Human Satisfaction as the Ultimate Goal in Ridesharing
HHuman Satisfaction as the Ultimate Goal in Ridesharing
Chaya Levinger, Amos Azaria, Noam Hazon
Department of Computer Science, Ariel [email protected], [email protected], [email protected]
Abstract
Transportation services play a crucial part in the de-velopment of modern smart cities. In particular, on-demand ridesharing services, which group togetherpassengers with similar itineraries, are already op-erating in several metropolitan areas. These ser-vices can be of significant social and environmentalbenefit, by reducing travel costs, road congestionand co2 emissions. The deployment of autonomouscars in the near future will surely change the waypeople are traveling. It is even more promisingfor a ridesharing service, since it will be easierand cheaper for a company to handle a fleet of au-tonomous cars that can serve the demands of differ-ent passengers.We argue that user satisfaction should be the mainobjective when trying to find the best assignmentof passengers to vehicles and the determination oftheir routes. Moreover, the model of user satisfac-tion should be rich enough to capture the travelingdistance, cost, and other factors as well. We showthat it is more important to capture a rich model ofhuman satisfaction than peruse an optimal perfor-mance. That is, we developed a practical algorithmfor assigning passengers to vehicles, which outper-forms assignment algorithms that are optimal, butuse a simpler satisfaction model.To the best of our knowledge, this is the first pa-per to exclusively concentrate on a rich and realisticfunction of user satisfaction as the objective, whichis (arguably) the most important aspect to considerfor achieving a widespread adaption of ridesharingservices.
The National Household Travel Survey performed in the U.S.in 2009 [Santos et al. , 2011] revealed that approximately83.4% of all trips in the U.S. were in a private vehicle (otheroptions being public transportation, walking, etc.). The av-erage vehicle occupancy was only . when compensatingfor the number of passengers (i.e., if two people travel inthe same vehicale, their travel distance is multiplied by two). This extremely low average vehicle occupancy entails a verylarge number of vehicles on the road that collectively con-tribute to carbon dioxide emissions, fuel consumption, airpollution and an increase in traffic load, which in turn requiresadditional investment in enlarging the road infrastructure. Inrecent years, ride hailing services such as Uber and Lyft havegained popularity and an increasing number of passengers usethese services as one of their main means of transportation[Wallsten, 2015]. Both Uber and Lyft are now also offer-ing ride-sharing options, and other companies, such as Super-Shuttle and Via, are explicitly targeted at customers who wantto share their ride.The deployment of autonomous vehicles in the near futurewill have a significant impact on the way people are travel-ing. The implication of this revolutionary way of transporta-tion is not fully known nowadays [Guerra, 2016], but it issafe to claim that autonomous vehicles will have a positiveeffect on the development of ridesharing services. Indeed, itwill be easier and cheaper for a company to handle a fleet ofautonomous vehicles that can serve the demands of differentpassengers. It can also rule-out some negative human-driverfactors, such as driver’s fatigue from the long travels and thedriver’s inconvenience from having multiple pick-up and dropstops along his route.The basic challenge of a ridesharing service is how to as-sign the passengers’ requests for a ride to vehicles and de-fine the routes for the fleet of vehicles in an optimal manner.This problem belongs to the generic class of Vehicle Rout-ing and scheduling Problems (VRPs), which have been ex-tensively studied over the past years, mainly in the oper-ation research and transportation science communities. Sev-eral variants with different characteristics have been devel-oped. For example, the initial formulation of the VRP as-sumes that the environment is static , i.e., all requests areknown before-hand and do not change thereafter [Dantzigand Ramser, 1959]. The more complex variants, includ-ing the rideharing problem, are dynamic , where real-timerequests are gradually revealed along the service operatingtime. In this setting the assignment of passengers to vehi-cles and the determination of vehicles’ routes may be adjustedwhen they are already in transit [Psaraftis et al. , 2016; Shen et al. , 2016]. Arguably, a major criterion that characterizeseach variant of the VRP is the objective function. It is verycommon to consider objectives from the service provider’s a r X i v : . [ c s . M A ] J u l erspective, for example, minimizing the total distance trav-elled [Secomandi, 2000], minimizing the fleet size [Diana etal. , 2006; Secomandi and Margot, 2009], or maximizing theservice provider’s profit [Campbell and Savelsbergh, 2005;Parragh et al. , 2014]. However, as noted by Cordeau andLaporte [Cordeau and Laporte, 2003], one should be inter-ested not only in minimizing the operating costs for the ser-vice provider but also in maximizing the quality of the serviceand the user satisfaction.Many works integrate quality of service and user satisfac-tion considerations as additional constraints of the problem.For example, a time window restricts the waiting time a pas-senger is willing to face before being picked up [Jaw et al. ,1986; El-Sherbeny, 2010], and it is usually combined with abound on the maximum user ride time [Paquette et al. , 2009].In addition, there are several works that combine the afore-mentioned operational objectives with the objective of max-imizing the user satisfaction (or its antonym, minimizing theuser inconvenience). The common interpretation for user sat-isfaction is the minimization of the total user on-board (ride)time and the total user waiting time [Psaraftis, 1980], theextra riding time due to ride-sharing [Lin et al. , 2012], orthe amount of deviations from desired departure and arrivaltimes [Fu and Teply, 1999; Yang et al. , 2013]. However, tothe best of our knowledge, there are no works that exclusivelyfocus on maximizing a complex user satisfaction function,which captures the traveling distance, cost, and other factorsas well.We investigate a comprehensive human-centric approachfor the ridesharing problem. Our basic claim is that the usersatisfaction should be the main objective of the ridesharingservice. Moreover, the model of user satisfaction should berich enough to capture the complex interdependencies amongseveral factors. Therefore, we develop a method for maxi-mizing a complex user satisfaction function.One approach for handling a rich objective function isto treat its factors as multiple objectives. Indeed, thereare several methods in the literature on VRP for handlingmultiple objectives. The most common approach is to ag-gregate the objectives into a single weighted-sum objectivefunction [Molenbruch et al. , 2017a], and advanced utilitymodel may be used for modeling the interactions betweenthe objectives [Lehu´ed´e et al. , 2014]. Additional strate-gies define hierarchical objective function [Schilde et al. ,2011], or return a set of non-comparable solutions whichdo not weakly dominate each other [Parragh et al. , 2009;Molenbruch et al. , 2017b]. Since our rich objective func-tion models user satisfaction, we propose a different, human-centric, approach. Specifically, we investigate machine learn-ing methods for modeling the rich satisfaction function fromreal humans.Modeling human behaviour is not a easy task, and a theo-retical model might fail to accurately capture real human be-haviour. We therefore ran experiments with actual humansand build a deep learning based function to estimate usersatisfaction. We introduce Simsat, an algorithm for assign-ing passengers to vehicles while maximizing a complex usersatisfaction function as the objective, for the ridesharing lastmile variant [Cheng et al. , 2014] setting. We show that Sim- sat outperforms optimal assignment methods that use a sim-pler objective function, indicating that it is more important toobtain a richer model of user satisfaction, than improving theperformance of the assignment algorithm. We will briefly review the current literature on the broad classof Vehicle Routing and scheduling Problems (VRPs), to placeour ridesharing problem in an appropriate context. The VRPwas first introduced by [Dantzig and Ramser, 1959]. Thegrowing body of research on routing problem over the past years has led to the development of several research commu-nities, which sometimes denote the same problem types byvarious names. In particular, the traditional VRP and someof its extensions deal with finding an optimal set of routesfor a fleet of vehicles to traverse in order deliver or pickupsome goods to a given set of costumers. We refer to the com-prehensive survey of [Parragh et al. , 2008a] on this class ofproblems, which they denote by Vehicle Routing Problemswith Backhauls (VRPB). A more recent survey, that also de-fines a taxonomy to classify the various variants of VRP by criteria, is given by [Psaraftis et al. , 2016]. A secondclass of problems, that is denoted by Parragh et al. as Vehi-cle Routing Problems with Pickups and Deliveries (VRPPD),deal with all those problems where goods are transported be-tween pickup and delivery customers. We refer to the sur-vey of [Parragh et al. , 2008b] on this class of problems.One subclass of VRPPD compromises the dial-a-ride prob-lem (DARP), where the goods that are transported are pas-sengers with associated pickup and delivery points. As notedby [Cordeau and Laporte, 2003], the DARP is distinguishedfrom other problems in vehicle routing since transportationcost and user inconvenience must be weighed against eachother in order to provide an appropriate solution. Therefore,the DARP typically includes more quality constraints that aimat capturing the user’s inconvenience. We refer to a recentsurvey on DARP by [Molenbruch et al. , 2017a], which alsomakes this distinction.A domain closely related to ride-sharing is car-pooling. Inthis domain, ordinary drivers, may opt to take an additionalpassenger on their way to a shared destination. The commonsetting of car-pooling is within a long-term commitment be-tween people to travel together for a particular purpose, whereridesharing is focused on single, non-recurring trips. Indeed,several works investigated car-pooling that can be establishedon a short-notice, and they refer to this problem as rideshar-ing [Agatz et al. , 2012]. We stress that in our ridesharingproblem, similar to the DARP setting, there is a central or-ganization that owns the vehicles, and they thus do not havetheir own travel plans. Informally, the ridesharing problem consists of a weightedgraph, requests given by passengers, each with an origin anda destination that are both nodes in the graph, and a set ofvehicles, each with a given capacity. All the vehicles are as-sumed to be operated by a central entity. In this paper wefocus on the last mile variant [Cheng et al. , 2014] setting.n this variant it is assumed that all the passengers are po-sitioned at the same origin location (e.g. airport), where allthe vehicles are also located, and must be taken to their fi-nal destination. The problem requires assigning travel routes(on the graph) to vehicles, in order to satisfy these passengerrequests while optimizing a given objective function. In ourwork, we concentrate on the objective of maximizing the usersatisfaction function.Let n be the number of passengers (and thus the numberof requests). We assume that the service provider incurs afixed cost per minute of travel, M , that encapsulates the fulloperation cost including any desired revenue. For example,if the fuel, tolls and any maintenance costs are estimated at x dollars per minute of travel, and the service provider commitsto receiving only a certain percentage of the user payment (asits revenue), ρ , M equals x − ρ . The service provider is free todetermine how to distribute this cost among the passengers,and in section 3.2 we discuss the properties of this paymentfunction.Every user, u ∈ U , is assumed to have a primary traveltime t o ( u ) and distance d o ( u ) , which are the time and dis-tance, respectively, it would take the user to reach her desti-nation had she received a direct ride. Consequently, we definefor each user a primary travel cost c o ( u ) = M · t o ( u ) . Wewill also add a fixed constant to t o ( u ) that represents waitingtime of the passenger had she received a direct ride. Givenan assignment, P , and a user u , the actual travel time of theuser is denoted by t P ( u ) , the actual travel distance is denotedby d P ( u ) , the actual amount paid by the user is denoted by c P ( u ) , and the user satisfaction is denoted by s P ( u ) . Our definition of the objective of the ridesharing problem isto find an assignment P that maximizes the sum on all usersatisfaction, i.e., (cid:80) u ∈ U s P ( u ) (or, equivalently, the averagesatisfaction n (cid:80) u ∈ U s P ( u ) ). For simplicity, in the last mileproblem we assume that a passenger traveling alone has somebaseline satisfaction level (“neither satisfied nor disatisfied”)from the trip. Satisfaction factors may include any or all ofthe following: • Travel cost . c P ( u ) . • Actual travel time . t P ( u ) . The travel time may alsodepend on other parameters, for example, a user maycare more about travel time during the weekend. • Extra travel time . t P ( u ) − t o ( u ) . • Actual travel distance . d P ( u ) . • Extra travel distance . d P ( u ) − d o ( u ) • Total number of passengers . Users may rather travelalone. The more passengers a ride may have the moreinconvenient it may become to each of the passengers. • User’s seat . In every vehicle with a constant capacitysome seats may be preferred over others. For example,if a vehicle with seats reaches its full capacity, mostpeople prefer siting on the front seat rather than the mid-dle back seat. • Working status / Occupation . Unemployed passengersmay be willing to travel longer in exchange for a lowercost. • User demographic information . Depending on theuser demographic group (e.g. age, gender, annual in-come), users may care more about the other factors. Forexample, young people may not mind traveling longerif they save a few dollars, but people in their 40’s maybe more concerned about their time. Some of this in-formation may be extracted from an image of the user.Note that this factor allows to define a more personal-ized function, since different users might end up withdifferent satisfaction values for identical rides.
Given n passengers and their destinations, who travel in a sin-gle vechile, an assignment P determining the drop-off path(i.e. the order of which the passengers are dropped-off), thetime it takes to reach each destination under this assignment( t P ( u ) ), the total cost of the ride-shared trip, and the cost( c o ( u ) ) and travel time ( t o ( u ) ) that each of the passengerswould have encountered had they traveled alone, the pay-ment function determines how much each passenger must pay( C P ( u ) ) when all passengers share the ride. We define thefollowing axioms on the payment function:1. The aggregated payment from all passengers should ex-actly cover the cost of the trip.2. n passengers traveling to the same destination split thetrip cost equally.3. Given two passengers with the same distance from thesource, the passenger who is dropped off second shouldpay less than the passenger who was dropped off first.4. Given a passenger that is ”on the way” to another pas-senger, both passengers should pay strictly less thanwhat they would pay had they traveled alone (eventhough, both passengers do not travel any longer).The following payment schedule satisfies all the axiomsabove: Let α, β ∈ R + , we define the user’s gain (given anassignment P ) as: g P ( u ) = ( αt o ( u ) + βc o ( u )) − ( αt P ( u ) + βc P ( u )) (1)If we define the inconvenience of a user, i P ( u ) , as αt P ( u ) + βc P ( u ) (and similarly, i o ( u ) is αt o ( u ) + βc o ( u ) ), Equation1 can be simplified to g P ( u ) = i o ( u ) − i P ( u ) . The paymentschedule sets c P ( u ) such that all passengers traveling in thesame ride have the same g P ( u ) , and the sum of c P ( u ) equalsthe cost of the whole ride. Equation 1 implies a simplifiedview of human behavior with the single concept of “time ismoney”, and it further assumes a linear relation between thetwo. In order to develop a more realistic human satisfaction model,we use machine learning techniques based upon data col-lected from humans. To this end, we solicited humansubjects from Mechanical Turk to obtain satisfaction leveldata. Based on this data, we use deep learning to build asatisfaction model. .1 Data Collection
The subjects were first asked to provide the following per-sonal details: year of birth, gender and whether they wereemployed or unemployed.Our satisfaction model tries to predict the relative satisfac-tion, that is, how much a passenger traveling by shared-rideis more or less satisfied than the same passenger traveling ina private ride. However, asking users to provide their rela-tive satisfaction is unrealistic, and we thus split every travelscenario into two sub-parts. In the first part we asked thesubjects to determine their satisfaction level from a direct pri-vate ride to some destination. In the second part the sub-jects were asked to determine their satisfaction level from ashared ride to the same destination. Specifically, in the firstpart of each scenario we described a direct private ride witha given time (random number between 5 minutes and onehour) and price (dollar per minute). In the second part ofeach scenario we described a shared ride to the same des-tination, where we varied the travel time and cost. Traveltime of a shred ride can never be shorter than a direct pri-vate ride, and we thus uniformly sampled a number from { , . , . , . , . , . , . , , , } and multiplied it bythe direct private travel time. The cost of a shared ride shouldbe lower than the cost of a direct private ride. In the optimalsharing scenario, assuming a 4 passenger vehicle (excludingthe driver), there could be 4 passengers traveling to the samedestination; in this case the cost will be reduced by a factor of4. We thus divided the direct private ride’s cost by a numberuniformly sampled from { . , . , . , . , . , . , , , } .In addition, we randomly sampled the number of additionalpassengers from { , , } , and we randomly sampled theuser’s seat from { front passenger, middle back, right back,left back } . The subjects could choose one of seven satisfac-tion levels on a Likert scale, between very satisfied (7) to verydissatisfied (1) with the middle being ‘neither satisfied nordissatisfied’ (4).Each subject was asked a total of six travel scenarios (eachwith a private and a shared ride). In order to eliminate sub-jects that may be selecting satisfaction levels at random, weadded two sanity check scenarios. In these scenarios the costof the shared ride was more expensive than the private ride.Since it is unreasonable for a person to be more satisfied witha shared ride, being both longer and more expensive, we havedisqualified any subject who expressed her satisfaction in thisquestion to be higher than her satisfaction from the privateride of that scenario. subjects failed one of these sanitytests, and were removed from our analysis. subjects refused to answer the personal questions andwere eliminated from our analysis as well. Of the remaining subjects were females and were males. Their ageranged from to with an average of . and a medianof . were employed while were unemployed. Eachof the subjects had scenarios, resulting with data-points. The average satisfaction for a private trip was . and a shared ride was . . Using the collected data we consider deep learning basedmodels with a varying depth to find a good satisfaction model, Model Train-error Validation errorLinear regression 1.680 1.7341 hidden layer 1.621 1.7112 hidden layers 1.592
Table 1: Train and validation error for the different model depths.Values indicate the root mean squared error (RMSE) of each model. that is, a model that will accurately predict user satisfactionlevels of a new user, based on different features of the userand the user trip. We split the data into train, validation (dev)and test sets. We use mean squared error to measure the per-formance of each model. The neural network depth variedfrom 1 (linear regression) to 4 (3 hidden layers). Each hid-den layer consisted of neurons. We used early stopping[Prechelt, 1998], i.e., we used the validation set to determinewhen to stop training. Table 1 presents the results obtained byeach of these models. Since the 2-hidden layers model per-formed best, we use it as our satisfaction function, this modelachieved a root mean squared error of . on the test set.We set the satisfaction of a private ride to ‘neither satisfiednor dissatisfied’ (4). In this section we present a stochastic algorithm for the lastmile variant. We assume that there are a sufficiently largenumber of vehicles so that any request could be satisfied, andthat the capacity of each vehicle is passengers. We comparethe performance of our algorithm, in terms of user satisfac-tion, to an optimal algorithm that uses a simpler satisfactionfunction. We show that the algorithm outperforms the optimalalgorithm, and that emphasizes the importance of capturing arich model of human satisfaction. We now present, a practical algorithm for assigning passen-gers to vehicles with the objective of maximizing the sum onuser satisfaction: Simsat (Algorithm 1). Simsat runs Floyd-Warshell on the graph at the initialization, to obtain the mini-mal travel time between every two vertices. Simsat then runsits main procedure as follows. Simsat shuffles all passengersand assigns every passenger to the vehicle that maximizes thecurrent satisfaction sum. For computing the total satisfac-tion of a single vehicle (the SatFunc function in Algorithm 1),we use the nearest neighbor algorithm (the greedy approach)for ordering the passengers drop-offs (based upon the Floyd-Warsell matrix). The main procedure is repeated multipletimes ( n ) and the assignment that yields the maximal totalsatisfaction is selected. The complexity of Simsat is clearly O ( n ) . The number of times the main procedure is repeatedcan vary; the more times it is repeated the higher the expectedperformance. Therefore, Simsat is an any-time algorithm. We use the following method to obtain the optimal assign-ment. First, the algorithm runs Floyd-Warshell on the graph. lgorithm 1:
The Simsat algorithm
Input : A graph (Graph), with source vertexPassenger destinations list (Passengers),A satisfaction function that returns the total satisfactionof all passengers in a vehicle (SatFunc),
Result:
An assignment of all passengers to vehicles.Compute Floyd-Warshell on Graph;MaxSum := 0; for i:=0 to n do Shuffle Passengers;SatSum := 0;Clear CabList; for
CurrentPassenger : Passengers do MAX := -1; for
CurrentCab : CabList doif
CurrentCab not full then
CurrentSat := SatFunc(CurrentCab);Add CurrentPassenger to CurrentCab;SatWithCurPass = SatFunc(CurrentCab);Remove CurrentPassenger fromCurrentCab; if (SatWithCurPass - CurrentSat) largerthan MAX then OptimalCab := CurrentCab;MAX := SatWithCurPass -CurrentSat endendendif
MAX < then Add CurrentPassenger to newCab;Add newCab to CabList; else
Add CurrentPassenger to OptimalCab;SatSum += MAX; endendif (SatSum > MaxSum) then
MaxSum := SatSum;OptimalFullAssignment := CabList; endend
Figure 1: A graph created from a map of the city of Toulouse,France.
The algorithm then solves a coin-change problem ([Harris etal. , 2008, p. 171]) to obtain all possible ways to split the num-ber of passengers into vehicles. For example, when n = 10 ,we get { , , , } , { , , , , } , { , , } etc. For each split-ting option the algorithm iterates over all possible assign-ments (we explicitly handle multiple vehicles with the samenumber of passengers, since it does not matter which groupof passengers travels in which vehicle if they are in the samesize). For example, for a group of { , , , , , } , it first iter-ates over all assignments of 4 passengers (there are (cid:0) n (cid:1) suchassignments), then, recursively calls the assignment functionwith { , , , , } and the remaining passengers. The recur-sive call iterates over all possible assignments of three peopleto two vehicles and preforms a recursive call with the remain-ing vehicles and passengers. For each vehicle, the algorithmcomputes all possible options for dropping off its passengers(this is done once for each set of users), and, based upon theFloyd-Warshell matrix and the satisfaction model, selects themost efficient travel order. We considered two different types of graphs, a randomly gen-erated graph and a more realistic graph, the city of Toulouse,France . The random graph was created by placing ver-tices uniformly on the plane. We then randomly chose a pairof vertices, and connected them with an edge with a probabil-ity that is proportional to their distance. The weight of eachedge was determined by the air-distance multiplied by a ran-dom number (uniformly sampled) between 1 and 2, to modeltopological variance. The graph of the city of Toulouse ispresented in Figure 1. This graph includes the actual dis-tances between the different vertices. We cropped the graphto , vertices, by running Dijkstra algorithm starting atthe airport, sorting all vertices by their distance from the air-port, and removing all farther away vertices (including thosethat are unreachable). obtained from .5
6 7 8 9 10 A v e r a g e U s e r - S a t i s f a c t i o n Simsat Optimal Gain Cost
Figure 2: Average satisfaction for 6, 7, 8, 9 and 10 passengers whenusing a random graph.
Being a last mile problem, we set the origin vertex to be thesame for all passengers. That is, in the random graph we ran-domly generated an origin vertex. In the city of Toulouse thegraph includes the Toulouse-Blagnac airport, and it was usedas the origin vertex. The destination vertices were randomlysampled for every passenger using a uniform distribution overall vertices.In the payment schedule we set α to . and β to forthe gain function. α and β were set according to the averageU.S. wage. That is, average annual hours worked per workerin U.S. at 2016 was 1783, and the average annual income inthe U.S. per worker in 2016 was $31,099. Dividing the twowe get $17.5 per hour, or approximately $0.3 per minute. Weset the average speed to 60 kph, and the cost per km traveldistance to $1 . We tested 5 assignment algorithms:1. The optimal algorithm with the full satisfaction function(developed in section 4).2. Simsat with the full satisfaction function.3. The optimal algorithm with simpler satisfaction func-tions:(a) Travel cost only.(b) Travel time only.(c) Time and cost according to the payment function(that is, the gain function is used as a substitute foruser satisfaction).All the algorithms were evaluated with the complete satisfac-tion function developed in section 4, regardless of the func-tion actually used by the assignment algorithm. Figure 2 presents the average satisfaction for 6,7,8,9 and 10passengers when using a random graph, and Figure 3 presentsthe results for the city of Toulouse. The results were obtainedby averaging over samples of passenger destinations. Theresults for the optimal method using travel time only wereomitted, as it constantly yields an average user satisfaction of (since it assigns a private vehicle to each and every passen-ger). As depicted in both figures, our satisfaction oriented as-signment method (Simsat) obtains results that are quite close
6 7 8 9 10 A v e r a g e U s e r - S a t i s f a c t i o n Simsat Optimal Gain Cost
Figure 3: Average satisfaction for 6, 7, 8, 9 and 10 passengers whenusing a map of the city of Toulouse, France. to the optimal assignment. Simsat’s average satisfaction levelis much closer to the optimal assignment than that of the op-timal assignments using a simpler user-satisfaction model.These results indicate that it is more important to obtain aricher model of user satisfaction, than improving the perfor-mance of the assignment algorithm. That being said, we donot disregard the importance of improving the performance ofthe assignment algorithm and do intend to pursue additionalalgorithms that may perform better.
Ridesharing has a true potential for improving the qual-ity of life for many people [Cici et al. , 2013], and it ispart of the general concept of sharing economy that is be-ing evolved nowadays. However, despite both Uber andLyft offering ride-sharing options, not many users elect toshare their rides with additional passengers [Koebler, 2016;RSG, 2017]. Following the statement by Carnegie [Carnegie,1936, p. 37], ”There is only one way to get anybody to do any-thing. And that is by making the other person want to do it.”,we believe that the key ingredient required for a widespreadadaptation of ridesharing is to focus on user satisfaction.The importance of the paper lies in its being the first to ex-clusively concentrate on a rich and realistic function of usersatisfaction as the objective, which is (arguably) the most im-portant aspect to consider for achieving a widespread adap-tion of ridesharing services. We use deep learning to modeluser satisfaction based upon data collected from actual hu-man subjects. We present a satisfaction oriented assignmentmethod (Simsat), and show that it outperforms optimal as-signments using a simpler user-satisfaction model. These re-sults indicate that it is more important to obtain a richer modelof user satisfaction, than improving the performance of theassignment algorithm.In future work we intend to extend our model to the moregeneral ridesharing schenario, where people may have differ-ent origins. We also intend to build a game that will simulatean actual ride for the subjects; this should allow us to obtainmore exact satisfaction levels. This game could include addi-tional travel information such as the other passengers in thetrip, and allow the subject to select her seat when enteringa vehicle. Since users will be playing the game more thannce, the satisfaction model can be further improved by per-sonalization, taking into account user’s feedback on previousrounds.
References [Agatz et al. , 2012] Niels Agatz, Alan Erera, Martin Savelsbergh,and Xing Wang. Optimization for dynamic ride-sharing: A re-view.
European Journal of Operational Research , 223(2):295–303, 2012.[Campbell and Savelsbergh, 2005] Ann Melissa Campbell andMartin WP Savelsbergh. Decision support for consumer di-rect grocery initiatives.
Transportation Science , 39(3):313–327,2005.[Carnegie, 1936] Dale Carnegie.
How to win friends and influencepeople . Simon and Schuster, 1936.[Cheng et al. , 2014] Shih-Fen Cheng, Duc Thien Nguyen, andHoong Chuin Lau. Mechanisms for arranging ride sharing andfare splitting for last-mile travel demands. In
AAMAS , pages1505–1506, 2014.[Cici et al. , 2013] Blerim Cici, Athina Markopoulou, EnriqueFr´ıas-Mart´ınez, and Nikolaos Laoutaris. Quantifying the poten-tial of ride-sharing using call description records. In
Proceedingsof the 14th Workshop on Mobile Computing Systems and Appli-cations , page 17, 2013.[Cordeau and Laporte, 2003] Jean-Franc¸ois Cordeau and GilbertLaporte. A tabu search heuristic for the static multi-vehicle dial-a-ride problem.
Transportation Research Part B: Methodologi-cal , 37(6):579–594, 2003.[Dantzig and Ramser, 1959] G. B. Dantzig and J. H. Ramser. Thetruck dispatching problem.
Management Science , 6(1):80–91,1959.[Diana et al. , 2006] Marco Diana, Maged M Dessouky, and NanXia. A model for the fleet sizing of demand responsive trans-portation services with time windows.
Transportation ResearchPart B: Methodological , 40(8):651–666, 2006.[El-Sherbeny, 2010] Nasser A El-Sherbeny. Vehicle routing withtime windows: An overview of exact, heuristic and metaheuristicmethods.
Journal of King Saud University-Science , 22(3):123–131, 2010.[Fu and Teply, 1999] Liping Fu and Stan Teply. On-line andoff-line routing and scheduling of dial-a-ride paratransit vehi-cles.
Computer-Aided Civil and Infrastructure Engineering ,14(5):309–319, 1999.[Guerra, 2016] Erick Guerra. Planning for cars that drive them-selves: Metropolitan planning organizations, regional transporta-tion plans, and autonomous vehicles.
Journal of Planning Edu-cation and Research , 36(2):210–224, 2016.[Harris et al. , 2008] John Michael Harris, Jeffry L Hirst, andMichael J Mossinghoff.
Combinatorics and graph theory , vol-ume 2. Springer, 2008.[Jaw et al. , 1986] Jang-Jei Jaw, Amedeo R Odoni, Harilaos NPsaraftis, and Nigel HM Wilson. A heuristic algorithm forthe multi-vehicle advance request dial-a-ride problem with timewindows.
Transportation Research Part B: Methodological ,20(3):243–257, 1986.[Koebler, 2016] Jason Koebler. Why everyone hates uberpool?,2016. [Lehu´ed´e et al. , 2014] Fabien Lehu´ed´e, Renaud Masson, Sophie NParragh, Olivier P´eton, and Fabien Tricoire. A multi-criteria largeneighbourhood search for the transportation of disabled people.
Journal of the Operational Research Society , 65(7):983–1000,2014.[Lin et al. , 2012] Yeqian Lin, Wenquan Li, Feng Qiu, and He Xu.Research on optimization of vehicle routing problem for ride-sharing taxi.
Procedia-Social and Behavioral Sciences , 43:494–502, 2012.[Molenbruch et al. , 2017a] Yves Molenbruch, Kris Braekers, andAn Caris. Typology and literature review for dial-a-ride prob-lems.
Annals of Operations Research , 2017.[Molenbruch et al. , 2017b] Yves Molenbruch, Kris Braekers,An Caris, and Greet Vanden Berghe. Multi-directional lo-cal search for a bi-objective dial-a-ride problem in patienttransportation.
Computers & Operations Research , 77:58–71,2017.[Paquette et al. , 2009] Julie Paquette, Jean-Franc¸ois Cordeau, andGilbert Laporte. Quality of service in dial-a-ride operations.
Computers & Industrial Engineering , 56(4):1721–1734, 2009.[Parragh et al. , 2008a] Sophie N. Parragh, Karl F. Doerner, andRichard F. Hartl. A survey on pickup and delivery problems.part I: Transportation between customers and depot.
Journal f¨urBetriebswirtschaft , 58(1):21–51, 2008.[Parragh et al. , 2008b] Sophie N. Parragh, Karl F. Doerner, andRichard F. Hartl. A survey on pickup and delivery problems. partII: Transportation between pickup and delivery locations.
Jour-nal f¨ur Betriebswirtschaft , 58(1):81–117, 2008.[Parragh et al. , 2009] Sophie N Parragh, Karl F Doerner, Richard FHartl, and Xavier Gandibleux. A heuristic two-phase solutionapproach for the multi-objective dial-a-ride problem.
Networks ,54(4):227–242, 2009.[Parragh et al. , 2014] Sophie N Parragh, Jorge Pinho de Sousa, andBernardo Almada-Lobo. The dial-a-ride problem with split re-quests and profits.
Transportation Science , 49(2):311–334, 2014.[Prechelt, 1998] Lutz Prechelt. Automatic early stopping usingcross validation: quantifying the criteria.
Neural Networks ,11(4):761–767, 1998.[Psaraftis et al. , 2016] Harilaos N Psaraftis, Min Wen, and Chris-tos A Kontovas. Dynamic vehicle routing problems: Threedecades and counting.
Networks , 67(1):3–31, 2016.[Psaraftis, 1980] Harilaos N Psaraftis. A dynamic programmingsolution to the single vehicle many-to-many immediate requestdial-a-ride problem.
Transportation Science , 14(2):130–154,1980.[RSG, 2017] Rsg 2017 survey results: Driver earnings, satisfactionand demographics, 2017.[Santos et al. , 2011] Adella Santos, Nancy McGuckin,Hikari Yukiko Nakamoto, Danielle Gray, and Susan Liss.Summary of travel trends: 2009 national household travelsurvey. Technical report, 2011.[Schilde et al. , 2011] Michael Schilde, Karl F Doerner, andRichard F Hartl. Metaheuristics for the dynamic stochastic dial-a-ride problem with expected return transports.
Computers &operations research , 38(12):1719–1730, 2011.[Secomandi and Margot, 2009] Nicola Secomandi and Franc¸oisMargot. Reoptimization approaches for the vehicle-routing prob-lem with stochastic demands.
Operations research , 57(1):214–230, 2009.Secomandi, 2000] Nicola Secomandi. Comparing neuro-dynamicprogramming algorithms for the vehicle routing problem withstochastic demands.
Computers & Operations Research ,27(11):1201–1225, 2000.[Shen et al. , 2016] Bilong Shen, Yan Huang, and Ying Zhao. Dy-namic ridesharing.
SIGSPATIAL Special , 7(3):3–10, 2016.[Wallsten, 2015] Scott Wallsten. The competitive effects of thesharing economy: how is uber changing taxis?
Technology PolicyInstitute , 22, 2015.[Yang et al. , 2013] Jianjie Yang, Jin Li, Yun Chen, Xintong Liu,et al. Multi-objective distribution model and algorithm for on-line shopping express logistics.