Procedia Computer Science 37 ( 2014 ) 382 – 389
1877-0509 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).
Peer-review under responsibility of the Program Chairs of EUSPN-2014 and ICTH 2014. doi: 10.1016/j.procs.2014.08.057
ScienceDirect
The 1st International Workshop on Information Fusion for Smart Mobility Solutions (IFSMS’14)
An Alternative Approach to Network Demand Estimation:
Implementation and Application in Multi-Agent Transport
Simulation (MATSim)
Enock T. Mtoi
a,∗, Ren Moses, PhD, PE
a, Eren Erman Ozguven, PhD
aaFAMU-FSU College of Engineering, 2525 Pottsdamer St, Tallahassee 32310, Florida
Abstract
This paper introduces a novel network demand estimation framework consistent with the input data structure requirements of Multi-Agent Transport Simulation (MATSim). The sources of data are the American Community Survey, US Census Bureau, National Household Travel Surveys, travel surveys from South East Florida Regional Planning Authority, OpenStreetMap and Florida Statewide Transportation Engineering Warehouse for Archived Regional Database. The developed framework employs mathematical and statistical methods to derive probability density functions and multinomial logit models for activity and location choices. The implementation of demand estimation process resulted into the creation of 1,200,889 agents (only those using cars). The scenario for the estimated agents was configured and simulated in MATSim. The results from the simulated scenario resulted in the expected morning, afternoon and evening traffic patterns as well as the desirable level of agreement between simulated and observed traffic volumes.
c
2014 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the Program Chairs of IFSMS-2014.
Keywords: Demand estimation; Multi-agent simulation; MATSim; Smart mobility
1. Introduction
In today’s economy, businesses and individuals are looking for ways of making their fiscal resources and workforce more efficient. However, traffic congestion dampens the efficiency and prosperity by imposing additional operating costs, slowing mobility and causing delays and by hindering efficient metropolitan services such as deliveries, public safety and maintenance. Traffic congestion in the United States in 2011 for instance, caused urban commuters to travel 5.5 billion hours more and to purchase an extra 2.9 billion gallons of fuel for a congestion cost of $121 billion1. Urban
traffic congestion also brings broader social and environmental costs, such as delayed delivery of goods, air pollution (particulate matter, black carbon, total nitrogen oxides, and nitrogen dioxide)2, micro-climate change and urban heat islands3, and increases in respiratory problems and lost productivity4. Furthermore, in large urbanized areas and in busy expressways, traffic infrastructures are already operating at near or full capacity. With today’s shrinking budgets,
∗Corresponding author. Tel.:+1-850-980-2755 ; fax: +1-850-410-6580. E-mail address: [email protected]
Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).
often limited funding is available to rebuild or expand an aging public transportation infrastructure, making it crucial to devise ways to optimize the performance of existing transportation assets. While the world is becoming increasingly urbanized5, urban travel demand leads to more cars into the roadways. With more cars added to the roads every day,
congestion alleviation and incident response have become the biggest issue for traffic operators and planners. There is a need to improve the speed of detection and clearance of incidents in order to avoid delays due to stoppage leading to miles of bottleneck. Timely, coordinated action to control traffic and to guide motorists, based on predictive modeling of the network, can alleviate painful conditions. And, in many cases, it can also help traffic operators to get ahead of problems and to stop them from happening. Without accurate and innovative simulation techniques it is very difficult to identify where improvements can be made, leaving the transportation system operations without the tools needed to analyze and devise mechanisms to get the most from the available resources.
Effective and efficient operations of transportation networks can have a direct influence on people’s pockets. It can also have a positive influence on environment, local economic activities and dramatically impact quality of life. Operational effectiveness and efficiency in transportation networks can be achieved with the use of user-behavior and system performance models and simulation tools. In transportation system performance analyses, travel demand models and simulation tools are crucial for the evaluation of congestion scenarios and for testing the viability of congestion mitigation measures. Nevertheless, the accuracy of travel demand models and network simulation tools largely depend on the data and techniques employed in network demand estimation.
Robust network demand estimation techniques are important for the analysis and modeling user-behavior and decision making dynamics in transportation network. User-behavior and decision making dynamics in transportation network are vital in modeling and simulation of user interactions, which in turn can be used to identify causes of traffic congestion and devise viable mitigation measures. The availability of data for modeling user behavior and decision making dynamics, is increasingly growing due to the increased uses of information and communication technology (ICT). But turning data into timely network performance and user behavior inferences for taking action is the new and encouraging approach in the world of intelligent transportation systems (ITS). Modern advances in ITS research and practice define an efficient congestion management policy as the one grounded on its capability to capture behavioral responses from the users of the transportation network system6. Systematic proficiency of activity scheduling and travel behavior modeling in a diversity of circumstances is essential for achieving effective transportation systems management.
In this paper we introduce the network demand estimation framework consistent with the input data structure re-quirements of Multi-Agent Transport Simulation (MATSim)7,8. To achieve better simulation results, agent-based
models require the input demand to be structured in such a way that agents are grouped into households, and even better grouped into social networks9,10,11.The demand estimation framework utilizes data from multiple sources as
initial inputs. In order to estimate the demand and store data efficiently several statistical and analytical methods were applied for information extraction and fusion. The estimated demand was consequently used for transportation net-work simulation in MATSim. Based on its flexibility to accept netnet-work demand from various demand estimators12,13,
MATSim was chosen for simulating the estimated demand.
Multi-agent based simulation models have been used widely and proved to be very useful in transportation mod-eling and simulation7,14,15,16. In such models, every network user is represented by an object in software, a so-called
agent. Each agent can have its own values for a given set of attributes, leading to very detailed description of a model-population. For this reason, such simulations are sometimes also categorized as microscopic. Examining the principal operations of multi-agent based transport simulation models, they are microscopic activity-based models. In multi-agent simulations, each agent may decide on its own, influenced by its specific set of attributes and how it behaves (given a set of rules). For this decision making process, additional complex models could be used, e.g. route and mode choice or departure time choice models.
2. Research Motivation
The process of model building, estimation of model parameters, derivation of sampling distributions and valida-tion of model scenarios are data-driven tasks requiring extensive and robust data of regional populavalida-tion and their socio-economic attributes, travel characteristics, time dependent traffic patterns and information regarding network congestion management strategies. In most reported practical applications of multi-agent models for the microscopic
traffic simulation12,16,17,18,19, demand was estimated from public interviews or household travel surveys. For instance,
in the applications of MATSim12,18,19, the initial demand has been derived from the agent’s (individual’s) chain of
activities and tours that include detailed information about departure time, destination and mode for each tour leg. The derivation of agents’ activity chains and tours in MATSim have largely depended on micro-census surveys18or
household travel surveys12,19. The use of micro-census survey data or household travel surveys in scenario simulation
can lead to desirable results provided the study region is small enough to be covered by surveys at a viable cost. However, the cost of conducting household travel characteristics surveys increases with the increase in the size of the study area. Because of high cost for conducting house-to-house surveys in large urbanized areas, modelers and analysts have generally been using data from a representative sample of households to capture travel characteristics for the entire region. This leads to under-estimation of transportation network demand if there are no additional data to supplement the sampled household travel surveys. Thus, this paper was motivated by the need to acquire data from different sources and apply data fusion techniques using advanced statistical techniques for enhancing synthetic demand generation. The techniques and steps used to develop and implement the proposed framework are described in the following sections.
3. Methodology
The need for innovative methods for network demand estimation and multi-agent based simulation techniques stem from wide ranging applications in modeling commuter behavior, network state prediction, and traffic management and planning. The achievement of efficient demand estimator and accurate estimated demand depends on the proper design of the framework and its capability to extract useful information from the input data. The following subsections give descriptions of the input data, the structure of the network demand estimation framework and the integration of the framework into MATSim for simulation.
3.1. Description of Input Data
The study area is a region containing three urban areas of Southeast Florida, i.e., Palm Beach, Broward and Miami-Dade counties. Multiple data sources with potential information on how to derive agents for the regional network demand were used. The data was collected from the following sources:
1. American Community Survey (ACS)20
2. US Census Bureau20
3. National Household Travel Survey (NHTS)21
4. South East Florida Regional Planning Model (SERPM) Region22
5. OpenStreetMap (OSM)23
6. Florida Statewide Transportation Engineering Warehouse for Archived Regional Data (STEWARD) Database The ACS data is based on questions asked of about three million housing units annually, such as income, education, occupation, and mode of travel to work. The ACS data in combination with census data were used to derive several demographic profiles which are crucial for the estimation of network demand and household’s or individuals’ attributes (i.e., agents’ attributes). The data from SERPM region and the NHTS data contained individuals’ travel characteristics such as origin, destination, departure and arrival time, chain of activities and tours, etc. For the agents to access the facilities where they accomplish the activities which triggered the trips, they need the street network (comprising of local streets and highways). The street network for the study area was extracted from OSM23by using Osmosis23(a Java library for accessing and modifying OSM data). The STEWARD database was used as an important source for accessing traffic counts data for validation of regional travel simulation outputs. The database contains volume, speed and occupancy data from automatic traffic count stations located in the main highway system.
3.2. Demand Estimation Framework
The framework developed (Figure 1) in this research employed several mathematical and statistical methods for the derivation of sampling distributions of users’ (i.e., agents’) behavior and travel characteristics for the initial network
demand generation. The processes of deriving sampling distributions of agents’ behavior and travel characteristics, largely rely on the quantity, quality and resolution of the available data of the region under study. Travel character-istics/travel surveys data from SERPM region and the NHTS data contained individuals’ travel characteristics such as origin, destination, departure and arrival time, chain of activities and tours within the trip. These are micro-information needed for the derivation of household and individual agent’s travel behavior. The data was processed to develop probability distributions for groups of agents with similar travel behavior, given the agents’ household characteristics. In a similar fashion, with agents’ household characteristics given, the logit models for agents’ activity and locations choices were developed . The probability distributions were developed on the basis of every household
h(h∈ H where, H is the set of all households in the study area), with socio-economic attributes s(s ∈ S where, S
is the set of all socio-economic attributes such as household income, number of vehicles, number of licensed drivers, family size and structure, number of household members with employment, etc.). The probability distribution func-tion,φct∈T∈C(sh ∈ S), entails that an agent from household h with socio-economic shis likely to make a trip with trip
characteristics c (whereC is the set of all trip characteristics such as number of stops, trip length, trip origin and destination, trip purpose etc) at time t∈ T (where T is the set of all time intervals within simulation horizon, typ-ically 24 hours). Using the probability distribution function, we can draw samples of agents from households and their trip characteristics. However, the agents still lack the choices of activities a (a∈ A, where A is the set of all possible activities) and the choices of locations l, viable to accomplish the chosen activities. Therefore, the activity and location choice logit models were developed. The logit models, ft(a∈A, l∈L)∈T (sh∈ S) were specified to predict that,
an agent from a household h∈ H with socio-economic attributes s ∈ S will choose to accomplish activities a ∈ A in locations l∈ L within the study area (where L is the set of all activity locations or facilities in the study area). This entire process takes place in the module labeled Statistical Methods\ Logit Models in Figure 1.
Fig. 1. Flow chart for modeling within-day network user activity and behavior dynamics
Since the travel characteristics data (survey data from SERPM region and the NHTS data) were collected for individual household members (agents) from sampled households in the study area, we derived a portion of initial agents population and their activities directly from the data. The agents’ homes and their activity locations were determined as x and y coordinates taking advantage of prior geocoding of households and trip destinations. However, not all the households were covered by the survey data from SERPM region and the NHTS data. Therefore, in order to estimate the population of agents for the households which were not covered by the surveys in the study region the probability functions, the logit models and the census data are used. The census data covers the entire study area but
the data is aggregated in groups with the similar socio-economic characteristics for every traffic analysis zone (TAZ) in the area. For every household, h in the TAZ, the probability functions draw samples of agents with their possible trips, given the household socio-economic attributes from the census data. This process takes place in the Trip Estimator module of Figure 1. The agents and their trips are distributed to the households which were not covered in the derivation of the initial portion of agents population using travel characteristics data. In the Initial Demand Generation module, the distributed agents and the agents derived from the travel characteristics data are combined to generate an initial synthetic population. The initial synthetic population generated, the land-use data (at parcel level), the street network data and the traffic counts are treated through the Fusion Process module. The outputs from the fusion process are the street network, population of agents and facilities (addresses of agents’ homes and activity locations derived from land use data and travel characteristic data) all in the form of Extensible Markup Language (XML)24
files. However, agents drawn from the census data were still lacking activities and activity locations. Therefore, the portion of initial synthetic population created from census data, was modified to include a set of agents’ activities they intend to accomplish and the locations where those activities are likely to take place. The activity choices, the location choices logit models and the ACS data were used to define the agents’ activities and activity locations. The XML files of population, network and facilities are included in this process in order to avoid duplication of agents and to provide the correct spatial attributes of activity locations with reference to the network and the facilities. After this phase in the steps of demand estimation process, files are finally written for the individual demand, the facilities and the network in the XML format ready for simulation in MATSim.
3.3. Scenario Configuration and Simulation in MATSim
The simulation scenario in MATSim is created with the notion that agents make independent decisions about their actions. Each network user of the real system is modeled as an individual agent. An important component of the input data is the agent plan (accessed from Individual Demand.xml) which encodes the intentions of an individual agent for the day. A plan contains the itinerary of activities the agent wishes to perform during the day, and the trips (including chains of tours making up the trip) the agent must take to travel between activities. An agent’s plan details the order, type, location, duration and other time constraints of each activity, and the mode (car, in our case), route and expected departure and travel times of each leg. The plans for all agents’ are simultaneously executed in the simulation of the physical system by the mobility simulation module (MobSim)7 which is in the EXEC step of Figure 1) within
MATSim. While traffic flow simulation is running iteratively between plans, there is a mechanism that allows agents to learn and update their behavior. Depending on the configuration of the simulation scenario, the system keeps in memory several plans for each agent, and scores the performance of each plan using the utility-based scoring function (implemented in the SCORING module). For every agent j, the score of an executed plan25is calculated using the
modified utility function based on the Vickrey bottleneck model26,27:
Uplanj = n i=1 Uper fj ,i+ n i=1 Utravelj ,i+ n i=1 Ulatej ,i (1)
where, n is the number of activities for a given plan, Uper fj ,iis the logarithmic form of positive utility for performing an activity i, Utravelj ,iis the linear form of negative utility for time spent traveling and Ulatej ,iis the linear form of negative utility for time being late(for details, see Charypar and Nagel28). Agents normally choose plans with the highest
score, re-evaluate plans with bad scores, and obtain new plans through the REPLANNING module.
The simulation scenario for SERPM region was configured for base year 2010. The region has a total population of 5,564,636 people according to 2010 census20. The demand estimation process resulted into the creation of 1,200,889
agents (only those using cars). The simulation was implemented on Dell XPS 8700 with 8.0 GB RAM, 8 CPU cores running on Windows 8. The Java codes for both demand estimation and MATSim engine were executed with Java Virtual Machine (JVM) version 1.7.0-21 (64-bit mixed mode) on eclipse29. The simulation reached the convergence
after about 300 iterations but the maximum number iterations was set to 400 iterations for more stable results as shown in Figure 2. With the allocation of maximum memory of 9102.25MB to the simulation configuration, the machine took 147 hours to execute the demand estimation process and run 400 iterations in MATSim.
In Figure 2, it shows that when iterations are executed, the worst and average scores drop to lower scores than the starting values and then increase gently towards convergence. This is the behavior confirming that, the simulated
Fig. 2. Update of Agents’ Scores as a Function of Simulation Iterations
agents are learning and updating their plans after every iteration. The worst and the average scores are the initial scores that agents evaluates and after replanning, they achieve plans with the best score which they re-evaluate and come up with plans having viable score for execution.
Fig. 3. Distribution of Agents by Time of Day
The observation of network traffic evolution showed the expected traffic patterns; morning peak (7-8 a.m) and afternoon peak traffic (5-6 p.m.). Traffic evolution is the measure of how the network links are loaded at a particular time of day in relation to agent plans. In Figure 3 we can see that, most of the agents depart their origin and arrive at their destinations around 7-8:00 a.m. This is a characteristic of regular commuters, mostly working class; they mostly
tend to leave their homes around 6-8 a.m., and since their trips are shorter, they form a balance between departures and arrivals within approximately the same time period. The results on Figure 3 also show an increasing traffic between 10 a.m. and 3 p.m. This could be the amount of traffic comprising shopping trips, shift workers’ trips, leisure trips etc. The evening peak traffic occurs between 5-6 p.m. and is lower than morning peak. The lower peak traffic in the evening compared to morning peak could be due to some part time workers leaving their working places during the day. This can further justify the increase of traffic in the mid-day (10 a.m.-3 p.m.). The amount of traffic experiencing severe congestion (stuck) increases during both morning and evening peak periods.
Fig. 4. Comparison of Simulated and Observed Volumes
In order to check the accuracy of the simulation model, traffic counts from STEWARD database were used. The comparison of volumes for simulated and observed vehicles was performed. The comparison of the simulated volumes with observed volumes from count stations (Figure 4) shows satisfactory degree of agreement between the simulation outputs and the actual traffic. The line of zero intercept in Figure 4 indicate that, on average the simulation model produces about 78.4% of the observed volume. In other words, using the line with intercept, the simulation model produces about 73 vehicles less than 74.19% of the observed volume from count stations. However,it is worth noting that, comparison of the simulated volumes with observed volumes was only performed for freeway links where count stations were available.
4. Conclusions and Future Research
This paper introduced network demand estimation framework consistent with the input data structure requirements of MATSim. The sources of data were ACS, US Census Bureau, NHTS, SERPM travel surveys, OSM and STEWARD database. The study area in this research was a region containing three urban areas of Southeast Florida. The devel-oped framework employed mathematical and statistical methods for the derivation of probability density functions and multinomial logit models for activity and location choices. The probability density functions and multinomial logit models were used to estimate the population of agents and their attributes for the households not covered by NHTS and SERPM travel surveys. There were a total of 1,200,889 agents estimated which was the sum of all agents estimated directly from NHTS and SERPM travel surveys, and those estimated using the probability density functions and multinomial logit models. The scenario for the estimated agents was configured and simulated in MATSim. The results from the simulated scenario showed the expected morning, afternoon and evening traffic patterns. Also, the comparison of the simulated volumes with observed volumes from count stations showed desirable level of agree-ment despite a slight underestimation of simulated traffic. Future directions for this research include the inclusion of
other data sources and optimization of the demand estimation framework in order to scale-down the computation cost. In addition to the reduction of computation cost, focus will be on development and implementation of modules for simulating dynamic toll pricing on high occupancy toll lanes using and assessing the effects of information exchange among the agents on mobility.
References
1. D. Schrank, B. Eisele, T. Lomax, Ttis 2012 urban mobility report, Texas A&M Transportation Institute. The Texas A&M University System. 2. T. Outlook, Focusing on co2 emissions from road vehicles, Technical report, Joint Transport Research Centre of the OECD and the
Interna-tional Transport Forum (May 2008).
3. Chicago, Chicago climate change action plan, Technical report, Chicago Climate Task Force (2007).
4. J. J. Kim, S. Smorodinsky, M. Lipsett, B. C. Singer, A. T. Hodgson, B. Ostro, Traffic-related air pollution near busy roads: the east bay children’s respiratory health study, American journal of respiratory and critical care medicine 170 (5) (2004) 520–526.
5. N.C.S.U, Mayday 23: World population becomes more urban than rural (May 2007). URL www.sciencedaily.com/releases/2007/05/070525000642.htm
6. IBM, Intelligent transportation solution for active traffic management, Technical Report GVF03008USEN, IBM Corporation, Systems and Technology Group (November 2008).
7. M. Balmer, K. Meister, M. Rieser, K. Nagel, K. W. Axhausen, K. W. Axhausen, K. W. Axhausen, Agent-based simulation of travel demand: Structure and computational performance of MATSim-T, ETH, Eidgen¨ossische Technische Hochschule Z¨urich, IVT Institut f¨ur Verkehrspla-nung und Transportsysteme, 2008.
8. A. Horni, D. M. Scott, M. Balmer, K. W. Axhausen, Location choice modeling for shopping and leisure activities with matsim, Transportation Research Record: Journal of the Transportation Research Board 2135 (1) (2009) 87–95.
9. T. Arentze, F. Hofman, H. van Mourik, H. Timmermans, Albatross: multiagent, rule-based model of activity pattern decisions, Transportation Research Record: Journal of the Transportation Research Board 1706 (1) (2000) 136–144.
10. K. W. Axhausen, Social networks and travel: Some hypotheses, Social Dimensions of Sustainable Transport: Transatlantic Perspectives (2005) 90–108.
11. J. K. Hackney, K. W. Axhausen, J. K. Hackney, J. K. Hackney, K. W. Axhausen, K. W. Axhausen, An agent model of social network and travel behavior interdependence, ETH, Eidgen¨ossische Technische Hochschule Z¨urich, IVT, Institut f¨ur Verkehrsplanung und Transportsysteme, 2006.
12. W. Gao, M. Balmer, E. J. Miller, Comparison of matsim and emme/2 on greater toronto and hamilton area network, canada, Transportation Research Record: Journal of the Transportation Research Board 2197 (1) (2010) 118–128.
13. A. L. Bazzan, K. Nagel, F. Kl¨ugl, Integrating matsim and itsumo for daily replanning under congestion, in: Proceedings of the 35th Latin-American Informatics Conference, CLEI, 2009.
14. P. Davidsson, L. Henesey, L. Ramstedt, J. Trnquist, F. Wernstedt, An analysis of agent-based approaches to transport logistics, Transportation Research Part C: Emerging Technologies 13 (4) (2005) 255 – 271, agents in Traffic and Transportation: Exploring Autonomy in Logistics, Management, Simulation, and Cooperative Driving. doi:http://dx.doi.org/10.1016/j.trc.2005.07.002.
URL http://www.sciencedirect.com/science/article/pii/S0968090X05000318
15. S. Gallanda, L. K. Ansar-Ul-Haque Yasarb, N. Gauda, T. Bellemansb, D. Janssensb, Simulation of carpooling agents with the janus platform, Journal of Ubiquitous Systems & Pervasive Networks 5 (2) (2014) 09–15.
16. S. Galland, N. Gaud, A.-U.-H. Yasar, L. Knapen, D. Janssens, O. Lamotte, Simulation model of carpooling with the janus multiagent platform, Procedia Computer Science 19 (2013) 860–866.
17. S. Galland, J. Buisson, N. Gaud, M. Gonc¸alves, A. Koukam, F. Guiot, L. Henry, Agent-based simulation of drivers with the janus platform, Procedia Computer Science 32 (2014) 738–743.
18. J. K. Hackney, F. Marchal, J. K. Hackney, J. K. Hackney, Model for coupling multi-agent social interactions and traffic simulation, ETH, Eidgen¨ossische Technische Hochschule Z¨urich, IVT, Institut f¨ur Verkehrsplanung und Transportsysteme, 2007.
19. S. Bekhor, C. Dobler, K. W. Axhausen, Integration of activity-based and agent-based models, Transportation Research Record: Journal of the Transportation Research Board 2255 (1) (2011) 38–47.
20. DOC, 2010 census, Technical report, US Department of Commerce, Census Bureau (2010). 21. USDOT, 2009 national household travel survey.http://nhts.ornl.gov (2009).
URL http://nhts.ornl.gov
22. ICPSR, Inter-university consortium for political and social research (2013). URL https://www.icpsr.umich.edu/icpsrweb/landing.jsp 23. OSM, Osmosis - openstreetmap wiki (2013).
URL http://wiki. openstreetmap. org/wiki
24. T. Bray, J. Paoli, C. M. Sperberg-McQueen, E. Maler, F. Yergeau, Extensible markup language (xml), World Wide Web Journal 2 (4) (1997) 27–66.
25. M. Rieser, K. Nagel, Network breakdown at the edge of chaos in multi-agent traffic simulations, The European Physical Journal B-Condensed Matter and Complex Systems 63 (3) (2008) 321–327.
26. W. S. Vickrey, Congestion theory and transport investment, The American Economic Review 59 (2) (1969) 251–260.
27. R. Arnott, A. De Palma, R. Lindsey, A structural model of peak-period congestion: A traffic bottleneck with elastic demand, The American Economic Review (1993) 161–179.
28. D. Charypar, K. Nagel, Generating complete all-day activity plans with genetic algorithms, Transportation 32 (4) (2005) 369–397. 29. R. F. Araujo, Getting Started with Eclipse Juno, Packt Publishing Ltd, 2013.
