Demand for New York City Taxicabs in 2013: Exploring the Potential for Ride-sharing

Transcription

1 Demand for New York City Taxicabs in 2013: Exploring the Potential for Ride-sharing AJ Swoboda (corresponding author) Department of Operations Research and Financial Engineering Princeton University 229 Sherrerd Hall (ORFE Building) Princeton, NJ T: Alain Kornhauser, Ph.D. Professor, Department of Operations Research and Financial Engineering Princeton University 229 Sherrerd Hall (ORFE Building) Princeton, NJ T: October 1, 2015 Princeton Autonomous Vehicle Engineering (PAVE ) Word Count: 5,135 Number of Figures: 8

2 ABSTRACT Over past decades, the automobile, its extensive infrastructure network, and their ability to together increase the capacity for personal rapid transportation have significantly contributed to the development of society. Yet the sustained dominance of today s transportation system will continue on a path toward higher fuel costs, more foreign oil dependency, denser traffic congestion, and more. This paper acknowledges that, while others search for next-generation physical technologies to solve this modern transportation problem, an effective solution may be possible with current technology on existing infrastructure dynamic ridesharing. Using highly-detailed data of all Taxi & Limousine Commission trips taken in Manhattan and the surrounding boroughs of New York City during 2013 as a proxy for on-demand, personal, door-to-door transportation demand of NYC, this research examines characteristics of the current system and investigates how various ridesharing policies could perform if they were implemented as an enhancement to the existing system. The data revealed three basic characteristics major findings resulted from analysis into the existing data. 1) Detailed precisely and as expected, demand fluctuated daily (peaking twice around 6 AM and 7 PM), weekly (peaking on Fridays and shifting to later hours during the weekend), and annually (dropping during the summer months and the winter holiday). 2) Surprisingly, NYC TLC demand did not correlate with precipitation levels throughout the year. 3) Roughly 90% of NYC TLC taxicabs recorded at least one fare on an average day, however very rarely did half of the taxicab fleet generate revenue at any given moment in the week, with only 32% doing so on average. Since the data included the precise origin, origin time, and destination of each individual trip, it was possible to analyze the vehicle use and corresponding societal implications if various ridesharing policies had been in effect in 2013 and if trip makers had responded positively to those ridesharing opportunities. Ridesharing origins were required to fall within the same 0.1-by-0.1 square mile to classify as a ridesharing opportunity. Although some ridesharing opportunities were discovered from some places at some times, it was surprising how rare they were. Under the most stringent of policies (little flexibility in terms of correlation of origin, origin time, and destination), where requirements included the origin times fitting within a 30 second time window and the destinations fitting within a 0.3-by-0.3 mile drop off point, few ridesharing opportunities existed. Had those trips been shared, the reduction of annual vehicle miles would have been a paltry 0.3%. To have achieved a 23% reduction in annual vehicle miles throughout 2013, NYC TLC would have needed to implement a ridesharing system that could include a wait time of up to five minutes, a common destination limit of up to five, and a drop-off zone measuring 0.5-by-0.5 miles. While the City-wide ridesharing opportunities were not very impressive, those opportunities were all concentrated from a few origin locations, namely the City's major transportation hubs. From those locations, substantial ridesharing opportunities existed. Not analyzed were ridesharing opportunities associated with pick-up-along-the-way.

3 INTRODUCTION The personal automobile has served to offer unparalleled mobility over the past decades and it's overwhelming superior service quality makes it the most popular transportation mode of choice throughout America and the rest of the world. In fact today, the automobile and its extensive network of roads indelibly defines transportation in the United States. Unfortunately, the evolution of this type of personal mobility has inflicted substantial societal costs. Since many individuals use vehicles as the lone driver (1), a large number of personal miles traveled (PMT) consume an equivalent amount of vehicle miles traveled (VMT). The societal costs associated with transportation energy consumption, dependence on foreign oil, increased pollution, congestion, etc. are more closely tied to the VMT metric than to PMT, so it is imperative that next generation transportation systems focus on minimizing VMT to diminish the negative externalities associated with transportation. Dynamic ridesharing, paired with Internet enabled technologies and comprehensive algorithms, can provide such a solution without requiring a complete upheaval of the already-extensive transportation infrastructure (2). In a sense, this is what conventional mass transit is about: serve a significantly higher number of PMTs without increasing the total VMT. However conventional mass transit does not work in a sufficient number of places. The fundamental reason mass transit does not work more ubiquitously is not due to the public's aversion to travel with strangers, but rather due to either of the following facts: 1) that individual trips are not sufficiently correlated in both space and time enough to allow for the opportunity of traveling together; or 2) that even if that correlation does exist, there is no possibility for strangers to utilize the same vehicle. This conundrum results in people tending to travel alone or with closely related individuals, even when inconvenient. Setting aside the question about people's willingness to travel with strangers, is PMT sufficiently correlated such that ridesharing could occur if a system was set in place to notify travelers of that possibility? This is a very difficult question for many reasons, least of which is that there is not sufficiently detailed and comprehensive database of trips to empirically test such a propensity. However trips within a relatively closed system can provide the basis for such a test. Trips taken by taxicab in NYC are essentially captive to cabs they are usually too long to walk, their passengers do not have access to personal automobiles, and other motivations incline the passengers to not use the bus or subway systems. In most cases there is not transportation mode choice, however they may be a dynamic ridesharing potential should the opportunity be made apparent. The New York City Taxi and Limousine Commission (NYC TLC) released detailed logs of every taxicab trip taken during the 2013 calendar year (3). The 2015 paper New York City Taxicab Transportation Demand Modeling for the Analysis of Ridesharing and Autonomous Taxi Systems by AJ Swoboda set out to empirically determine the relationship between trip correlation (the level of service offered) and ridesharing potential, hoping to address the extent to which there is rideshare potential amongst people who use cabs to satisfy their travel need. A 2013 paper written by Talal Mufti laid the groundwork for synthetic transportation analysis. Through simulating all trips taken in New Jersey by drawing on aggregate data from the 2010 U.S. Census, the 2010 American Community Survey, and other sources, Mufti was able to model how ridesharing and the use of autonomous taxis would work in the state of New Jersey (5). Alexander Wyrough, in the 2014 paper A National Disaggregate Transportation Demand Model for the Analysis of Autonomous Taxi Systems, improved and expanded on Mufti s model, as well as extended it to the entire United States (6). By exploring the nature and feasibility of a ridesharing model built off of actual real-life data very similarly structured to the theoretical, synthesized trip data of (5) and (6) the research results could lead to a more

4 realistic, and therefore more useful, understanding of the potential for a ridesharing framework throughout the United States. A thorough discussion of the background, data processing, and activity of the NYC TLC is included in (4), as well as more detailed descriptions of the process and breakdown of the ridesharing algorithms used during simulation and its connections and implications related to national ridesharing. This paper, a condensed version of (4), focuses on the main highlights of taxicab activity and the key findings generated through automated ridesharing models. TAXICAB PICKUP ACTIVITY ACROSS 2013 The demand for taxicabs in New York City can be most accurately represented by the distribution of taxicab pickups that occur. The day-by-day visualization of taxicab pickup frequency in the top of Figure 1 provides a general understanding for the ebb and flow of taxicab demand over the course of an entire calendar year. Taxicab usage in 2013 was affected by seasonality it experienced a drop during the warmer, summer months and the winter holiday and it fluctuated more sharply day-by-day over the course of a week. Debunked more thoroughly on page 28 of (4), most of the overall trend s irregularities existed due to lack of passenger demand during major holidays, while the sharpest decline in activity (in early August) resulted from a lack of drivers due to out-of-the-ordinary traffic closures. No immediate correlation could be found between precipitation levels in this Northeastern city and taxicab demand. To account for the weekly trending patterns explained in Figure 1, this weather analysis was also conducted for each specific day of the week where, again, no significant relationship was found. TAXICAB PICKUP ACTIVITY BY WEEK Perhaps a more enlightening insight is derived from the examination of an average 24-hour day of NYC taxicab trips. However individuals routines differ based on the day of the week, so rather than inspecting an average day, the entire 2013 year was compressed into an average week to unveil characteristics of demand that would be indiscernible at a larger scale. Figure 2 explores the frequency of taxicab pickups, separated by NYC borough, during each minute of the week from 12:00 AM on Monday to 11:59 PM on Sunday Manhattan accounted for 90.3% of all taxicab originations over the course of 2013 and followed a logical pattern of demand as one traces through the week considering to- and from-work travel as well as nightlife and weekend activities. Queens almost-constant pickup demand from 6 AM through midnight of each day was heavily influenced by the presence of JFK and LGA airports, which together alone made up 3.5% of all NYC TLC 2013 taxicab pickups. The remaining 6.2% of all activity was

5 broken down amongst the other four boroughs as the following: Brooklyn (3.1%); the rest of Queens (1.5%); the Bronx (0.9%); and Staten Island (0.8%). Further explanations begin on page 21 of (4). FLEET ANALYSES Understanding how the existing fleet of taxicabs in New York City conducts business and serves inhabitants allows for the discovery of inefficiencies within the transportation system. On an average day in 2013, 90.8% of the total taxicab fleet (12,480 medallions) recorded at least one trip, indicating that there was a cause for not all taxicabs to be able to serve customers each day. One possible explanation would be due to vehicles being serviced or repaired. The ebb and flow of the bottom of FIGURE 1 Annual Snapshot: Manhattan Rainfall; NYC TLC Daily Demand; NYC TLC Fleet Activity Proportion

6 FIGURE 2 Per-Minute Trip Demand during 2013 by Weekday Figure 1 can be compared to the oscillation in the top plot in Figure 1 to understand whether a shortage in taxicab activity correlated more closely with a decrease in demand or in supply. However perhaps a more insightful approach to understanding taxicab activity on a standard day is to consider how many taxicabs are generating revenue at a given moment during a day. For such an analysis a standard week with no extreme events was selected: Monday, March 4th through the end of Sunday, March 10th (as can be verified by observing this week-long span in both plots of Figure 1). For each minute in this selected week, every taxicab in the process of transporting a passenger from pickup to drop-off was tallied. Figure 3 displays the proportion of the taxicab fleet that actively produced revenue at any given moment during the week. Although ridesharing solutions addressed in Figure 3 Proportion of Taxicab Fleet Generating Revenue this paper do not specifically address manners in which taxicabs can be generating revenue more consistently, this figure is very telling of the inefficiency of the current NYC TLC system. Very rarely in a standard week was half of the taxicab fleet generating revenue at a given minute, with only 32% of the fleet transporting a customer on average. It is infeasible to expect human drivers to complete constant back-to-back fares without taking personal breaks. Additionally, these independently-thinking

7 drivers do not necessarily know the optimal repositioning location to find a new customer upon the termination of a taxicab fare. A connected network of vehicles has the potential to tell a different story. PIXELIZATION OF NEW YORK CITY GEOGRAPHIC AREA This paper discretized the 2013 NYC TLC data into a grid of 84,970 adjacent pixels in order to reduce the complexity of a ridesharing analysis. The original location components of the data involved specific latitude-longitude (lat-long) coordinates of each trip s pickup and drop-off. The coordinate system s accuracy is usually an asset, but due to its continuous nature it would have resulted in essentially an infinite number of lat-long combinations throughout the entire region of analysis covering the 850-square-mile area of land surrounding New York City s five boroughs. Each pixel was measured to be 0.1-by-0.1 miles in size, a small enough assignment so as to not sacrifice the high level of detail included in the data. Further technical procedures for the dividing and tracking of the data can be found in Chapter 4 of (4). It is worth noting that the majority of NYC TLC pickups and drop-offs were concentrated in Manhattan, while the drop-offs over the year were more uniformly dispersed amongst the other boroughs. 3D heat maps of the frequency of pickups and drop-offs by NYC TLC cabs over the course of 2013 can be found in Chapter 4.2 of (4). These 3D visualizations use the same scale and can appropriately be compared side by side. As mentioned earlier in this report, LaGuardia Airport and John F. Kennedy International Airport possessed high quantities of taxicab traffic relative to other non- Manhattan areas of New York City. Although it may prove difficult to implement a broad-sweeping ridesharing system overnight, introducing the public to such a program at specific hot spots would be a feasible approach towards achieving an overarching goal of ubiquitous ridesharing. Airports, popular train stations, and city destinations would likely prove to be efficient starting points for such a next-generation transportation system. Chapter 5.4 of (4) gives contextual information surrounding the most popular pickup and dropoff locations in NYC during A more detailed analysis revealed that the top-five pickup locations all possessed an almost-constant flow of taxi demand throughout the course of an average 24-hour day. It makes sense that the most popular locations in New York City for taxicab pick-ups consistently contain individuals looking for a ride, but this could not possibly be the case for every pixel location in the city s five boroughs. And for an effective ridesharing program to exist, it must focus on the locations with sufficient volume of demand in order to successfully match individuals together in dynamic rideshares. Upon further inspection of the almost 85,000 total pixels in the geographic area, it was found that 2,249 of the pixels made up 99 percent of all taxicab pickups over 2013 and that just 294 pixels accounted for 50 percent. RIDESHARING: METHODOLOGY, POLICIES, AND RESULTS Dynamic ridesharing can provide a solution to today s overwhelmed and inefficient transportation system, which struggles with issues of pollution, congestion, and more. As mentioned earlier, in order to quantify the degree to which a ridesharing transportation system can mitigate these negative byproducts, the single most complete metric to consider is vehicle miles traveled. Ridesharing Methodology It will be valuable to first address some major conceptual assumptions and explanations of terms used throughout the remainder of this paper.

8 1) This ridesharing analysis was approached pixel-by-pixel, rather than using discrete lat-long coordinates, meaning each passenger s departure was identified by a 0.1-by-0.1 square mile. When calculating distances the research assumed each departure occurs from the pixel s centroid, and thus it may require an individual to walk a small distance to reach the pickup. 2) A ridesharing policy refers to an overarching, theoretical algorithm that determines how a rideshare match can be identified. Whereas a framework identifies the specific geographicboundary and time-window combination of a given policy. Geographic boundaries refer to the relative sizes used when determining if individuals were heading to the same destination. A superpixel is a 9-pixel square and a macropixel is a 25 pixel grid, translating to squares that are 0.3-by-0.3 and 0.5-by-0.5 miles in size. A time window refers to the amount of time that a taxicab with one passenger would wait for another passenger to show up with a rideshare match before departing. For this research, the time window ranged from 30 seconds to 5 minutes. Eight distinct frameworks were chosen for analysis, and evaluated under each policy, but there is no limit to the number that could be studied. 3) Average departure occupancy (ADO), percent reduction in taxicab usage (%TaxiRed), and percent vehicle miles saved (%vmiles) are three important metrics to quantify the effectiveness of ridesharing implementations. ADO represents the number of passengers involved in an average taxicab departure. %TaxiRed refers to the percent fewer taxicabs summoned for trips over the course of the year, while still satisfying the entire demand for taxis. Most important, %vmiles represents the percent of vehicle miles that would not have been driven if the ridesharing policy had been implemented in NYC during ) To prepare the trip files, the 2013 NYC TLC data was sorted chronologically by the trip origination date and time and each lat-long location was assigned the corresponding pixel identification. Once this preparation was complete, the simulation began with the following logic. An unobserved individual (rider-a) enters a taxicab and a timer begins, set to the length of whatever the pre-determined time window is. Any other individual who arrives at the same pixel within rider-a s time window and who s trip satisfies a ridesharing policy s matching requirements is recorded as joining rider-a s departure. Once the time window passes, rider-a and all of the other matching riders are logged as being in a single departure and set off toward their destination (or destinations). The methodology for determining whether a potential rideshare satisfies the ridesharing policy s requirements will be explained later, but note that it varies based on the policy parameters. The top picture of Figure on page 44 of (4) gives a sample for how the algorithm s logic searches and iterates through the dataset of taxicab trips this sample framework consists of a megapixel (MP) geographic boundary and a time window of 300 seconds. The bottom half of Figure in (4) presents what would be the finalized departure list of the trip files from the top half. 5) The term common destination (CD) refers to the number of destinations a ridesharing taxicab would be driving to before all passengers are dropped off. If CD = 1, then all trips within the rideshare departure are going to the same destination (determined by the geographic boundary size). For CD = x where x > 1, a departure would travel to x unique destinations. In order to take into account the potential inconvenience added to any member of a departure, an additional destination would only be added to the existing departure if it added no more than a specific percentage of distance to the existing total trip distance this metric is referred to as circuity (CIR). An evolution of this CIR metric is another metric called directional circuity (dcir), which eliminates the possibility of adding any other trip with a destination the opposite direction as the first destination by creating a dividing line that splits the geographical area into two sides. The orientation of this split is either North-South or East-West depending on the geographical relation between Origin and Destination #1. The process, including diagrams, is included in more detail in Chapter (page 65) in (4).

9 6) To best understand the feasibility of a ridesharing implementation in NYC for taxicabs, there must be a well-defined scope of operation the service level of the ridesharing program. A clear choice was to conduct analysis for a service level consisting of the 2,249 pixels that represented 99 percent of all taxicab activity in For other points of comparison, 50 pixels were hand selected, which represented all major entrance and exit points of major transportation hubs in NYC Penn Station, NY Port Authority Bus Terminal, Grand Central Station, LGA, and JFK. A figure on page 53 of (4) gives the exact geographical location of each of the selected pixels overlaid onto a map. Lastly, service level considerations were made consisting of the pixels contained in the top- 5, top-100, and top-1000 groupings once pixels were sorted by taxicab pickup activity in descending order. These were added to the analysis to give a snapshot as to how effective a given policy and framework may be as a ridesharing program would expand from initial implementation to more widespread usage throughout the city. Ridesharing Policies and Results: (CD = 1) This subsection focuses on a (CD = 1) ridesharing policy and considers the benefits and effectiveness of implementing it over various service levels for eight frameworks a superpixel or macropixel geographic boundary combined with time window lasting 30, 60, 120, or 300 seconds. Python code included in Appendix B.5.1 of (4) shows how the ridesharing simulation was conducted, with the code in B.5.2 consisting of the subsequent analysis of the rideshare departure pairings. As implied, this policy is concerned only with a single common destination between all passengers in the rideshare. Assuming that trips B and C fit within the specified time window of trip A, the left-hand side of Figure 5 gives a visual example of what a (CD = 1) departure assignment could consist of. A departure s total length can be computed as the length of the longest trip in the departure list, since a vehicle would need to travel no further than that destination. So, the number of vehicle miles saved can be determined by subtracting the distance of the rideshare from the sum of the individual, original trips. With this example scenario, the number of vehicle miles saved is the sum of the trip A and trip C distances. Table 1 (CD = 1) Ridesharing Policy at Two Service Levels

10 Table 1 shows the results of the analysis after running a simulation with the (CD = 1) ridesharing Figure 5 Example Logic of (CD = 1) and (CD = 3, dcir) Ridesharing Policies policy. When comparing the percentage values from the major transportation hubs and all NYC pixels, it is important to note that the transportation hub trips account for roughly 10% of all NYC trips. By comparing relative %vmiles and %TaxiRed metrics in the table, it can be concluded that implementing a (CD = 1) ridesharing program would be roughly two times as effective at just the major transportation hubs than if it were implemented throughout NYC. The plots consisting of Figures 8(a) and 8(b) are cumulative percentage reduction plots, displaying the total reduction in vehicle miles as a percentage of original vehicle miles given an increasing number of pixels added to the overall ridesharing system. Furthermore, the addition of pixels along the x-axis is in descending order of individual-pixel vehicle mile reduction, meaning the first pixel on the x-axis is the most productive at reducing departure vehicle miles. So to repeat for clarity, one reads these plots by visualizing how the pixel-by-pixel increases in service level of the ridesharing policy corresponds to a decrease in %vmiles. The value at the tail end of each curve aligns with the associated value listed under All NYC Pixels: %vmiles in Table 1. Figure 6(c) presents ridesharing system efficiency in a different manner by plotting effectiveness curves of differently-sized service levels. Each effectiveness curve displays the degree to which each individual rideshare framework reduced total vehicle miles. The x-axis of this plot is a discrete ordering of rideshare frameworks sorted by increasing inconvenience to the passenger it was assumed that it is less convenient to wait longer before departing and that this is more severe of an inconvenience than being included in a destination consisting of a slightly larger geographic area. This figure highlights how the (CD = 1) policy places more value on the geographic size of a destination than on the time window, given by how the undulating effectiveness curves always dip when switching to a superpixel framework with the next-largest time window.

11 Figure 6 (CD = 1) Results Ridesharing Policies and Results: Research Evolution The results presented in the previous section begin to shed light on the overall effectiveness of implementing a ridesharing program within the NYC TLC service. However they are difficult to appreciate without a comparison to other ridesharing policy results. (4) provides much further detail surrounding the evolution and progression of various ridesharing policies. To save time and space, only the major lessons learned will be mentioned here. Increasing a ridesharing policy from (CD = 1) to (CD = 3, CIR = 0.2) brought forward a significant increase in effectiveness, particularly with shorter time windows. The 30-second time window frameworks witnessed a %vmiles increase by an average factor of 13, while the effectiveness of the 300-second time window frameworks increased by a factor of 8. This large jump makes sense, as the policy s algorithm has more wiggle-room to find optimal rideshare combinations when more common destinations exist per departure. Taking directionality into policy consideration shifting from (CD = 3, CIR = 0.2) to (CD = 3, dcir = 0.2) improved %vmiles results, but only slightly. However this result is still promising, as the

12 application of direction circuity would continue to enhance the overall ridesharing system effectiveness while also greatly pleasing passengers. A possible explanation for this increase could be that the previous algorithm would have added trips to a certain departure as long as the circuity constraint was not surpassed, even if a better assignment could have existed. Directional circuity logic helps ensure each trip addition is the best assignment. Reducing the inconvenience level experienced by passengers from (CD = 3, dcir = 0.2) to (CD = 3, dcir = 0.1) resulted in a corresponding drop in performance. It makes sense that a more stringent policy experienced underperformance across all frameworks (especially the ones with less lenient parameters). Further research would need to be conducted to gain insight into the emotional difference for a rider experiencing dcir = 0.1 as opposed to dcir = 0.2; this would allow for a better understanding of the significance of the drop in policy effectiveness. Lastly, the exploration of (CD = 5, dcir = 0.1) and (CD = 5, dcir = 0.2) produced interesting findings when compared with the results explained above. The addition of 2 more common destinations resulted in a jump in policy effectiveness while the greater dcir constraint allowed for even greater performance. An intriguing difference between these policies and their CD = 3 counterparts was that the shorter time-windows experienced much smaller of an effectiveness increase. A likely reason being because, on average, there are simply not enough passengers to make a dramatic difference in filling all of the departure constraints if the time windows are small. It is also important to note that the performance of (CD = 5, dcir = 0.1) exceeded that of (CD = 5, dcir = 0.1), leading to the conclusion that a plausible best practice for NYC taxicabs could consist of a policy with a high number of common destinations and a low circuity ratio. Ridesharing Policies and Results: (CD = 5, dcir = 0.2) Table 2 (CD = 5, dcir = 0.2) Ridesharing Policy at Two Service Levels

13 Figure 7 (CD = 5, dcir = 0.2) Results Although the previous subsection traced through the important findings from all of the result sections in (4), this subsection exists to give a quantitative and visual comparison between a basic ridesharing policy (CD = 1) and a complex ridesharing policy (CD = 5, dcir = 0.2). Table 2 and Figure 7 show how the most inclusive framework in the best ridesharing policy would have witnessed a 23 percent reduction in vehicle miles for all taxicab trips over the course of Although any double digit reduction in %vmiles and %TaxiRed when considering the total miles driven by all taxicabs over the course of the year would be helpful in solving the issues presented earlier in this paper, it is an underwhelming result. It is particularly underwhelming when considering some of the assumptions made to reach it: all individuals would share a ride if they could; all individuals would be willing to wait up to five minutes before departing; all individuals would be willing to potentially walk additional short distances to their destination. CONCLUSION

14 The United States is in desperate need of a reduction in vehicle miles traveled. Dynamic ridesharing appears to be a potential solution that effectively minimizes the amount of infrastructural and technological change required. Yet developing an effective ridesharing analysis across the United States to effectively test such a concept be it from real or synthesized data is a major undertaking. This research serves to provide a piece of analysis grounded in the real world with an eye toward answering questions about New York City and also be used as a basis of extrapolation to larger, national models. By applying ridesharing concepts to the 2013 NYC TLC data set, this research unveiled flaws in the traditional taxicab system and identified characteristics of a ridesharing program that could reduce the sheer quantity of vehicle miles driven in order for society to function. With bold assumptions made during ridesharing simulations, the reductions in VMT (while leaving PMT unaffected) were not nearly promising enough to warrant the implementation of such a system, even in as dense of a city as New York City. These uninspired results still raise further questions when considering such a system for the city. If a cost-effective ridesharing system were implemented, would more inhabitants hail taxicabs and therefore increase the system s ability to function? What would the success of a ridesharing program do to the other forms of public transportation in the city? How dense would a city need to become in order to allow for dynamic ridesharing to prove extremely effective at reducing total personal vehicle miles? Nevertheless, these underwhelming findings associated with the 2013 NYC TLC system prove to be troubling news for the future ridesharing networks installed throughout the United States, where personal automobile travel becomes significantly more complex. REFERENCES (1) Federal Highway Administration. Summary of Travel Trends: 2009 National Household Travel Survey. In: U.S. Department of Transportation (2009) (2) A. Di Febbraro, E. Gattorna, and N. Sacco. Optimizing Dynamic Ride-Sharing Systems. In: Transportation Research Record: Journal of the Transportation Research Board 2359 (2013), pp (3) Chris Whong. FOILing NYC s Taxi Trip Data. Mar url: (4) Swoboda, AJ. New York City Taxicab Transportation Demand Modeling for the Analysis of Ridesharing and Autonomous Taxi Systems. Princeton University, June url: NYC_aTaxiRideSharingPotential2015.pdf (5) Talal R. Mufti. Synthesis of Spatially and Temporally Disaggregate Person Trip Demand: Application for a Typical New Jersey Weekday. Princeton University Master s Thesis, url: (6) Alexander Hill Wyrough. A National Disaggregate Transportation Demand Model for the Analysis of Autonomous Taxi Systems. Princeton University Senior Thesis, url: Wyrough_NoationalDisaggregateTripSynthesizer.pdf