Resilience and Reliability

D. Gama Dessavre, G. Gongora, A. Garcia Tapia, L. Perez Estrada, A. Gilgur, A. Zavala Martinez, J. Manfredonia, D. Brandao, F. Capela

Dr. Jose E. Ramirez-MarquezSchool of Systems and Enterprises

Gabriela N. Gongora Svartzman - Urban Transport Systems ResilienceModeling and Visualizing Reliability In an Urban Bicycle Sharing Program. The Case of NYC

15

Reliability Model

Error)A)No)available)spaces)to)return)bicycle.

No)available)bicycles)to)take.

Error)B)

15

Reliability Model

Error)A)No)available)spaces)to)return)bicycle.

No)available)bicycles)to)take.

Error)B)

23

Results WM33(ST(and(7th(ST((Manhattan)

21

Reliability Model

Reliability model formulated from the occurrence probability of two errors. Error A: no available spaces to return. Error B: No available bicycle to take.

Using the behavior per station and simulating the demand according to Uniform, Triangular and Normal Distributions the probability of occurrence of each error is calculated per hour of the day during Weekdays and Weekends.

Figure 5: Results transferring to other stations

indicating when it would be a lower commute timeto be worth to change stations. The simulation alsoconsiders that the delay in one station affects the up-coming stations, i.e. there is a cumulative delay effectcarried through the stations.

From the previous image the outcome looks posi-tive, but when the passengers are considered i.e. theymove stations thereby the average number of passen-gers waiting at each station increases, the result is notthat positive. Figure 5 shows these results. The in-crease in passengers at each station (considering theyall go to one station or the other one) is very drasticfor a single station to hold in real life and have capac-ity for all of them on the upcoming trains. Which inturn would mean higher commuting times.

6 CONCLUSIONS AND FUTURE WORK

The simulation showed that delays on a single sta-tion can have a catastrophic effect on average wait-ing times, and this is only one subway line and onestation problem. If all stations throughout the systemwere tested for this kind of delays the effects wouldspread out the same way and the amount of passen-gers to handle would not be bearable.

It is important to have in mind how much these de-lays can affect daily passengers. A 10 minute delaycan seem insignificant but when it happens repeatedlythroughout the day it is clear that the effect can evenmean hours for the daily commuter.

In terms of progressive delays, it is establishedfrom the simulation that they do play an importantrole in average waiting times for passengers and spe-cially in the amount of passengers that will accumu-late in a platform. In the worst case scenario this canbe up to 2 extra minutes of waiting, but in the big pic-ture, if there are 10 or more stations ahead of this one,2 minutes will propagate through the stations creatinga delay of at least 20 minutes for the last station. Thisis of major concern to the MTA and it would be inter-esting to have real-time data to verify that their newpolicies on platform controllers can achieve these best

case scenarios.The model developed might have some limitations

but can simulate the subway system in NYC by sta-tion and line propagation. The contributions of thiswork are in the simulation itself. Figuring and fittingthe correct probability distributions, determining thevariables for the model and setting the conditions tosimulate many different disruption scenarios.

Future work can include exploring other disrup-tions, either micro-events (e.g. a game at the Mets-Willis Stadium) or the enforcement of new policies(size of platforms, carbon tax policies or redistribu-tion of budgets for repairs).

The simulation model already states variables forhandling capacities, and future work would be to ob-tain more data about train capacities and behavior atdifferent hours of the day to experiment how manypassengers can really get into the trains and how willthe waiting and commute times be affected.

REFERENCES

Alsger, A. A., M. Mesbah, L. Ferreira, & H. Safi (2015). Useof smart card fare data to estimate public transport origin–destination matrix. Transportation Research Record: Journalof the Transportation Research Board (2535), 88–96.

Barry, J., R. Newhouser, A. Rahbee, & S. Sayeda (2002). Originand destination estimation in new york city with automatedfare system data. Transportation Research Record: Journalof the Transportation Research Board (1817), 183–187.

Carey, M. & S. Carville (2000). Testing schedule performanceand reliability for train stations. Journal of the OperationalResearch Society 51(6), 666–682.

Gordon, J., H. Koutsopoulos, N. Wilson, & J. Attanucci (2013).Automated inference of linked transit journeys in london us-ing fare-transaction and vehicle location data. TransportationResearch Record: Journal of the Transportation ResearchBoard (2343), 17–24.

Hussein, Z. S. & B. K. Abbas (2014). Simulation model to man-agement for trains movement. International Journal of Com-puter Applications 93(2).

MTA. (2016). Turnstile Data. http://web.mta.info/

developers/turnstile.html. [Online; accessedNovember-2016].

MTA, F. & F. M. T. Authority. (2016). Average WeekdaySubway Ridership. http://web.mta.info/nyct/facts/ridership/ridership_sub.htm. [Online; accessedNovember-2016].

Reddy, A., J. Kuhls, & A. Lu (2011). Measuring and controllingsubway fare evasion: improving safety and security at newyork city transit authority. Transportation Research Record:Journal of the Transportation Research Board (2216), 85–99.

Schmocker, J.-D., S. Cooper, & W. Adeney (2005). Metro ser-vice delay recovery: comparison of strategies and constraintsacross systems. Transportation Research Record: Journal ofthe Transportation Research Board (1930), 30–37.

Stasko, T., B. Levine, & A. Reddy (2016). A time-expanded net-work model of train-level subway ridership flows using ac-tual train movement data at new york city transit. In Trans-portation Research Board 95th Annual Meeting, Number 16-4090.

Zhao, J., A. Rahbee, & N. H. Wilson (2007). Estimating a railpassenger trip origin-destination matrix using automatic datacollection systems. Computer-Aided Civil and InfrastructureEngineering 22(5), 376–387.

3.2 Performance Measures

The main performance measure in this study will bethe average waiting time for the train, this measure-ment will be obtained after several simulation runs(the amount needed for a 95% confidence interval).The second performance measure will be the numberof passengers waiting at the platform, before a trainarrives. If the trains get delayed it is more likely thatmore passengers will arrive at the station.

4 MODELING APPROACH

4.1 Data Collection

Data collection was not an easy task for this and therewere different steps for it. The first step was gatheringarrival times of the subway lines, per station. Unfor-tunately, after examining the observed data (gatheredfrom a real-time feed from the MTA) and the sched-ules set on paper by the MTA, these two don’t match.Either the MTA has not posted an updated versionof their schedules on-line or the trains in reality dorun very different to the ideal times put in paper orsome stations has been undergoing severe schedulingchanges due to repairs. To have a more accurate viewof reality in this simulation the observed data will beused in the model to calculate the process generatorsused in the model. The data collected for the observedvalues comes from the MTA‘s real-time feed.

The information for the passengers at this stationwas also obtained through the MTA, as they haveopen turnstile data about the passengers going in andout of the station. This number can vary sometimes,as time slots for the turnstiles are wide and its hard toguess which passenger corresponds to each time slot.To make the model simpler the passenger arrivals willbe modeled as a triangular distribution.

4.2 Model Development

The problem stated in this work will be modeled andsolved through a Discrete Even Simulation. Figure 1shows the state diagram for the proposed simulation,and Table 1 explains the notation (terms and distri-butions) used in the model. All simulations are con-structed in python programming language, with thehelp of the “simpy” module.

In a general description, the model has the follow-ing steps:

1. Collect real-time data of the subway trains ar-riving at each station, and calculate the waitingtimes between trains.

2. Get the data for people entering the subway sta-tions at different times.

3. Determine a probability distribution that best fitseach type of data collected (in steps 1 and 2). The

Figure 1: State diagram describing simulation.

probability distributions must fit the behavior ob-served at each station.

4. Perform a goodness of fit test (chi-square in thiscase) to the previous step. If approved then theprobability distribution and its parameters werechosen correctly.

5. Use the probability distributions chosen as pro-cess generators within the discrete event sim-ulation. For example, a triangular distributionwill be used to sample the number of passengersarriving at the station‘s platform every minute,while the � distribution will help sample theinter-arrival times of trains arriving at each sta-tion.

6. Adjust parameters of probability distributionsaccording to times of day as well.

7. Use Figure 1 to construct the discrete event sim-ulation. The variables are stated in the diagram.The simulation will carry a clock and the timesfor arrival and departure of trains will be kept.The inter-arrival times and number of passengersarriving are produced by the process generators.The statistical counters aid in retaining the num-bers relevant or the statistics of the simulations.

8. To make the model more accurate, it will be run200 times first, and from these simulations sta-tistical values will be derived to determined thenumber of simulations required for a 95% con-fidence interval. If this confidence interval is notacquired then the 200 simulations will be incre-mented.

9. Once n is calculated, n number of simulationsare run. Notice that every single one of the n isdone with a different seed to make sure resultsare simulated and not always the same.

10. A general delay variable will be modified everyn number of simulations.

11. Results can be analyzed from previous steps.

Commuting Time Variations and Reliability of Subway Systems in Case of Disruptions. The Case Study of NYC

Commuting times for riders can be altered by small delays such as dwell times, up to massive events (e.g. hurricanes). This work proposes a micro-event insight into commuting time variations through Discrete Event Simulation. For validation purposes New York City, and in particular line 7 of the subway was used as a case study. Simulating and Visualizing Emergency Departments For Improved Waiting Times and Patients’ Decision Making ProcessThis work identifies bottlenecks in the ED processes, tests scenarios and provides decision making suggestions to ED’s policies through an Agent Based Simulation. The findings of the simulation were used to developed a visualization tool.

Economic Assessment of Disaster Impact on Urban Resilience

Andrea Garcia Tapia -

Cities are complex systems composed of socio-technical and socio-ecological subsystems. The main components of interest are:

• Political ( quality of institutions)• Services ( access to health and education)• Social ( safety nets and community cohesion)• Infrastructure (transport, energy , water)• Economic & Finance ( credit access , insurance)• Ecological ( air, water and soil quality, biodiversity)

The main objective of this research is to understand how disruptions in ecological, services and infrastructure subsystems affect the social, political and economic subsystems. In order to asses an economic impact metric that enables better decisions by disaster management policy makers.Fig 1. Layers of Networks in a city, based on Dicken (2011) and Meerow et al (2016)

Araceli Zavala - Strategic Configurations for Multi-echelon Supply Chain Resilience Networks Against Disruptions

Vendor

Central

Inter. 2Inter. 1 Inter. 3

Field 11

Field 12

Field 21

Field 22

Field 31

Field 32

(100,200,30) (125,50,30) (100,125,30) (125,175,30) (100,50,30) (125,100,30)

(75,250,30) (75,150,30)(75,300,30)

(75,700,90)

Lead time 180 days

Lead time 120 days

Lead time 120 days

Stock 174 Stock 46 Stock 58 Stock 77 Stock 27 Stock 48

Stock 42 Stock 87Stock 172

Stock 58

I = $74,200A = 0.6528 Vendor

Central

Inter. 2Inter. 1

Field 11

Field 12

Field 21

Field 22

Field 31

Field 32

(100,200,30) (125,50,30) (100,125,30) (125,175,30)(100,50,30) (125,100,30)

(75,250,30) (75,300,30)

(75,700,90)

Lead time 180 days

Lead time 120 days

Stock 86 Stock 26 Stock 56 Stock 74Stock 26 Stock 45

Stock 144 Stock 196

Stock 135

I = $77,075A > 0.95

Multi-echelon supply chains are sensitive to disruptions, so a rapid network configuration is essential to keep providing an optimal service at the lowest possible cost. This research focuses on possible network configurations and its economic impact due to disruptions. • The model demonstrates the post-disruption resiliency at each supply chain network node along with the investment

necessary to restore network operations. One of the objectives of the research is on analyzing how to reallocate inventory through the network when a node is no longer capable of providing any service.

• The Central location is the most critical in terms of the post-investment needed to restore the supply chain followed by the Intermediate location with the greater demand to satisfy.

• The fields that come from an Intermediate node are more critical and need more investment than fields that comes directly from the Central.

Alex Gilgur - Resilience Metric as the Inverse of Sensitivity of Stabilization Time to Size of impact

ModelRoot Definition Simulation

Experimental Study

!"#$

#$ = &' ∗ !"*+ = 1/&'

Understanding Emotions in Communities

Joe Manfrediona - Visualizing Relationships between Mood and Facial Expression• Our mood affects how we feel as well as how we

behave every day. • Most of the current studies focus on how facial

expression translates to mood, but not how our mood impacts our ability to perceive it in others.

• This research focuses on:(1) exploring the relationship between mood state and

its effects on how we interact with others and…(2) visualizing the clusters of people who behave

similarly in this regard.• It also crowd-sources data using a gamified web

app (above) to measure the effect playing a ‘game’ has on our mood.

• The research employs unsupervised machine learning techniques to discover and visualize mood subpopulations within the respondents.

Fernanda Capela - Emotion Classification and Visualization

Use of NLP tools, such as SentiWordNet, for developing a method to identify specific emotions in fragments of crowdsourced text. A term sense is more likely to transmit an emotion if their distance in the polarity/subjectivity graph is smaller.

Visualizing Emotions in Music’s Crowdsourced Interpretations

Finding Patterns in Pictures with Similar Emotional Classification

Which characteristics of an image are responsible to arise emotions in the viewer? And do similar emotions means similar characteristics? This research will:

• Explore image recognition tools for captioning and description of pictures;

• Use Natural Language Processing to identify the emotional weight of those images, according to their labels;

• Group pictures that have the same emotional results together to explore their similar features and identify patterns;

• Build a model to predict emotions in any picture, without the need for the labeling step.

Danilo Brandao - Emotion Detection from Visual SourcesImages are a reflection of one’s instantaneous state of mind. Every time we take a picture or draw something, we are embedding our personal emotional signature into that image. This research seeks answers for the following questions:

• Are we able to identify the emotions from a visual piece and teach a machine to recognize them in an automated fashion?

• What is the effect of these emotions in users’ online behavior?• What components of images shared in online social media are

responsible for triggering emotions in the viewer?

Crowd Computing and Human-based Computation

Luis E. Pérez Estrada - Human-based Computation for Solving Complex ProblemsOur work explores how to leverage human skill and intuition into problem-solving efforts, focused on hard problems that are difficult for a computer or a single person to solve, or problems whose restrictions can not easily be codified as an optimization function.

We utilize the Crowd Computing framework and game design principles to make interfaces that enable participants to contribute solutions to a given problem. We have tackled combinatorial optimization problems like the Robust Facility Location Problem, Scheduling problems, vector and graph Clustering and Refugee Aid Deployment Policy

Text Analytics: Topic and Narrative Visualization

A novel visual exploratory text analytic system called NarViz was presented in this works. It can help users rapidly view, explore, and analyze the topic structure and management that are part of a single te

Storytelling is an integral aspect of human perception of reality.

This work presents a tool to visualize the narrative structure of textual data. Event time lines are an effective way to present stories and provide context to an audience.

Extending the idea of time lines, the visual representation proposed in this work can be used to understand the story that a text is trying to tell, in an intuitive and efficient manner.

Dante Gama Dessavre - NarViz: Narrative Visualization System Propagation of Topics in the MediaSome topics are more important than others for the news maedia. The importance of a topic dictates the number of articles published related to it.

This work analyzes how those topics are chosen, based on topic modeling algorithms.

After collecting a big corpus of news articles, latent dirichlet allocation was run as the topic model.

The network was created joining the outlets, articles and topics. So each topic has a network of how it was spread.

Center, Hoboken / Stevens Institute of Technology Global Network