Technique for Explaining Causes of Congestion Using Pedestrian Simulation

2. Pedestrian simulation

Past attempts at reducing congestion at facilities where people congregate have involved both direct and indirect measures for controlling pedestrian flow. Direct control measures have included restricting entry numbers, upgrading the capacity of facilities that provide services, and making changes to the site layout. An example of an indirect control measure, on the other hand, is the provision of signage to indicate the level of congestion. Another, used at places such as restaurants, is to issue coupons redeemable at off-peak times. While indirect control has attracted interest in recent times for its advantages of being inexpensive to implement and unlikely to detract from user satisfaction, a problem with the approaches is that their effects can be difficult to estimate.

Having recognized the importance of indirect measures for congestion mitigation, such as providing information about the extent of congestion or offering coupons and other incentives, Fujitsu Laboratories went on to develop a pedestrian simulation that is able to evaluate both direct and indirect control measures. Specifically, this involved developing agent models that replicate the visitor choice behaviors that determine how visitors to a site move about, for example, destination choice based on their preferences as well as on how far away and how crowded the location is, and information perception and information processing of how the choice changes in response to their acquisition of information at the site. By using artificial agents implemented with these models, the simulation is also able to evaluate indirect control measures. Fujitsu Laboratories and Fujitsu have been jointly developing techniques for enabling the precise simulation of people’s movements, with 25 patents currently pending in Japan and overseas, including for the agent models that make it possible to evaluate indirect control measures.

4. Evaluation experiment

The new micro-dynamics analysis technique was applied to a pedestrian simulation at an airport to verify its effectiveness. In the evaluation experiment, firstly, the micro causes of congestion were identified by applying the technique to a simulation log. Next, after these causes had been eliminated, the level of congestion was evaluated.

Table 1 lists the results of the analysis. The aim of the analysis was to reveal the causes of congestion that occurred near a restaurant after it opened (from 9 a.m. to 10 a.m.). The degree of similarity between the set of agents involved in the congestion and the set of agents having each combination of characteristics was evaluated using the F-value. The higher the F-value for the characteristics, the better the characteristics represent the set of agents involved in the congestion. The parameters were grouped into attributes, actions, and cognition, then the agents were clustered according to the grouped parameters. The table includes columns: “Arrival and departure time” showing the time agents arrive at and depart from the airport, “Imperative purpose” meaning activities agents must do, “Purpose list” meaning activities agents want to do but are not absolutely necessary, “Places visited” showing places agents visited in time sequence, “Areas passed through” showing areas agents passed through, and “Information obtained” showing signage or other sources agents got information from. “Precision” indicates the proportion of agents involved in the congestion contained in the row, among all agents included in the row (namely, agents having a set of characteristics corresponding to the row). “Recall” indicates the proportion of agents involved in the congestion contained in the row, among all agents involved in the congestion. “F-value” is a statistic that combines Precision and Recall.

Table 1 Characteristics of agents involved in congestion.

The following explanation uses as an example the most important micro cause (i.e. the row with the highest F-value). Characteristic 1 denotes agents for whom a meal is an imperative purpose, for whom the visit sequence goes from entrance to restaurant to exit, and who viewed the information board closest to the restaurant (where the restaurant location was given). In other words, this indicated that the people involved in the congestion are those who have a purpose of eating a meal, who go directly from the entrance to the restaurant, who view the information board closest to the restaurant and enter the restaurant following it. These results can be used to suggest, for example, that showing a location other than the restaurant on the information board closest to the restaurant could be a good way to reducing congestion. This is because changing the information board should prompt some of the people who would otherwise have gone straight to the restaurant to instead go somewhere else, so it can be anticipated that this will reduce congestion by dispersing the people more widely.

Figure 3 shows the congestion mitigation that results when the agents identified as the most important micro cause are excluded. The horizontal axis is the time of day and the vertical axis is the number of agents waiting at the restaurant. The red band indicates the time period covered by the analysis. The blue baseline represents the number of agents waiting at the restaurant when nothing is done to reduce congestion while the grey line represents the number after the agents identified by this technique as the micro cause have been excluded. In this experiment, 72 agents were excluded out of a total of 2,000. The results indicate that congestion can be reduced by excluding the agents identified as the most important micro cause.

Figure 3 Congestion mitigation by excluding agents identified as a micro cause.