3. Artificial Intelligence Supporting Sustainable and Individual Mobility: Development of an Algorithm for Mobility Planning and Choice of Means of Transport
Mobility planning is rarely individually tailored. Instead people have to make an active effort to adapt standard solutions to their requirements. Routing apps like Google Maps allow for personalization only by saving important places like home and a workplace but do not allow the user to influence the routing suggestions or choice of mode of transport based on preferences, limitations, or situation. It becomes even more difficult when different means of transport are to interact since most routing applications offer very little multimodal optimization aside from the last mile. Thus, the objective of this article is to present a concept for the utilization of artificial intelligence and regression models in order to enable individual and sustainable mobility planning. To achieve this objective, initially existing routing and mobility planning applications are examined and are conceptually expanded in order to outlay the benefits of personalized route planning. The concrete objective alongside with a method for the development of a new solution is summarized. An algorithm fulfilling these objectives based on multiple linear regression is conceptualized. Relevant factors with coefficient are identified, as well as necessary data sources and interfaces. This algorithm is then implemented in a limited prototype as a proof of concept. Finally, this prototype is tested based on a set of mobility scenarios in order to validate the achievement of the defined objective.
3.1 Introduction
Routing apps have a major impact on the choice of mode of transport for most people—especially for non-routine trips and, due to varying conditions and contexts, even for routine trips. Thus, they play a major role in spreading the adoption of sustainable modes of transport. They shall therefore be the starting point of this article: first we take a short look at the state of the art and then we draw a vision of what routing apps could offer. This serves as a guideline for the development of a corresponding method and algorithm for route planning.
3.1.1 Routing Apps: What They Provide Today
Google Maps is one of the most important services offered by Google (Brandt, 2016). It provides two essential functions: searching routes for different modes of transport and navigation. When looking for a route for an upcoming trip, Google Maps compares a trip by car, public transport (bus and train), walking, cycling, and flight and sometimes includes local offers such as cabs or electric scooters. Users can find out estimated travel times and distances and, in case of multimodal connections, the transfers. If the user wants to travel by car, by bicycle, or by foot, Google Maps provides detailed navigation. It also takes into account current traffic or closures. Since Google evaluates the cell phone data of all users for its route optimization, the navigation is widely considered to be precise, especially with regard to delays and obstacles. A recent confirmation for this has been provided (La Rocco, 2016; Leicht, 2018). The main competitors of Google Maps are Apple Maps, Here WeGo, ReachNow, and OpenStreetMap. The functionality of all these applications and their restrictions in regard to personalization is very similar.
Advertisement
3.1.2 A Vision for Routing Apps: Individually Tailored, Sustainable Mobility
None of these existing routing apps allow the user to influence process or optimization parameters. The user may only enter a few travel parameters: date, time, and the choice of means of transport. These limited parameters significantly restrict the individualization of travel planning in exchange for a simplistic user experience. For example, there is no way to include physical restrictions, presence or absence of privately owned vehicles, mobility subscriptions, or simply an aversion to or preference for a means of transport. There are no situation-dependent parameters either, such as weather conditions, luggage, or fellow passengers. However, all these parameters strongly influence the optimal travel option. For example, if the user needs to utilize the travel time for work on a long trip, traveling by train is more advantageous than it would be otherwise. The number of people traveling also has a strong impact on the pros and cons of a means of transport.
Parameters that influence the decision for routing and means of transport are countless. So far, they are not sufficiently covered in any existing application. However, existing technologies could already enable the consideration of person- and situation-dependent parameters. Moreover, artificial intelligence and a self-learning algorithm could be able to learn from past decisions and behavior of each user and independently reflect these results in a profile, which could be considered for future route planning. The advantages thereof can be attributed to three areas:
1.
Significantly increased comfort for the traveler: By taking into account the personal preferences and situational conditions, recommendations can be made on an individual level. As a result, route planning adapts to the traveler.
2.
Less effort for the traveler: Until now, a traveler has to independently compare different routes and means of transport (combinations) through different applications and websites. A new route planning application that includes individual context and independently is able to learn can make things easier for the traveler, since it automatically recommends exactly the most suitable option.
3.
Multicriteria optimization is possible: Until now, optimization according to a single criterion was the rule with routing apps. Their use was associated with a high level of manual effort on the part of travelers when comparing routes and means of transport. At the same time, they only provided little information. The routing apps (or the users themselves) thus tried to find a route or connection that fulfilled one criterion, e.g., as cheap or as fast as possible. However, if the optimization is automated by a new routing app and sufficient information is available, multicriteria optimization is possible. Then an app could, among other factors, simultaneously optimize for sustainability, costs, and travel duration (Scholz, 2018).
Thus, traveling can be more comfortable; the process of planning a trip is less complex, and more sustainable modes of transport are more accessible and useful. This vision of a routing app will be realized, albeit initially in a limited state, in a prototypical routing app called “EmiLa.”
3.2 Objective
The goal of the future application and therefore the algorithm is to popularize low-emission mobility for a wide range of users by establishing it as an optimization criterion without disregarding other factors. Giving emissions, among other factors, an adequate weight in the decision-making process is intended to nudge the users toward more environmentally friendly means of transport. Thereby, reducing mobility-induced emissions, a contribution can be made to reduce the impact of climate change.
Advertisement
The algorithm needs to be able to process all these types of information and based on them find an optimal connection, possibly multimodal, on an optimal route for a requested change of location from A to B, possibly including intermediate stops. The objective of the optimization is the fulfillment of a main goal, while secondary goals must be reached and conditions must be fulfilled. For the more limited first prototype of the app EmiLa, the main goal is to minimize the emissions of the journey, while secondary objectives are the shortest possible travel time, the lowest possible travel costs, and taking into account the personal profile and situation-dependent parameters (weather, as few transfers as possible, etc.).
3.3 Development of the Algorithm for Personalized-Quantified Routing Including Self-Learning Units
Based on the set objective, this chapter will elaborate on the fundamental idea and structure behind the algorithm as well as the process of data integration necessary in order for the algorithm to give informed recommendations.
3.3.1 Concept and Structure of the Algorithm
The requirement of optimizing based on numerous variables without a common metric represents the key challenge for the intended algorithm. First of all, each singular criterion needs a quantified scale standardized to the same dimensions. Additionally, the weighting of the criteria needs to be incorporated in an adaptable way in order to personalize the decision-making process. Also, it allows for a continuous learning process in the prioritization of optimizing a decision that is based on different factors with no existing common metric to assess them. This problem can be formulated similarly to a multiple linear regression, as described, for example, in Draper and Smith (1998):
y = estimated value of the dependent variable, i.e., overall score for a route and mode of transport. For the prototype application EmiLa, this is branded as “EmiQ” (emission quotient).
β0 = intercept on the y-axis, not relevant for this application as there is no baseline score.
β1 to βn = coefficient for each factor, i.e., the weight of the different factors included in the overall score. For the prototype application EmiLa, this is branded as “QuPeR” (quantified personalized routing).
x1 to xn = value of the independent variable, i.e., score for each singular factor (emissions, costs, travel time, etc.).
u = remaining error, not relevant for this application as this is not calculated.
Furthermore, the general model for linear regression needs to be extended for this use case (excluded factors from before are omitted):
p1 to pn = prohibitive coefficient, representing (individual) exclusion rules or strong aversions to a certain mode of transport. Examples for this are the exclusion of driving a car without a license or refusal to ride a bike in the snow. This application of multiple linear regression enables weighting the different factors through discrete individual coefficients. The metric and scaling of each factor have to be solved separately as only the result thereof can be processed as the x variable.
3.3.2 Metric and Scaling of the Factors
A central requirement for enabling the optimization of the chosen mode of transport based on the previously stated variety of factors is quantifying the degree of satisfaction for each factor for each available option. Thus, the factors need a common metric and a common scale. The length of the scale is negligible, so long as it is uniform across all factors. A grade system of 1 to 5 has been chosen for the prototype EmiLa, 1 representing the best possible score and 5 the worst possible score. An important requirement regarding the metric that arose during initial tests was to assess the factors relative to the set of available options. For example, the overall duration and the duration per kilometer vary strongly depending on the length of a trip. One unified scale for all trips would distort the score for this metric. For EmiLa, this has already been solved for the first three factors:
Duration of the trip: The fastest available option always receives the best score of 1, regardless of the absolute value overall or per kilometer. Even if an option may be slow by certain standards, it has to be adequately expressed that it is the fastest one available. Based on a small-scale user test, the worst possible score of 5 is given for any duration at least three times as long as the fastest possible option. In between those values is a linear scale. This can be summarized in the following resulting pseudo-formula for each option:
Cost of the trip: The assessment of the cost works very similar to the assessment of the travel duration. Unlike the duration, there are options that achieve a value of zero (e.g., walking), requiring an alternative calculation for distances that can be traveled by those means of transport. In these cases, the cost advantage of free options should be reflected with the best score of 1 without ignoring the differences among the other options, whereof the cheapest one receives a score of 3 and values at least three times as high receive a score of 5. Thus, the following two pseudo-formulas for each option have been created, the first without free travel options and the latter with free travel options:
Emissions caused by the trip: The scoring of the emissions has been designed differently from the costs and duration. A universal scale better fulfills the goal of measuring the adequacy of a mode of transport compared to a relative scaling. Especially for overseas trips, flights would receive a perfect emissions rating due to the lack of a more environmentally friendly option. Thus, a universal scale based on the current emission levels of the most widely used modes of transport has been defined, returning the best score of 1 for zero emissions and the worst score of 5 for emissions of 150 grams of CO2 equivalents per kilometer and above. This threshold for the worst score represents a relatively modern car with only one passenger. Unlike the other two factors, the scale is also separated into two parts from 0 to 30 and from 30 to 150 grams per kilometer. This serves the purpose of adding a slight advantage for modes of transport with very low emissions, thus reducing the elasticity of the scale in the lower range. Again, two pseudo-formulas result from this reasoning:
All three of these scales represent merely a starting point and are to be further adjusted based on more large-scale testing and studies. The approach of defining one metric per factor and then adjusting the scale between 1 and 5 based on user studies can be applied to any further factors as well, even qualitative ones with certain adjustments (e.g., comfort based on number of transits and a scoring model for different modes of transport). The prohibitive factors require less intricate scales, as they are only intended to eliminate certain modes of transport in certain conditions or favor them in others. Resulting scales are to be defined on a personal level though, primarily from preferences stored in a user profile. An exemplary factor may look as follows:
Thus, all options including walking in the rain automatically receive the worst possible overall score, if the user has made that choice in their profile.
3.3.3 Utilizing Machine Learning for Improving the Algorithm
The two main elements are the metrics for each factor x and the corresponding coefficients β. While the metrics of the factors are to be refined through empirical studies, the coefficients are intended to be improved continuously and user specific through machine learning. Only an initial calibration for the coefficients is predefined within the algorithm as that is necessary for any results to be calculated before there is any historical data of user choices to use for optimization. Afterward, the comparison between the recommended mode of transport and the one chosen by the user will serve to calculate vectors for each coefficient, similar to the method described in Tao et al. (2006), i.e., if the starting coefficients are βemission = 0.5, βduration = 0.25, and βcosts = 0.25 but most users opt for a more expensive faster option, the weight of βduration and βcosts needs to be adapted accordingly. The coefficient for emissions, βemission, will not be as dynamically adjusted according to user choices, as one of the main purposes is to encourage the wider usage of sustainable modes of transport whenever the drawbacks are within reasonable bounds.
As stated, the continuous optimization through vector-based machine learning can be conducted on an individual as well as on a global level. When nearly all users demonstrate different priorities in their choice of mode of transport than suggested by the algorithm, the global coefficients will be adjusted. If only an individual repeatedly demonstrates their divergent priorities though, these can be reflected as a user-based coefficient adjustment stored in the user database and retrieved whenever the specific user sends a request.
3.3.4 Application of the Algorithm in EmiLa
The first implementation of the previously elaborated concept is a prototype of the aforementioned routing application EmiLa. It serves as a proof of concept for the developed algorithm and also allows the adjustment of the predefined factors and metrics based on a wide sample of real-world user tests.
The routing application EmiLa compares the available means of transport, routes, connections, and combinations thereof in order to find the one that represents the optimum in relation to the multitude of defined factors. In the current first stage, the prototype only includes the comparison of costs, duration, and emissions. It can be expanded further based on the algorithm described in the previous paragraphs. To achieve this, the necessary data on the factors needs to be obtained from third parties (e.g., mobility service providers).
Personal mobility-related data stored in a user profile to derive prohibitive coefficients is not yet incorporated in the first prototype due to the current lack of user profiles. This will be the next addition when the prototype will be further developed. Those prohibitive coefficients are comparatively simple though, so they were not of primary interest for the first tests. Decisions previously taken within the application are not used yet either for the same reason. Before gathering adequate amounts of data through large-scale testing, the vector-based machine learning would not only have been very limited; it also implies a high risk of overfitting the model (Tao et al., 2006).
3.3.5 Data Integration into the Application
Once the presented algorithm is implemented in the prototype, the required data sources will be integrated for the various decision criteria, i.e., the factors emissions, costs, and travel duration for the first prototype of EmiLa. That data is retrieved from third parties, such as different regional, national, and international mobility service providers, via an application programming interface (API). Figure 3.2 illustrates this process.
Fig. 3.2
Integration of third party data through API
×
EmiLa retrieves the information from the interfaces of third-party providers, which is then processed within the EmiLa application for route planning. Data on the emission per kilometer of the different available vehicles is stored in an internal database. EmiLa evaluates the resulting recommendations for a requested route by the results of the EmiQ algorithm. Thus, the user can retrieve the result via any Internet-enabled terminal device in any web browser (Ott, 2018). In order to implement the algorithm beyond the first prototype stage, EmiLa primarily needs to integrate information from mobility service providers. Therefore, a wide variety of additional information such as weather data and occupancy information from parking garages will be integrated. Traffic data, availability of charging stations, and many further data sources are also beneficial. Apps for virtual meetings can be another interesting tool to integrate or recommend within EmiLa. This will avoid certain trips entirely and thus save costs, emissions, and travel time. A list of possible interfaces is given in Table 3.1.
Table 3.1
Possible API to implement for EmiLa
Provider
Content/data
Google Maps
• Navigation
• Travel time, distance, and routes by car, by bike, and by foot
• Travel time and routes by local public transport
• In certain areas, myTaxi and E-scooter-sharing options
OpenStreetMap
• Navigation
• Travel time, distance, and routes by car, by bike, and by foot
• Slope
• Fuel/energy consumption
• Road type and properties
HERE WeGo
• Navigation
• Travel time, distance, and routes by car, by bike, and by foot
• Travel time and routes by local public transport
• Travel time, distance, and routes by car, by bike, and by foot
• Travel time and routes by local public transport
• Payment
• Ride-hailing
• Carsharing.
• Availability, reservation, and payment of parking
• Charging infrastructure
Local transport associations, e.g., transport and tariff association, Stuttgart
• Travel time and routes by local public transport
Scooter-Sharing, e.g., lime, Voi, and tier
• Availability and cost of shared e-scooter
Carpool services, e.g., BlaBlaCar, MiFaZ, and Simply Hop
• Available carpool or passengers
Carsharing, e.g., Share Now, Stadtmobil, Flinkster, and local providers
• Availability and cost of shared cars
Bike-sharing, e.g., RegioRad, Smoove, Citybike, Call a Bike, Lidl-Bikes, Deezer, Nextbike, and local providers
• Availability and cost of shared (e-)bikes
Long-distance busses and trains, e.g., FlixBus and FlixTrain
• Travel time, distance, costs, and routes
Car parks, parking spots, e.g., at public transport hubs, Contipark, Deutsche Bahn, and GitHub
• Availability and cost of parking
Weather data providers, e.g., Deutscher Wetterdienst
• Current weather and forecast
Virtual meeting and conference services, e.g., GoToMeeting, Microsoft Teams, and Zoom
• Digital meetings to connect with customers, partners, colleagues, etc.
3.4 Testing of the Algorithm
In order to validate the practicality and usability of the algorithm and its recommendations within an application, the prototype of the routing app EmiLa has been developed and tested based on real-world scenarios. These first tests shall serve as a basis for possible adjustments to the included factors as well as the subsequent expansion of the algorithm and interfaces within the application in iterative loops.
3.4.1 EmiLa Testing Results
First tests were carried out for the travel planning of a selection of hypothetical domestic trips. For international trips, too few mobility service providers have been integrated in the app so far to allow for meaningful tests. Since public transport is shaped by national and regional providers, the implemented API of German providers are of little to no use for mobility planning in other countries. Tests will be conducted based on the following scenarios, representing common types of trips over varying distances and for different lengths of stay:
1.
Long-distance trip for one person between urban areas with one overnight stay (Munich-Berlin-Munich): arrival on Sunday evening and return on Monday evening.
2.
Long-distance trip for one person between a structurally weak region and an urban area without overnight stay (Wittenberg-Stuttgart-Wittenberg): arrival on Tuesday morning and return on Tuesday evening.
3.
Medium-distance trip for two people between two rural areas for a whole week (Calw-Radolfzell on Lake Constance-Calw): arrival on Monday morning and return on Friday evening.
4.
Short-distance trip for one person between a suburban and an urban area for a workday (Boeblingen-Stuttgart-Boeblingen): arrival on Tuesday morning and return on Tuesday afternoon.
5.
Short-distance trip for two people within a village (Dagersheim-Dagersheim): departure Saturday morning and return Saturday morning, about half an hour later.
Figure 3.3 shows an exemplary view of the options recommended by EmiLa as well as the preliminary user interface based on the first trip of the first scenario (departure from Munich to Berlin on a Sunday at 17:45): the best option according to EmiQ is a connection by bus, which takes longer than a connection by train but causes less emissions and costs only about one third compared to the train. The emission values are made comprehensible by a traffic light system since most users have little experience judging emissions based on grams per kilometer. The private car with one passenger scores worst with an overall EmiQ value of 3.67, which is the result of high emissions, the second longest travel time, and the highest costs.
Fig. 3.3
Route planning for Sunday, 27/09/2020, at 5:45 p.m. from Munich to Berlin
×
The resulting recommendations for all scenarios and trips are summarized in Table 3.2.
Table 3.2
Resulting recommendations for the testing scenarios (mode of transport for the longest distance is denoted)
Car (emissions high, costs 0,5 EUR, duration 0:02 hours)
These tests show that the application is generally capable of algorithmically comparing and evaluating travel options with respect to the defined parameters. Travel time and the costs were manually compared to different sources and have proven to be sufficiently accurate. Overall, the recommendations are comprehensible and useful: train, car, and bike are all recommended in scenarios where they represent the best balance between emissions, costs, and travel duration. In particular, the first rank for the car when traveling with two people between two rural areas reflects the particularities of a well-balanced choice mode of transport. No single mode of transport or in some cases even a car with an internal combustion engine is overall a good choice when public transport is too inefficient along a certain route.
Another observation based on these results is the distinct impact of personal preferences, especially in case of huge differences in duration and costs. An example thereof is the trip from Wittenberg to Stuttgart, where the second option is 27 EUR more expensive but 2 h faster. Depending on the individual situation, the order of the first and second recommendation might just as well be the other way around. Such individual calibrations to the recommendations are possible by individually adapting the coefficients of the factors defining their impact on the overall result on a user-specific level, as has been outlined in the third chapter.
3.5 Conclusions
The testing results underline the potential of the developed algorithm and application as a contribution to establishing widespread use of sustainable mobility. With no additional effort required from the user, a balanced recommendation for a mode of transport is made, taking into account emissions, costs, and travel duration. Thus, utilizing more sustainable mobility without relevant sacrifices in other areas becomes as simple as using any other existing routing app, providing cognitive-rational motives for the users. Additionally, emission-related challenges and gamification features will provide emotive motivation for developing more sustainable habits. Nonetheless, the choice is made by the user and not the app itself. This encourages further testing regarding the influence of the recommendations on the users’ decisions, i.e., the effectiveness of the application for encouraging the use of sustainable modes of transport as well as the preferred handling of conflicts between user preferences and rational, sustainable recommendations.
The prototype used for the previously presented testing of the algorithm is still a very limited implementation, especially in regard to individualization and user specificity. Further development of the application will be required in order to verify the practical viability regarding the inclusion of more decision criteria as well as the vector-based user-specific machine learning on the basis of Itskov (2019) and De Mello and Ponti (2018). Data security and privacy protection will become a more defining concern during the necessary storage and utilization of user data. Once these remaining functionalities have been implemented, a large-scale user test could provide the final confirmation of the developed algorithm.
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
Artificial Intelligence Supporting Sustainable and Individual Mobility: Development of an Algorithm for Mobility Planning and Choice of Means of Transport
