This paper presents a data-driven agent-based simulation of individual mobility based on spatio-temporal data from mobile phones. The model developed is embedded within the CityScope framework, a platform used as decision support system for city development. This work analyzes the Andorra visitors’ flow and traffic congestion through an agent-based visualization using different representation, abstraction, and interaction features.

Introduction

Telecom data coupled with other data sources—such as social media—can help us understand human behavior at spacial, temporal, and social level. These un- precedented rich sources of data allow us to study how people move and, thus, how our society behaves. Previous research [5] show that such insights can be used to design interventions to improve our daily lives and even visitors’ expe- rience in the scope of tourism strategies, which is the case study for Andorra. However, the results of the analyses are not always comprehensible for non- experts. CityScope is a visualization framework, developed by the City Science Initiative at MIT Media Lab, that serves both as an urban data observatory and urban exploratory decision support system for city development. Cityscope is a next generation, tangible, augmented reality platform that helps to (1) visualize and understand the meaning of complex urban data and inter-relationships, (2) simulate the impact of multiple interventions and (3) support decision making in a dynamic, iterative and evidence-based process. CityScope helps non-experts to engage into conversation through visualizations that synthesize the analyses in a coherent manner on the physical model of their cities. An agent-based Model (ABM) has been used for simulating the actions and interactions of autonomous agents in this work. The central idea of the model is to show emerging patterns in visitors’ behavior during specific events [4]. This model leads to insightful visualizations that show how visitors and locals move across the country.

The remaining of this paper is organized as follows. Section 2 gives a general overview of the framework. Section 3 describes the input data, 4 describes the ABM model. Section 5 shows the visual results of the simulation and, finally, section 6 discusses further research.

Overview 

Andorra, located between Spain and France in the middle of the Pyrenees, is a country with a population around 78,000 people4 that welcomes more than eight million visitors a year. According to the statistics provided by the Departament d’Estad ́ıstica d’Andorra, the tourism sector accounts for 80% of the GDP of the country. Andorra has two types of visitors: (1) tourists, which stay over at least one night and (2) same-day visitors, which enter and leave the country the same day. The presented model has mainly been developed to simulate the movement of visitors across the territory and gain understanding on this industry. Modeling people’s flow can help us assess the actual impact of visitors in terms of traffic congestion, energy consumption, consumer spending, among others. The current model focuses on visitors’ attendance at the events held in the country as well as traffic congestion levels. The following two events in 2016 have been analyzed: (1) Cirque du Soleil, (2) Le Tour de France.

The resulting model is projected on the Andorra CityScope table, which is a 3D model of the two main cities of Andorra (see Fig. 1).

The model has been implemented using different environments: Processing [6] and the GAMA platform [3]. The latter offers a more generic framework that extend the model’s functionality and allow us to process other types of data. 

A data driven model

Figure 2 represents the different elements obtained from the input datasets: (1) telecom data, (2) road network, and (3) amenities. The ABM uses this different types of data to characterize both static and dynamic agents.

Telecom data: Andorra Telecom provided a three-year collection—from 2014 to 2016—of anonymized Call Detail Records (CDR). Observations in these records are triggered by any kind of action with a mobile phone (i.e., phone call, text message, cellular data). From the features describing each observation, we ob- tain the location of the cell towers involved in the action and, thus, compute the origin and destination of each agent. We can also assign the country of residence to the agents.

Road Network: Agents do not move in a straight line; their trajectories are constrained by the actual road network. Therefore, agents move along a graph topology, which is provided by Open Street Maps. Roads can be of different type (primary, secondary, residential, and pedestrian) allowing only certain be- haviors. For instance, vehicles are not allowed in pedestrian areas. Roads can be either one-way or bidirectional, but not all agents can go in both directions. The congestion level is updated during the simulation according to the number of agents present on the road and it can be modified to emphasize specific patterns (e.g., for traffic congestion).

Amenities: Amenities are places where agents may go, such as restaurants, hotels, or points of interest. Their geolocation have been gathered from TripAd- visor, Yelp, and the Andorra Tourism office. 

Model Description

Figure 3 corresponds to the simulation of a regular day—used as a benchmark in the analyses. Every simulation represents a full day and runs until all the observations from the CDR data set are processed.

The model deals with two types of agents: (1) static agents (i.e., buildings, roads, and amenities) and (2) dynamic agents (i.e., people and vehicles). People are represented by solid circles and vehicles by stroke circles; their color vary according to the country of residence—red refers to people from Spain, blue refers to people from France, and white refers to people from other countries.

At city scale, the two main, central cities of the country—Andorra la Vella and Escaldes-Engordany—are explicitly displayed using an ABM driven by Ge- ographic Information Systems (GIS) data (see map in Fig. 3). The rest of the territory is conceptually represented by clusters, which correspond to the two cities located near the border (i.e., Sant Juli`a de L`oria near the Spanish border and Pas de la Casa near the French border) and the parishes of Canillo, Encamp, Ordino, and La Massana (see pie charts in Fig. 3). The emerging structures show people’s flow from one city to another giving a general view of the activity at a country level. 

ABMs have successfully been applied to study emergences from a wide range of adaptive system made of individual entities, contributing to an easier and deeper understanding on local interactions, variability among entities, adaptive behaviors, and environmental states [1]. Lately, ABMs have also been used as a data visualization tool since they give the possibility to interact with the representation [2].

In the presented model, dynamic agents have a set of variables assigned that influence their behavior whenever a change occurs, either in its own state (e.g., when the agent arrives at its destination, it stops) or in the external environment (e.g., when a road is full, the agent can take an alternative path). The set of variables is composed of (1) country of residence, (2) origin location—defined randomly or using telecom data—, (3) preferred destination—generated by a decision making submodule—, (4) distance traveled, (5) speed of movement, and (6) passable streets. Agent’s trajectory is determined by an Origin-Destination (OD) matrix. The OD matrix is computed using the location of the cell towers where the action with the mobile phone was originated and terminated. The destination of the agent is set to the closest amenity to the cell tower where the action terminated. Depending on its speed (time difference between origin and destination location), the agent will be considered as a walking person (solid circle) or as a vehicle (stroke circle). The model is implemented with an enriched GIS data where the static agents provides information to the dynamics agents in order to adapt their movement such as amenities’ capacity and working hours, events, and direction of roads. Agents adapt themselves to both (1) congestion traffic and (2) amenity occupancy.

Congestion traffic. If congestion is too high, a pathfinding is called to recal- culate an alternative route. If a road is busy, then the agent will recompute the shortest path to its destination avoiding this road.

Amenity occupancy. Once agents reach their destination, they stay there for a few iterations.The number of iterations is defined by the average time spent on those places. The amenity size increases (or decreases) according to the number of agents currently in the location. Depending on the amenity occupancy, the agent might recompute its destination. If the amenity assigned as destination is full, the agent will select another amenity close to its initial destination. The chosen amenity also depends on the agent’s country of residence and the language affinity of the amenities.

Results

The emerging patterns display the actual dynamics of the city providing a urban planning tool that goes beyond the traditional ones that are usually focused on land uses and sociological static data extracted from surveys. The ABM visualization shows different patterns of movements from visitors revealing the structure of the city as a complex system. The following subsections describe (1) raw and (2) aggregated results, which highlight helpful information regarding Cirque du Soleil and Le Tour de France. 

Raw Results

When running a simulation on the Andorra CityScope table, one can imme- diately identify three main elements: (1) city representation defined by static agents (i.e., buildings, amenities, cell towers, roads), (2) people’s movement de- fined by dynamics agents, and (3) amenities’ density.

As mention in Section 4, the number of people that are present in the ameni- ties evolves during the simulation. The amenity size increases (or decreases) according to the number of agents currently in the location. This helps identify which and when places are popular or busy and isolate them. For instance, Fig. 4 shows the activity during the Cirque du Soleil on July, 16. The wide white circle spots the location where the show was taking place. This was the most dense place at that time; ticketing for the event was 5,174 attendees. According to the stats from the Andorra Turisme report, the average attendance per performance was 4,540 people in 2016. Overlapping layers from different days and/or editions of the event—while dynamically running them—is a useful way to display the stats and discuss the numbers.

Aggregated Results

Unlike Cirque du Soleil, both Le Tour de France is an outdoor events and do not has any kind of ticketing that helps assess attendance. To this end, aggregated data can be visualized on the CityScope table resulting in heatmaps that summa- rize global activity in the city and provide an attendance estimate. Figures 5(a) and 5(b) show occupancy levels for Le Tour de France on July, 12. The starting line was in Escaldes-Engordany, which corresponds to the hottest area in Fig. 5(b)—large red concentration on the left side of the image. Comparing this kind of visualization to the activity of a regular day, one can understand which events bring more people to the country. In addition, we can identify where these vis- itors go and what they do. This could be used to efficiently plan events or find new ones that help spread visitors across the territory. Figure 6 shows congestion levels. Focusing on the roads only, the movement of agents representing vehicles can be translated into another view based on traffic density. 

Discussion and Further Research 

Data Accuracy from CDR to RNC 

The CDR data sets only allows to trace visitors’ movement based on the geolo- cation of the cell towers. To improve the accuracy of the ABM, Andorra Telecom is collecting a new source of data from the Radio Network Controller that can greatly improve the accuracy of the geolocation of the devices. Observations for this data source are triggered (1) by any kind of action with your phone (phone call, text message, cellular data), (2) when the user moves and the network de- tects the device changes cells or technology (i.e., 2G, 3G, 4G), or (3) when the user is static and the update network timer expires.

Data Laboratory 

Other parameters such as proximity, hour in the day, price, or attractiveness factor could be implemented to enhance how agents’ destination is selected. Further work includes integrating data coming from sensors deployed in stores to study how visitors move and behave inside buildings. This model will also be used to replay agents’ behavior in order to understand the city dynamics and lead to more efficient urban designs. To this point the CityScope table is coupled with an interactive table that allows the user to modify the structure of the city and immediately see the impact on agent behavior.

Acknowledgment.

This work has been developed within the framework of collaboration between MIT Media Lab City Science and Fundacio ActuaTech.

References

1. Robert John Allan. Survey of agent based modelling and simulation tools. Technical report, 2009.
2. Agent-Based Visualization: A Real-Time Visualization Tool Applied Both to Data and Simulation Outputs. A Grignard, A Drogoul. The AAAI-17 Workshop on Human-Machine Collaborative Learning
3. Arnaud Grignard, Patrick Taillandier, Benoit Gaudou, Duc An Vo, Nghi Quang Huynh, and Alexis Drogoul. Gama 1.6: Advancing the art of complex agent-based modeling and simulation. In International Conference on Principles and Practice of Multi-Agent Systems, pages 117–131. Springer, 2013.
4. Volker Grimm, Eloy Revilla, Uta Berger, Florian Jeltsch, Wolf M Mooij, Steven F Railsback, Hans-Hermann Thulke, Jacob Weiner, Thorsten Wiegand, and Donald L DeAngelis. Pattern-oriented modeling of agent-based complex systems: lessons from ecology. science, 310(5750):987–991, 2005.
5. Yan Leng, Larry Rudolph, Alex ’Sandy’ Pentland, Jinhua Zhao, and Haris N Kout- sopolous. Managing travel demand: Location recommendation for system efficiency based on mobile phone data. arXiv preprint arXiv:1610.06825, 2016.
6. Casey Reas and Ben Fry. Processing: a programming handbook for visual designers and artists. Number 6812. MIT Press, 2007.