First, he introduced recursive route choice models and estimation algorithms. The recursive logit model has many advantages and applications, but there are also some challenges, such as IIA property, expensive calculation cost, and inapplicability to stochastic networks. He tried to solve these problems. In the nested recursive logit model, he introduced a scale parameter in random terms to model network correlation. On stochastic networks, he developed a method to estimate the RL model efficiently by decomposing stochastic networks.

Then, he explained the relation between recursive models and inverse reinforcement learning (IRL). These two are equivalent and share the same objective: learning from human behavior. But they are from different communities: transportation modeling and machine learning. He said the transportation modeling community cares about the interpretability of linear utilities, but prediction performance of nonlinear rewards is important for machine learning. He suggested that we can benefit from the machine learning community to develop new models/algorithms for route choice/ activity-based choice modeling. He showed some implications and directions from the machine learning community, including nonlinear utilities, no destination situation such as pedestrian behaviors, cooperative/competitive behaviors of travelers, etc.

Lastly, he reviewed various Markov Decision Process (MDP) frameworks and positioned the RL in the MDP framework. He introduced distributional MDP frameworks, which would be promising to capture the correlation between utilities and go beyond the RUM models.

Her research is about land ownership behavior models considering decision dynamics and interaction. The study aims to understand landowner's behavior well, extend the RL model, and broaden its applicability. First, she explained how to digitize old paper-based Cadastre data and convert them to trajectory data. Then, she introduced the dynamic land-owning pattern choice model based on a dynamic discrete choice model. The landowner's decision history is regarded as state choice paths. The value function is calculated by backward induction, and the model is estimated by MPEC to reduce calculation costs. The likelihood ratio is sufficient for the whole period. We can interpret the time discount rate and parameter values of the different periods. In the discussion, the way to model the actual situation of land acquiring patterns, estimation method, and calculation cost are discussed with Dr.Oyama and Dr.Mai.

She models residential choice behavior after the Great East Japan Earthquake using the RL model. On a macro-scale, victims moved to a larger city located closer to their hometown. On a micro-scale, she revealed that each person experienced dynamic residential choices influenced by their situation, such as life stages, the progress of reconstruction, and subsidy. She formulated these dynamic residential choices as discounted recursive logit model, where the residential state is defined by time, location, and family structure. Estimation results showed that distance between cities is the most important factor. She is now trying to estimate the time discount rate of different types of residents because it is insightful when analyzing decision-making after disasters. In discussion, the meaning of using the RL model, the importance of capturing victims' heterogeneity, and an estimation method are discussed with Dr. Mai, Dr. Hara, Dr. Oyama, and Prof. Hato.

As an introduction of the international research workshop, Prof. Hato started with the example of the gridlock phenomenon which occurred just after The Great East Japan Earthquake, in Tokyo metropolitan area. In the collaborative research with Mr. Oyama, they conducted an analysis on the drivers' behavior in damaged networks. In such a context, drivers have little spatial information and tend to choose routes shortsightedly. In order to model these decision-making dynamics, they proposed a discounted recursive logit model. Also, he introduced on-demand routing optimization method using network contraction, which will contribute to the new transportation system in the depopulated areas. He concluded that route choice model with sequential discount rate is beneficial for understanding the decision dynamics in unsteady networks.

He analyzes the tsunami evacuation destination choice process, using the Spatial Correlated Logit model. He conducted a case study of the Great East Japan Earthquake based on the data collected from Kesennuma city, in Miyagi prefecture. In order to describe an understudied aspect of the evacuation process, he improved SCL model so that it can account more accurately for spatial correlation among alternatives. The results suggest the existence of a high degree of correlation among unobserved attributes of zones, and empirical findings suggest that factors such as OD distance by mode, OD altitude difference, number of buildings and number of officially designated shelters are statistically associated with evacuation destination choice. Also, and from a theoretical perspective, it addresses the issue of spatial correlation in discrete choice models. A spatially correlated logit model is estimated, where the allocation parameter is specified as a function of proximity and inter-zone altitude difference to capture more adequately unobserved similarities among alternatives in the specific context of tsunami evacuation, and the proposed model yielded better statistical fit than the traditional MNL.

Focusing on the interaction generated at the time of disaster situation, he clarified how much the interaction influences the decision to evacuate. He improved network formation model in order to describe a network formation in a small village under the disaster risk, and he formulated interaction in a small community by introducing Cross Nested Logit model (CNL). After that, he evaluated the influence on evacuation timings using local interaction model, while considering the positive & negative interaction between a robust person and a weak (disabled or elderly) person. In the formulation of an evacuation model, the utility function contains social utility which is defined by a human network and weight of interaction, and here, the weight of interaction can catch the asymmetric (altruistic) interaction between a robust person and weak person. He also carried out a case study in the mudslide case in Niihama city, occurred in 2004. The estimation result suggests that evacuation alert for the weak from rescue and number of the direct interaction encourage their evacuation and the robust people hesitate to evacuate before the weak evacuate. Through numerical simulation, he had two suggestive conclusions. First, staying trends will be amplified by gathering people who have low evacuation probability. Second, evacuation of a weak person has a positive impact on others' evacuation because the robust are able to start their evacuation.

Tomer began his presentation with the introduction of the precedence research on evacuation behavior and explained about the Haifa fire event which occurred on November, in 2016. He conducted a web-based survey to collect information about the evacuation. Based on the information, he estimated binary evacuation choice model. The result shows that people who live with children under 12 are more likely to choose evacuation. In addition to that, the strong social network effect is observed because there was strong tendency to travel together to the shelter.

Ido talked about his preliminary research on activity patterns and intra-household interactions in evacuation behavior, he aims to implement activity-based models for disaster, while focusing on intra-household interactions during the evacuation. In Haifa fire (2006) case, he carried out some preliminary analysis and there was a big difference in behavior between adults with children and adults without children because adults with children tend to make more stops to pick up their children, therefore, it is necessary to consider intra-household interactions.

In the presentation, she developed a frequency based transit assignment model considering that online information of predicted arrival times is available to passengers. Some public transportations have introduced technologies which informs the passenger of predicted arrival times, and transit assignment plays an essential role in planning and management of transit networks. The methodology is considering two types of available online information: (1) full, where the arrival times are available for all intermediate stops in the candidate paths, and (2) partial, where the arrival times are available at the boarding stop only. Passengers are assumed to consider the estimated arrival times together with the expected travel time when choosing their path. The assignment procedure includes the finding of attractive paths, setting of route choice decision rules for different cases of predicted arrival times, and the probability calculation for these different cases. The developed model is illustrated by an application for the Winnipeg network. In comparison to the well-known optimal strategies method, the suggested model produced significantly different assignment results and a notable reduction in the total travel time. The results demonstrate the potential impact of online information on travel behavior and route choice and emphasize the need for its consideration in planning models.

Mr. Parady introduced his research “THE EFFECT OF SEAWALLS ON TSUNAMI EVACUATION DEPARTURE TIME: A CASE STUDY OF THE 2011 GREAT EAST JAPAN EARTHQUAKE”. Using data from a survey of survivors of the 2011 Great East Japan Earthquake and Tsunami, he evaluated quantitatively the effect of seawalls on evacuation departure time. A mixed effect Cox proportional hazard model was estimated, using time-dependent covariates to account for the changing nature of tsunami risk. Findings suggest that on average, the presence of an effective seawall (defined as a seawall higher than the forecast tsunami height at any given time) delays evacuation by 29.8%. Furthermore, irrespective of forecast tsunami height, higher seawalls were associated with 11.6% and 15.2% delays in departures for walls of 4 to 8 meters and over 8 meters, respectively. These results provide quantitative evidence of the existence of a false sense of security deriving from the presence of seawalls, a hypothesis that to the best of the authors’ knowledge has only been discussed qualitatively. Given that timely evacuation is one of the main factors affecting survival probability, the effect of seawalls on evacuation delay or non-evacuation is an important policy consideration.

Mr. Urata introduced an optimization approach for the dynamic strategy of evacuation and picking-up behavior to respond to tsunami risk. He started by suggesting that preparation and “pick-up” behavior is an important factor of evacuation delay in the 2011 Great East Japan Earthquake. His research purpose is to propose an optimal dynamic strategy for reducing the evacuees’ risk with a pick-up behavior. Assuming that a liner city and simple network, he estimated the risk evaluation, He showed dynamic control for picking-up behavior on bottleneck intersection. Because of some people who take a risk and pick up their family members, congestion can occur. A deductive and sequential approach to reach an optimal dynamic strategy with a low computational cost was proposed. He indicated that using simple control approach which shows just a stop timing, evacuees can understand easily.

As a recent research, Mr. Yaginuma proposed a road evaluation method in the case of a disaster. He began with introducing that earthquakes can occur in various places in Japan, participants from Israel were very surprised that disasters were predicted over a wide area. Cost by Benefit evaluation norm is usually used in Japan, but this norm considers only normal everyday life and the emergency road vulnerability is ignored. He insisted that the evaluation of connectability is important and developed two new models. He calculated the log sum variable for the OD pair, calculated travel time, and evaluated road network. First of all, as a simple case, he showed an example result assuming that the vulnerability of the road is distributed. Next, taking into consideration the shutdown risk of roads and bridges endangered by Tsunami or other disasters, he applied it to actual highway network and showed the result.

Mr. Yamano talked about Probe Person (PP) technique using GPS. He showed an analysis about the walking speed based on PP technique. And then, he introduced new approach of travel behavior analysis based on information theory using sensor data and PP data simultaneously. He showed the experiment of OD demand estimation and suggests the difference between including error and error-free in measurements. His model will be able to take into account various error in measurements, sampling, and estimation.

His presentation theme was “Network Design Problem considering system time minimization and road safety maximization”. For network design, various kinds of problems need to be solved. We decide the objective function on upper level considering and the user equilibrium on a lower level. Also, we have to estimate the system time of different project combinations. He presented a solution for the multi-objective NDP. The objective criteria were minimizing the system time and maximizing the safety in the network. The model was applied for a large set of candidate projects on a real-size network, which demonstrates its ability to provide solutions for real problems. Since the crash prediction model is location-specific, implementation of the model on different networks will necessitate the use of alternative CPM. In order to find the Pareto front, the multi-objective optimization NSGA-II was used. No major differences were found using different population sizes.

Mr. Toledo proposed real-time control for transit systems with transfers. These days, because integrated public transportation systems improve delay, new real-time operations control is required. The road was assumed to be under single line control or multiple lines control. In each case, he showed the result of the case study. He developed a system for real-time transit control based on a prediction. In multiple line case, he set optimization problem to investigate the total delay in the system. He assumed the speed and holding time constraints. Considering holding and speed change, this approach accounts for transfers and vehicle capacities. Moreover, the transit can be evaluated as on-going against no and headway control. As next steps, he was trying to improve the predictions and the robustness.

He introduced his research about methods for demand transportation and public transport (bus) operation plan. In order to enumerate solution patterns to the demand plan of the demand transport, he proposed a method using ZDD. First, he expressed the operation route of demand traffic by a graph structure and enumerated all subgraph structures satisfying certain constraints from the given graph structure. Next, using compressed ZDD structure, it is possible to get not only the best solution but also other good solution. In the traditional method, only the best solution was obtained. However, in the actual planning, the best solution may be unable to run. Therefore, he suggested that this approach would be useful for traffic planning under such circumstances. The usefulness was confirmed in a case study. In the bus operation plan, we have to consider fairness and economy of the plan together. He proposed an evaluation model that quantifies the potential public transportation demand by sub-district within a municipality based on DEA and clarifies concrete benchmarks in the district. In addition, he applied the developed model to the actual city and examined the district to focus on promoting the use of public transportation.

Ms. Hara introduced "Easy Ride" which is the autonomous drive vehicles service developed by Nissan & DeNA. This autonomous drive vehicles are set very accurate digital map and detect signals and obstacles. It enables users to move to where they want to go with the application. Before the field test at Minatomirai area in Yokohama city where there are many kinds of people (residents, tourist, and worker), she checked the market volume in this area using Person Trip data and the SP survey they conducted. She pointed out that this result is static and what they want to know is the demand changing dynamically. Then, how to manage and how to calculate the demand dynamically are very big issues. And social understanding is also a big issue.

Professor Tomer introduced micro-simulation for developing Autonomous Vehicles (AVs) to evaluate expected effects on traffic flow, safety, and externalities. In the simulation, driving models need good algorithms for autonomous driving that is longitudinal-lateral movement. Then, there are some challenges that are how to set variants, how to treat sensing and communications, and how to idealize human driving in semi-AVs. They conducted the experiment in which they checked Adaptive Cruise Control(ACC) state and vehicles movement. Applying these data to Mixed logit choice model, they analyzed the relation between ACC transitions and some parameters. After that, they run a simulation to check the harmonic speed changing ACC and C-ACC penetration rates, demand level, and automated minimum leader headway settings.

Professor Bekhor posed questions that are "How will autonomous vehicles and traditional ones coexist?" and "How will the introduction of autonomous vehicles affect travel habits in general and the mode choice in particular?" These questions include whether AVs will induce people to travel more and whether they will benefit. Each intermediate scenario which is the result of introducing AVs can be seen as a parametrized combination of User Equilibrium (UE) and System Optimum (SO). He proposed the methodology using the dynamic representation of the problem. The initial calculation based on it using MSA (multi-scenario approach) showed the different tendency of moving from UE to SO in each example network.

Mr. Hayakawa showed the traffic situations with AVs in which there are complicated interactions between the administrator, service operators, and customers. Then, he focussed on the sequential mechanism for a service operator that aims to maximize social welfare. There, the objective is to maximize social welfare which is defined as the summation of the users’ utility based on Activity model under time-space constraints of users and under capacity constraints of the road network. To solve the problem, he used ZDD (Zero-suppressed binary decision diagram) and MSA. In numerical study, he showed that his proposed mechanism can remain high efficiency when increasing the number of agents. Finally, he suggested that his mechanism can be easily extended to an auction-based implementation by which the social optimal states are achieved.

Travel behavior is changed under uncertainty. Especially, we focus on extreme weather events and research the route choice behavior during them in β-Scaled recursive logit model framework.Checking the data in Tokyo, we realize that travel time parameter increases indicating difficulty to assess difference of travel time in links during heavy rainy day, that sequential discount rate β decreases indicating tendency towards the myopic decision making behaviour, and so on.

We establish fundamental algorithm to realize MaaS(Mobility as a Service) which allocate traffic resources optimally to users, considering heterogeneous users and slow traffic using auction-based approach. We have some problems to express this, but we introduce activity-state network which is an extended time-space network and time-constraint, and formalize activity-chain auction based on VCG mechanism. Moreover, We propose an algorithm, which is based on graph theory, to mitigate the computational complexity. Simulating based on this model, it can be said that by considering slow traffic, especially for cruising agents, the total social welfare can be increased.

In Dynamic Traffic Assignment (DTA), we have to deal with Dynamic User Equilibrium(DUE), which is the assumption that the travel time of travels on all used paths is equal at user equilibrium in order to evaluate path cost. Firstly, variational inequalities formulation is adopted for DUE. However, this approach imposes high computation burden. Therefore, surrogate modeling is adopted to develop a DTA solver considering macroscopic traffic dynamics. It describes the mapping from path flow to path cost and can be viewed as a local approximation of mapping for the solution point. It performs better when the distance between sampled data and the solution are small. By iteratively solving the DUE problem, the accuracy of surrogate model will be improved accordingly.

Transportation demand management (TDM) has problems that we cannot solve easily; congestion. Dedicated bus lane is regarded as one of the most promising solutions. In this study, road space allocation of the dedicated bus lanes is optimized quantitatively by using 3D-MFD (Macroscopic fundamental diagram). This approach adopts both static optimization model and simulation analysis model. In the Geneva case, around 15% of the space to the bus lanes is the most efficient. In the future study, it is useful to investigate the universality of this approach further with the Tokyo case, by improving the aggregate logit model or calibrating by the detectors data.

In this session, Prof. Michel introduced various ongoing research projects in EPFL such as the analysis of traffic system of Ghana, the analysis of behavior of pedestrian in EU, the technology for detecting pedestrian in autonomous driving, the optimization of a timetable of Swiss Federal Railways and so on. He finished it with an big project which aims at "World without Car" by installing the moving walkway throughout the city.

At last, he gave an advice to young researchers that they pay attention to both updating technology and idea and problems in the world and match them.

Mr. Oyama had a presentation about the route choice model estimated simultaneously with the route measurement model by GPS data. This approach is to resolve the 'loop' that the route choice data for estimating parameters of the route choice model is made by using parameters of route choice model and the assumption that error variance is constant among all links. Since estimating the proposed model corresponds to solving a fixed point problem, structural estimation method is applied. The effectiveness of proposed model and the estimation method is validated by twins experiments. When estimating parameters by using the real data, the dependency on each link of measurement error variance was shown. Also, the improvement of the estimation result was shown. As an example, travel time seemed to be significant from the result of one-way model while the arcade dummy variable was most likely to be significant and travel time was not significant from the result of structural estimation.

Mr. Seo had proposed a new survey method for collecting travel data which is suitable for a long-term and extensive investigation. The proposed method estimates the type of the detectated activity, and questions to the traveler when the estimation confidence is lower than a particular rate.

The most likely type of the activity and the estimation confindence are calculated from the historical data based on Naive Bayes assumption.

The answer by the traveler or the estimated type of the activity is used as the traveler-specific historical data to the learning process of the model for estimation of the activity type. The proposed method is validated by using existing travel survey data. The error rate of the estimation stayed 5%, which is given by admin, and the frecuency of questions was almost halved after 12 days. The proposed method would be the survey method which reduces the burdan of travelers and keeps the quality of data high.

When we model the evacuation behavior, it would be important to acount for heterogeneity in expected utility and choices of the evacuation start time. Mr. Urata had formulated the sequential choice model in which a decision maker chooses to evacuate or to stay there at every term. The utility of staying at time t includes the expected utility when choosing to stay at time t (EV). In the evacuation situation the recognized expected utility would not be the same as the expected utility satisfying the Bellman equation, and may have some variance. Therefore, the new constraint Ω, the range of vairiance, is introduced. This model is validated by evacuation data in Rikuzentakata. The algorithm to estimate the model is constructed based on SQP method. The finel likelihood of the proposed model is higher than that of static model or dynamic model with no heterogeneity. Also, distribution of recognized expected utility in time and space was shown.

Mr. Oyama had a presentation about the route choice model estimated simultaneously with the route measurement model by GPS data. This approach is to resolve the 'loop' that the route choice data for estimating parameters of the route choice model is made by using parameters of route choice model and the assumption that error variance is constant among all links. Since estimating the proposed model corresponds to solving a fixed point problem, structural estimation method is applied. The effectiveness of proposed model and the estimation method is validated by twins experiments. When estimating parameters by using the real data, the dependency on each link of measurement error variance was shown. Also, the improvement of the estimation result was shown. As an example, travel time seemed to be significant from the result of one-way model while the arcade dummy variable was most likely to be significant and travel time was not significant from the result of structural estimation.

Mr. Fukuda had a presentation about firms Erelocation behavior based on trading and financing relationship. Historically, transportation costs have been decreasing and population have been concentrating to a few larger cities. However, now firm’s relocation is mainly moving locally to not only larger cities. To clarify the mechanism of relocations, he focused on continuous relationship of firms including trading and financing transactions.

Using the database of Teikoku Databank, which is the largest credit research company in Japan, he analyzed firms Erelocation behavior, trading and financing network at first. Then, he applied nested logit model to explain location (municipality) selection and estimated parameters. As a result, the proposed model was more effective to relocation to medium-sized cities and of small & medium enterprises.

Mr. Varghese focused on activity opportunities and changing travel patterns, and presented the case study in India as a Developing Nation. The first part is about public transport crowding. In this part, he explained about the crowding of Ahmedabad city’s bus systems, and how to measure and calculate it. And in the second part, he focused on activity travel pattern in Mumbai. As the ICT technology has progressed, activity travel behavior of people who use mobile internet information can be changed. He proposed the framework to combine socioeconomic and spatial factors with ICT use in behavior modelings. And as the future work, he aims to develop an inter-regional destination choice model. It might be significant to better understand effects of policies related to long distance travel, such as infrastructure provisioning and facilities allocation, to activity opportunities.

In pedestrian networks, activities can be generated context-dependently such as spatial attributes, social and activity interactions. To capture sequential decision making of pedestrians, Mr. Oyama modeled dynamic scheduling behavior based on the concept as energy, which is a personal resource for engaging activities and used as discrimination function of activity generation model. Using GPS data collected from Probe Person (PP) surveys in Matsuyama city, he stochastically inferred sequences of activities in pedestrian network considering measurement errors. In his model, walking links is assumed as an activity as well as staying at nodes, and these activities were connected with certain spaces through the bayesian approach. As a estimation result, it was clarified that the energy consumption and gain processes depend on some behavioral and spatial context variables.

Prof. Jan-Dirk Schmöcker provided a keynote lecture regarding smart transit with real time information considering bus bunching.

Firstly he revealed why bus bunching arises when passengers have real time information (RTI). Because boarding rate is influenced by travelers given RTI, even without exogenous delays bus bunching arises.

Secondly, he transformed classic route choice problem with hyperpaths formulated by Spiess and Florian into multiple local demon game assuming risk-averse users and demons aiming to maximise delay. And then based on it, a smart transit with RTI in bus services can be transformed into single local demon game. Not only with formulations but also with actual effects, it was shown that the travelers behave in different ways at each departure point depending on how much well informed in advance they are.

In conclusion, in smart transit which gives “good Einformation to travelers, they can benefit from it. However, there is also a risk that the system can be outsmarted by the travelers. Therefore, even more information and management of network may be required.

Dr. Keyvan-Ekbatani focused on gating control, which is a practical tool for mitigating congestion in urban networks, and evaluated its efficiency using network or macroscopic fundamental diagram (NFD or MFD). When gating is located upstream of protected network (PN) to keep the network under-saturated, time-delay to approach PN is induced.

To deal with time-delay, Dr. Keyvan-Ekbatani proposed the developed feedback regulators, and tested gating performances using this concept. In experiment in urban network of Chania (Greece) and San-Francisco, the results showed that gating strategy acted efficiently and leads to significant improvements in mean speed and delay at the overall network level.

Ms.Kasahara’s research focused on the Demand control management for One –way Car sharing service named “Choi-mobi Eby applying price optimization. By using Probe Person data that is activity diary and trajectory data by GPS, and use result data, she formulated the profit maximization problem by changing the price of this service. In this formulation dynamic programming method was applied. As a result of the numerical example, it was confirmed that price control have effect for the demand control. Furthermore micro simulation was calculated considering the availability of service in the real network in Yokohama, Japan and evaluate the pricing strategy.

Professor Morikawa’s lecture focused on the bounded rationality in travel behavior modeling and decision making strategies. First, he gave an overview of the three types of rationality and 6 type of decision-making strategies. Then, he showed the framework for the bounded rationality with latent classes. Furthermore, two case studies were introduced. First one was about the mode choice between car and dynamic P&R using a semi-ordered lexicographic rule including both compensatory and non-compensatory decision making rules. Second one, more psychological approach, was by applying mental accounting theory, he aimed to analyze the effect of Ride Point Program in the context of discrete choice travel behavior.

Dr. Chikaraishi developed a model of shopping behaviour in a neighbourhood with social interactions. It shows retail attractiveness as endogenous. The important point of this model is to deal with market/non-market interactions not through “average Ebehaviour but through “aggregate Ebehaviour. This model is estimated by NFXP algorithm. By using this model, an empirical analysis was performed using 10 networks in Hiroshima city. It confirmed significant endogenous effect but it did not give multiple equilibrium.

Main point in the discussion part is whether new elements can be applied to that model or not, such as history/strategy of supply side (i.e. supermarket), interactions on floor space and average/personal income information. Since the available data is limited and developing the estimation method is difficult, it is challenging but possible to fulfil some of the requirements.

Dr. Fukuyama proposed a data-oriented approach for choice set generation of pedestrian route choice behavior. In this research, she applied Branch and Bound (Prato and Bekhor, 2006) and sampling method using data distribution from the probe person surveys. Branching rules in her study consist of destination directivity, keep-direction rate and detour rate. She concluded this approach is able to be applied for individual choice set generation which reflects the characteristics of individual behavior when they have enough amount of trajectory data.

Personal Mobility Vehicles (PMVs) have high potential as an intermediate mode between walking and cars. However in most cases PMVs are of limited use, such as in the airports, shopping malls, golf courses etc. This preliminary study analyzed conflicts between pedestrian and PMVs, In particular, how did PMVs users avoid pedestrians.

In this experiment, a Robstep M1 vehicle was used (Max speed is 15km/h, Min Speed is 9km/h). Position data was gathered through video monitoring, while an android app was used to measure wheel speed. 30 monitors conducted 24 tries each. It was hypothesized that the PMV will travel along the shortest path (timewise) and adapt his behavior to avoid the coming pedestrian in two ways: by changing its direction angle, or adjusting speed. Preliminary findings suggest that all else equal, rear avoidance is preferred to front avoidance.

The analysis of travel behavior under disaster situations consists mainly of three main stages 1) evacuation participation and departure time choice 2) Destination choice and 3) route choice. Although there is enough evidence to suggest that these phases are interdependent, due to model complexities most studies model these phases independently. This presentation focused on this gap in the literature by proposing a discrete-continuous model of Tsunami evacuation destination choice and evacuation timing. Data from the Great East Japan Earthquake is used to estimate the model. Travel behavior data from the day of the earthquake was gathered by the Ministry of Land, Infrastructure, Transport and Tourism (MLIT) of Japan via a travel diary. In terms of the methodology used, based on the premise that both outcomes might in fact be jointly determined, a discrete-continuous model is formulated. Destination choice is formulated as a multinomial logit model, where the choice set is generated as a function of location altitude in order to account for tsunami risk. Evacuation timing is estimated as an accelerated failure time survival model. Results validate the use of a discrete continuous approach as several error correlation parameters were statistically significant. Among factors found to affect destination choice were OD distance, OD altitude difference, which is estimated as a quadratic function, average slope of destination and total number of buildings in destination. Regarding evacuation timing, significant variables include the use of car for evacuation, altitude at origin knowledge of refuge locations, and age.

Mr. Seo’s presentation is aim to acquire volume-related info over a wide range, and proposes a new method for estimating traffic states using probe vehicles with spacing measurement equipment. Then validates the method under an actual condition in Tokyo. Mr. Seo reports that high resolution information can be acquired where enough number of probe vehicle exist. And mentions a system model of microscopic travel behaviors and pedestrian space as future works.

In recent years, the concept of urban development is changing into which promote the population concentration to cities center, like a compact city, from a viewpoint of the breakaway from the society of automatic automobile dependency. However, it is not clarified about the causal relationship which built environment gives to traval behavior of individuals and households. This is because the right causal effect cannot be explained in cross-sectional data since selection bias occurs by the background covariates of a specimen.

To overcome these limitations, Dr. Troncoso had an eye to the "propensity score" approach (Rosenbaum and Rubin, 1984). This score is the conditional probability of treatment given observed covariates and can explain the causal effect which eliminated the selection bias. This approach is originally used under binary treatment regimes, but can be generalized under continuous treatment regimes (Imai and Van Dyk, 2004).

This study focus on making interactions under a disaster in a settlement. An interaction between two people is influenced not only by two but also by others in group. The process of making interactions are modeled as a group decision-making. This model capture a spatial correlation of interactions. The correlations arise from spatial risks and travel costs. The correlations are formulated by spatial divisions. The model structure is a network-GEV model (Daly & Bierlaire(2006)). In this study, a case study is shown. The utility of making interactions are set by a altruistic utility which is a difference of risk.

Dr. Iryo's presentation discussed the issue of traffic evacuation directions using the Great East Japan Earthquake as a case study. The presentation challenged the assumption of a unique destination used in Dynamic Traffic Assignment (DTA) for evacuation analysis. Dr. Iryo argued that in the case of disasters trip purposes can in fact be heterogeneous, because the only purpose is not evacuation. This paper was tested using evacuation data from some cities in Tohoku during 3.11. Findings in fact suggested that the direction of evacuation trips was quite clear (homogeneous), but also showed the existence of other trips such as picking up family members, that are rather heterogeneous.

Mr. Urata’s presentation focused on understanding collective behavior during disaster situations. And explain how to use a dynamic discrete choice modelling framework to understand the process of collective behavior formation. Mr. Urata further argues that collective behavior predictions can be helpful in developing an efficient local evacuation management plan and an adequate information propagation system.