicting bottleneck position, a new idea is proposed to extract the influence of longitudinal alignment on traffic flow. Based on the existence of exchange interactions in headways, the characteristics of traffic flow are condensed to a pair of variables, called hyperparameters, by hierarchical Bayesian estimation. The estimated hyperparameters are used to estimate flow. Specifically, a proxy traffic flow is generated, and finally the applicability of the idea of model exchange interactions in time headway is discussed.

Use of Particle Filtering to Establish a Time-Varying Car-Following Model

Methods for predicting the flow of vehicular traffic have been studied in order to anticipate short-term changes in service level. However, if the target route is an expressway with capacity bottlenecks (e.g., sag sections), it can be difficult to predict when a breakdown in the traffic flow will occur. There is a need to model the traffic dynamics from the free-flowing state to a congested state. Although previous studies have treated the parameters of traffic-flow models as being static, it is likely that they are actually time varying. This variation may be either random (i.e., white noise), influenced by longitudinal alignment, or both. To assess the critical traffic-flow state at a sag section, we use traffic-flow data collected from a driving simulator, these being more homogenous than actual flow data. Each participant repeated the course five times and from the second to fifth run followed a lead car corresponding to the same participant’s previous run. We estimate time-varying parameters to assess the influence of longitudinal alignment. To counteract operational randomness, we calculate the average parameters of the five repeated car-following runs for each participant. To minimize the computational cost, we use particle-filtering methods rather than Markov-chain Monte Carlo methods. Finally, we suggest future improvements to flow-breakdown modeling.

Car-following model with multiple predicting and controlling modules based on assumptions of anticipation behavior

Anticipation of driver actions may be a key factor in car-following behavior. Although many car-following models have been proposed, most of them are regarded as instantaneous feedback controllers. This paper considers taking anticipatory actions into account, as in feed-forward controllers. This paper retrofits this onto a previously proposed model assuming that drivers adjust acceleration to maximize instantaneous utility, adding predicting and controlling modules associated with different target times. The weight of each predicting module is updated depending on the circumstances of the driver over the last few seconds. This paper verifies that the proposed model improves estimation accuracy in terms of the root mean squared error of headway distance by calibration with observed car-following data from probe cars and video images. The goal of this model is future application in exploration of bottleneck phenomena.