Scholarly article on topic 'An Agent-based Approach for Modeling Real-time Travel Information in Transit Systems'

An Agent-based Approach for Modeling Real-time Travel Information in Transit Systems Academic research paper on "Materials engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"Travel information" / Agent-based / "Traffic simulation" / Prediction.}

Abstract of research paper on Materials engineering, author of scientific article — Oded Cats

Abstract Real-time travel information (RTI) systems are rapidly developed and deployed worldwide using the abundance of instantaneous data and dissemination means. This paper presents a framework for a multi-agent simulation model that emulates the generation and dissemination of RTI. The evolution of transit reliability influences both the performance of RTI generation schemes and the potential benefits that such information could yield. An iterative within-day network loading and a day-to-day learning process represent both service provider and service user ability to apply and adapt their strategies based on past performance and predictions. A case study illustrates model capabilities by applying BusMezzo, an agent-based simulation model of vehicles and travellers. The proposed model facilitates the analysis of alternative prediction schemes as well as the impact of their provision on system performance.

Academic research paper on topic "An Agent-based Approach for Modeling Real-time Travel Information in Transit Systems"

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedia Computer Science 32 (2014) 744 - 749

The 3rd International Workshop on Agent-based Mobility, Traffic and Transportation Models, Methodologies and Applications (ABMTRANS)

An agent-based approach for modeling real-time travel information in transit systems

Oded Catsa*

a Department of Transport Science, Royal Institute of Technology KTH, Teknikringen 10A, Stockholm 100 44, Sweden

Abstract

Real-time travel information (RTI) systems are rapidly developed and deployed worldwide using the abundance of instantaneous data and dissemination means. This paper presents a framework for a multi-agent simulation model that emulates the generation and dissemination of RTI. The evolution of transit reliability influences both the performance of RTI generation schemes and the potential benefits that such information could yield. An iterative within-day network loading and a day-to-day learning process represent both service provider and service user ability to apply and adapt their strategies based on past performance and predictions. A case study illustrates model capabilities by applying BusMezzo, an agent-based simulation model of vehicles and travellers. The proposed model facilitates the analysis of alternative prediction schemes as well as the impact of their provision on system performance.

© 2014 TheAuthors.PublishedbyElsevierB.V.This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/3.0/).

Selection and Peer-review under responsibility of the Program Chairs.

Keywords: Travel information; Agent-based; Traffic simulation; Prediction.

1. Introduction

The performance of transport systems is the outcome of numerous interactions between agents who pursue their travel strategy and interact with other agents. The uncertainty associated with the outcome of these complex interactions has implications on travel reliability. The latter is an important determinant of system performance and is associated with high welfare costs. Predictions of future system states facilitate more proactive operations and adaptive travel strategies by system operators and users,

* Corresponding author. Tel.:+46 87908816; fax: +46 8212899. E-mail address: cats@kth.se.

1877-0509 © 2014 The Authors. 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/).

Selection and Peer-review under responsibility of the Program Chairs.

doi: 10.1016/j.procs.2014.05.485

respectively. Real-time travel information (RTI) systems are rapidly developed and deployed worldwide capitalizing on the abundance of instantaneous data as well as dissemination means.

RTI could be classified based on the travel attribute that is provisioned, the prediction horizon and comprehensiveness as well as the dissemination mean. In the context of transit, the most common RTI refers to the next vehicle arrival at stop. However, RTI could provide information on additional travel attributes such as expected in-vehicle time or on-board crowding conditions. Information could be disseminated through on-site signs or personal mobile devices. The dissemination mean determines the availability and comprehensiveness of the information and hence the trip stage decision that the information can support. Finally, RTI can vary from descriptive (e.g. "disruption on the metro line"), through quantitative ("the metro runs 20 minutes behind schedule") to predictive ("the metro is expected to arrive 15 minutes behind schedule") and even prescriptive ("passengers heading to the university are advised to take the tram line"). These different levels of RTI require an increasing complexity of generation and dissemination strategies.

Conventional transit assignment models that investigate the impact of RTI on travellers' decisions assume that: (a) RTI is perfectly correct; (b) travellers perceive the RTI to be perfectly credible1,2. Modelling the impact of RTI is therefore considered as equivalent to the assumption that travellers have perfect knowledge concerning the provisioned travel attribute. Conventional approaches are therefore inadequate for analysing the impacts of RTI. Moreover, RTI was considered universal with respect to its availability and impact. MILATRAS, a transit simulation model, considered the impacts of RTI concerning vehicle arrival times on en-route travel decisions3. However, RTI is assumed to be universally available and perceived as credible. BueMezzo, a transit simulation model, illustrated the potential time savings from providing RTI on service disruptions on a small sub-network4.

In order to evaluate the impacts of various RTI systems it is necessary to model transit system dynamics through the adaptive decisions taken by individual vehicles and travellers. This paper presents an agent-based model which enables the generation of RTI based on individual vehicle progress and prediction schemes that are embedded into the transit simulation model. The explicit modelling of RTI generation as function of dynamic supply conditions enables the analysis of alternative dissemination strategies and predictions schemes and their impact on travellers' decisions and ultimately on travellers' flows. Note that RTI is therefore not equivalent to modelling the impact of perfect information. The performance and the prognosis depend on the interaction between numerous vehicle agents whilst the impact of RTI provisioned depends on its incorporation into the route choice of travellers' agents.

2. Modelling framework

The agent-based transit operations and assignment model, BusMezzo, emulates transit performance. Simulated data can be processed in order to generate predictions on future transit conditions that are disseminated to travellers. Traffic dynamics are represented in a mesoscopic level of detail, while passengers' decisions and transit vehicle operations are modelled microscopically. Passengers' progression in the network is dictated by the decisions they make in reaction to transit conditions, such as transit vehicle arrivals at stops, denied boarding and information provision. At the same time, travellers' decisions affect transit performance through the effect of passenger flows on crowding, dwell times and their secondary implications on service reliability5. Service reliability is particularly important in this context since its influences both the performance of RTI generation schemes and the potential benefits that such information could induce.

An agent-based modelling framework for RTI is presented in Figure 1. The model consists of a day-to-day learning model that executes an iterative within-day traffic and network loading model6. The framework presents only the components directly relevant to RTI. The agent-based transit simulation model mimics the progress of individual vehicles and travellers based on the outcome of traffic dynamics, transit operations and travellers decisions. The current state of all simulated agents at a certain time instance corresponds to the availability of instantaneous data such as traffic counts, vehicle positions and

passenger counts. The within-day loop results with outputs concerning the performance of the transit system and passenger flows for individual trips which can consequently be used for adjusting the planned service. In addition, by comparing predictions against the respective travel attributes, service provider can evaluate the accuracy and the reliability of RTI prognosi7. The contribution of system evolution through iterative network loading is thus twofold: the accumulation of historical travel attributes, and; the ability to improve the RTI generation scheme itself through day-to-day improvements.

Day-to-day learning model

Fig. 1. Real-time information in the agent-based transit model framework

3. Generation

The RTI generator could deploy various methods for predicting future system states based on current

and historical simulated data as well as the planned service. Prediction schemes could be based on simple prognosis rules or consists of machine learning algorithms where the weights given to various inputs are optimized based on its capability to reproduce past performance. Information and communication technologies are essential for data collection; transmission and processing in a central database and; dissemination to digital displays on-site and through virtual platforms.

Automatic data collection methods such as automatic vehicle location (AVL), automatic passenger counts (APC) and automatic fare collection (AFC) facilitate the generation of real-time predictions concerning transit performance. Methods for generating RTI include regression models, artificial neural networks, K-nearest neighbours and statistical pattern recognition, among others. These methods require inputs concerning current vehicle positions, historical running times and dwell times, and timetables.

The dynamic and disaggregate representation associated with agent-based simulation model facilitates the inclusion of various information sources in the RTI generator. The prediction of future states of the transit system could take into account the following information sources: (a) planned service - timetables and vehicle schedules. The planned service could be used as the reference when making predictions such

as expected dispatching times, arrival times and travel time between stops. Vehicle scheduling can identify the potential propagation of delay; (b) current state of the system - the current positions of all transit vehicles are essential for specifying the initial values for future prediction such as computing downstream trajectories for each vehicle; (c) previous within-day states - the progress of previous vehicles/travellers could be embedded into prediction schemes. The time it took the previous vehicle(s) to traverse a certain segment could be indicative of local traffic dynamics. Furthermore, the simulation can compute the travel time experienced by earlier travellers and obtain similar information to what available from crowdsourcing; (d) previous day-to-day states - the iterative network loading procedure mimics day-to-day variations in travel conditions which can correspond to the accumulation of historical data. The historical data could be filtered based on external variables or internal system states.

Alternative RTI generation schemes could be formulated for predicting a certain travel attribute for trip k on day d by specifying the following function:

<d = f(<d ;<d Vk = 1..k;<d Vk = 1..d) (1)

Where ntk d is the respective planned/scheduled value and n^ d is the actual travel attribute with trips and days sorted by order of occurrence.

Table 1. Simulated versus real-world data sources for generating real-time information

Simulation data Real-world data collection method

Timetables Public information

Vehicle schedules Operators' database

Positions of transit vehicle agents AVL

Number and speed of non-transit vehicle agents Traffic counts

per link

Number of traveler agents at stop AFC, in case ticket validation is required

upon entrance/arrival

Number of traveler agents on-board APC/AFC

Travel time experienced by vehicle agents GPS, plate recognition, mobile phones

Travel time experienced by traveler agents AFC in case of closed system,

crowdsourcing service

The availability of all information concerning current and historical states of the transport system allows assessing the potential value from its incorporation in the prognosis. Table 1 lists the simulated data that could be given as input to prediction schemes and the corresponding data collection methods that are available in practice. While simulated data is available instantaneously for all network elements, in reality transport systems do not collect all data, do not transmit it in real-time (e.g. post-processing of AFC), or have limited coverage (e.g. traffic counts from selected streets, APC from a subset of the fleet).

4. Provision and adaptation

The impact of RTI is ultimately manifested by reducing travel uncertainty and enabling travellers to take more informed decisions that will result with travel time savings. In order to capture such potential impacts, travellers' progress has to be modelled as a dynamic choice process which allows the incorporation of RTI in their en-route decisions. For example, travellers may adapt their departure time, travel mode or access stop based on pre-trip information and reconsider what line to board or where to transfer based on the RTI availability along the journey. However, the consideration of RTI does not imply that travellers rely solely on it in its presence. Instead, travellers may adapt their travel decisions (e.g. initial stop, boarding line, interchange stop) based on their travel characteristics (e.g. trip purpose,

scheduling constraints), their familiarity with the service, past experience with alternative travel paths and their perception of travel information quality.

The within-day loop yields experienced travel attributes per individual traveller which are accumulated in traveller's travel memory and will influence traveller's perception of the service and thereof future expectations and decisions. In a repetitive choice context, travellers are able to construct perceptions concerning service uncertainty and address them in their travel plans, including trip departure time. Moreover, service users can also appreciate the performance of RTI systems although only based on their directly experienced travel attributes8.

The perception of service and RTI reliability evolve at the individual traveller level based on a day-today learning procedure which determines how different information sources influence travellers' expectations and decisions. The reliability of the RTI is evaluated based on the comparison of provisioned value versus the corresponding experienced value. Past experience of the RTI can therefore influence the credibility travelers' associate with this information source and hence the extent of travel adaptation. The expected travel attribute of traveller n with respect to network element j could be then conceptualized as:

j =yZalnj] ^{PK, RTI, X} (2)

Where ajn is the normalized credibility coefficient associated with information source 1 and xe^ is the

expected service attribute based on prior-knowledge, RTI or accumulated experience.

The RTI dissemination scheme determines its availability on a given travel decision. BusMezzo supports the specification of RTI availability for individual stops or vehicles to reflect the availability of on-site displays. In addition, the share of travellers that have access to RTI can be specified in order to represent an imperfect penetration rate of mobile devices. The availability of RTI is determined then by the conjunction of location and individual availability while its comprehensiveness may vary from local (e.g. waiting time at this stop), hub (e.g. waiting time for each stop within a certain distance) to network-wide.

5. Illustration

In order to demonstrate the potential value of the modelling framework outlined above, the Metro network in Stockholm, Sweden was used as a case study. The morning peak was simulated in detail, including 700 vehicle trips and 150,000 travellers. The case study considered two scenarios: (a) No-RTI, where all travellers rely solely on prior-knowledge; (b) Network-RTI, where RTI concerning the remaining waiting time is available to all transit travellers at all times. For each scenario, 10 simulation runs were performed with an execution time of less than 1 minute per simulation run on a standard PC.

Fig. 2. On-board occupancy on the southbound Metro Line 14 with and without RTI

As expected, the provision of RTI results with time savings. However, the provision of RTI may also have undesired effects as illustrated in Figure 2. Each curve corresponds to a chain of consecutive trips with scheduled departure times from the terminal between 7:30 and 8:00. Under the provision of RTI, the actual system conditions are subject to more fluctuations and passenger decisions deviate from expectations that are based on average values. The fluctuations in passenger loads increases the probability of on-board and at-stops discomfort and capacity concerns.

6. Discussion

The rapid development and deployment of RTI services requires tools for evaluating their performance and impacts. Transport prediction models have the potential to change profoundly both the way the transit system is operated (e.g. real-time bus scheduling, fleet management strategies, dynamic fare structures) as well as the way it is consumed (e.g. personalized information, multimodality with shared modes, crowd-sourcing). Conventional approaches lack the capability to represent system dynamics in the adequate level of detail required for analyzing the generation, dissemination and consumption of RTI.

BusMezzo, an agent-based simulation model of vehicles and travelers, facilitates the analysis of alternative prediction schemes as well as the impact of their provision on system performance. BusMezzo could therefore be used as a test-bed for RTI schemes and generate synthetic data for their validation. The behavioral components on the model need to be adequately calibrate and validated. The impact of RTI on travel decisions depend on information acquisition processes and degree of adaptation.

The modelling tool presented in this paper can assist policy makers and system operators in evaluating the impact of RTI provision, compare alternative RTI generation and dissemination strategies. Furthermore, it can help prioritizing investments for certain information services, locations of information displays or type of information.

Acknowledgements

The author thanks Zafeira Gkioulou for valuable contribution to the development of the learning scheme References

1. Nuzzolo A, Russo F, and Crisalli U. A doubly dynamic schedule-based assignment model for transit networks. Transportation Science, 2001; 35, (3); 268-285.

2. Nokel K, Wekeck S. Boarding and alighting in frequency-based transit assignment. Transportation Research Record, 2009; 2111; 60-67.

3. Wahba M. MILATRAS - Microsimulation learning-based approach to transit assignment. Doctoral Dissertation, University of Toronto, Canada, 2008.

4. Cats O, Kousopoulos HN, Burghout W, Toledo T. Effect of real-time transit information on dynamic path choice of passengers. Transportation Research Record, 2011; 2217: 46-54.

5. Toledo T, Cats O, Burghout W, Koutsopoulos HN. Mesoscopic simulation for transit operations. Transportation Research Part C, 2010; 18: 896-908.

6. Cats O. Multi-agent transit operations and assignment model. Procedía Computer Science, 2013; 19; 809-814.

7. Cats O, Loutos G. Real-time bus arrival information system - an empirical evaluation. Proceedings of the 16th International IEEE conference on Intelligent Transport Systems (ITSC), 2013;

8. Ettema D, Timmermans H. Costs of travel time uncertainty and benefits of travel time information: Conceptual model and numerical examples. Transportation Research Part C, 2006; 14; 335-350.