DE102020207080A1 - Method and device for creating a network design for a route-based transport system with artificial intelligence - Google Patents

Abstract

The invention relates to a computer-implemented method for creating a route network layout for an on-demand transport system based on a geographical map, in particular based on a predetermined city map of an urban area, with the following steps: - Providing (S11) a route network based on Layouts (Spre) trained GAN model with a data-based generator model (11) that is trained to generate route network layouts (Sgen) depending on a random vector (Z), and a data-based discriminator model (12) that is trained to to distinguish generated and existing route network layouts (Sgen, Spre) from one another; - Use (S2, S3, S12) of the GAN model to determine a route network layout for an existing geographical map.

Description

Technical area

The invention relates to an on-demand transport system, in particular for rapid passenger traffic (PRT: Personal Rapid Transit). In particular, this invention relates to a public transportation system with autonomous vehicles on a given route network.

Technical background

On-demand transport systems linked to the route network for passenger transport can provide a large number of autonomous, in particular rail-bound vehicles, which move along a specified route network, such as a rail network. The route network can have stops for boarding and alighting, which are connected to one another via corresponding connecting routes, if necessary with branches / switches.

Route networks can be represented by a graph with nodes and edges connecting the nodes, the graph being adapted to a geographic map so that the nodes and edges are geographically assigned to a city map with stops and connecting routes.

Various methods are known for optimizing the transport efficiency between the individual stops, which determine an optimal timetable for an existing route network. A forecast of the time-dependent demand is essentially taken into account.

However, a further aspect for optimizing a transport system is the construction of the route network, which can be adapted in particular depending on the need for transport capacities and the available transport capacity through the positioning of stops and connecting routes. The layout of the route network can be based on the possible locations for stops and the possibilities of connections between the stops.

A conventional approach to creating a route network for such a transport system consists in creating a suitable route network layout, executing a simulation and evaluating the resulting simulation result using expert knowledge. By iteratively changing the route network, i.e. by adapting the locations for stops and the courses for the connecting routes, the route network is optimized until the simulation result meets the requirements.

Disclosure of the invention

According to the invention, a method for creating a route network layout for an on-demand transport system based on a predetermined city map of an urban area according to claim 1 and a corresponding device according to the independent claim are provided.

Further refinements are given in the dependent claims.

• - Bereitstellen eines auf bestehenden Streckennetz-Layouts trainierten GAN-Modells mit einem datenbasierten Generatormodell, das trainiert ist, um Streckennetz-Layouts abhängig von einem vorgebbaren Zufallsvektor zu generieren, und einem datenbasierten Diskriminatormodell, das trainiert ist, um generierte und bestehende Streckennetz-Layouts voneinander zu unterscheiden; und
• - Verwenden des GAN-Modells zum Ermitteln eines Streckennetz-Layouts für eine vorgegebene geografische Karte.

According to a first aspect, a computer-implemented method for creating a route network layout for an on-demand transport system based on a geographical map, in particular based on a predetermined city map of an urban area, is provided; with the following steps:

• Provision of a GAN model trained on existing route network layouts with a data-based generator model that is trained to generate route network layouts depending on a predeterminable random vector, and a data-based discriminator model that is trained to identify generated and existing route network layouts from one another to distinguish; and
• - Using the GAN model to determine a route network layout for a given geographic map.

The above method enables a manual iterative process for determining a route network layout based on an agent-based simulation to be replaced by a method which directly generates a route network layout with good or optimized properties. For this purpose, a generative adversarial network (GAN) is used to train a data-based generator model to generate a route network layout for a geographic map, in particular a city map, depending on existing route network layouts based on different city maps. This makes it possible to determine a route network layout with optimized properties in a direct and automated approach.

With the help of an unconditioned GAN model, a route network layout can be created without explicit specifications, but only on the basis of existing route network data. Generative adversarial networks represent self-learning machine learning models. They have two trainable data-based models, in particular two neural networks, a data-based generator model and a data-based discriminator model. The generator model and the discriminator model are trained together by an adverse training, with the two networks are trained against each other.

The discriminator model is optimized in order to distinguish known route network layouts from a training data set from route network layouts that are generated by the generator model. In contrast, the generator model is trained in order to make the generated route network layouts less distinguishable from the known route network layouts of the training data set.

According to one embodiment, the route network layout can thus be created by generating a possible route network layout based on a city map and checking the generated route network layout using the trained discriminator model, with a possible route network layout being a suitable route network layout for the City map is provided when the discriminator model recognizes the possible route network layout as an existing route network layout.

It can be provided that the GAN model trained on existing route network layouts is provided with a generator model that is trained to generate route network layouts as a function of a random vector and specification information that includes at least a city map, and with a discriminator model which is designed to distinguish generated and existing route network layouts from one another as a function of respectively assigned default information; wherein the route network layout is generated with the aid of the trained generator model as a function of default information for the specified city map.

In such conditioned GANs, additional specifications, such as existing city maps and class labels, can be taken into account so that route network layouts can be generated specifically for the desired class by the generator model. In essence, the above method represents a possibility of using generative adversarial networks to provide a route network plan based on a given existing city map depending on specifications, since route network maps of other cities have already been created with the help of expert knowledge. As a result, a route network with optimal properties can be obtained in a direct and automated approach.

With the help of a GAN model, a route network plan can be created with and without additional specifications, so that the GAN model can create a new route network plan that contains the observed training data distribution based on logical data from the operation of the existing route network. Furthermore, a GAN can be provided, the route network plans with predetermined specifications, such as. B. environmental conditions or other requirements are taken into account. In this way, the expert knowledge can be trained in a trained route network model, which can create optimized route networks based on sufficient city maps.

• - Stadtinformationen, wie beispielsweise Bevölkerungszahl, Bevölkerungsdichte und/oder Zahl der Pendler,
• - eine Satellitenbildinformation, insbesondere in Verbindung mit extrahierten quantifizierbaren Merkmalen,
• - einer Topologie-Information der durch den Stadtplan abgedeckten Stadtfläche, die ein Höhenprofilm des Untergrunds angibt,
• - bestehende weitere öffentliche oder private alternative Transportmöglichkeiten nebst Fahrplaninformationen sowie
• - zulässige für das Streckennetz nutzbare Flächenbereiche.

In particular, the specification information can furthermore comprise one or more of the following specification information:

• - City information, such as population size, population density and / or number of commuters,
• - a satellite image information, in particular in connection with extracted quantifiable features,
• - topology information of the urban area covered by the city map, which indicates a height profile of the subsurface,
• - existing other public or private alternative transport options along with timetable information as well as
• - Permissible areas that can be used for the route network.

Provision can be made for additional boundary conditions to be specified in the generator model in order to determine a route network layout for the geographical map as a function of the boundary conditions.

In particular, fixed locations for stops can be specified as boundary conditions, in particular with the aid of a placeholder value that indicates whether a location of a stop should be defined or can be freely selected.

Furthermore, a boundary condition for creating the route network layout can be taken into account in that the discriminator model identifies route network layouts that do not meet the specified boundary condition as negative examples that are used to train the generator model.

• - Bereitstellen eines auf bestehenden Streckennetz-Layouts trainierten GAN-Modells mit einem datenbasierten Generatormodell, das trainiert ist, um Streckennetz-Layouts abhängig von einem Zufallsvektor zu generieren, und einem datenbasierten Diskriminatormodell, das trainiert ist, um generierte und bestehende Streckennetz-Layouts voneinander zu unterscheiden;
• - Verwenden des GAN-Modells zum Ermitteln eines Streckennetz-Layouts für eine geografische Karte.

According to a further aspect, a device, in particular a data processing device, is provided for creating a route network layout for an on-demand transport system based on a geographic map, in particular based on a predetermined city map of an urban area, the device being designed to:

• - Providing a GAN model trained on existing route network layouts with a data-based generator model that is trained to generate route network layouts depending on a random vector, and a data-based discriminator model that is trained to distinguish generated and existing route network layouts from one another ;
• - Using the GAN model to determine a route network layout for a geographic map.

Figure list

• 1 ein beispielhaftes Streckennetz eines öffentlichen Nahverkehrssystems für eine beispielhaft angegebene Stadt.
• 2 ein Blockdiagramm zur Veranschaulichung des Aufbaus eines GAN-Modells;
• 3 ein Flussdiagramm zur Veranschaulichung eines Verfahrens zur Verwendung des GAN-Modells zur Erstellung eines Streckennetz-Layouts;
• 4 ein Flussdiagramm zur Veranschaulichung eines weiteren Verfahrens zur Verwendung des GAN-Modells zur Erstellung eines Streckennetz-Layouts.

Embodiments are explained in more detail below with reference to the accompanying drawings. Show it:

• 1 an exemplary route network of a public transport system for an exemplary city.
• 2 a block diagram to illustrate the structure of a GAN model;
• 3 a flowchart to illustrate a method for using the GAN model to create a route network layout;
• 4th a flowchart to illustrate a further method for using the GAN model to create a route network layout.

Description of embodiments

In the following, methods are described for creating a route network for an on-demand transport system based on a geographic map, such as a city map. The on-demand transport system can in particular be a capsule transport system, a local public transport system or the like, in which transport vehicles can be requested individually at stops. The transport system can also use a fleet of autonomous vehicles, such as B. Robotaxis and the like, which move along a predetermined route network, in particular a rail network.

The route network provides fixed stops at which people can get into or out of the vehicles or capsules. Connection routes along which the transport vehicles move are provided between the stops. The route network is based on a set of nodes (stops), edges that represent the connections between the stops. Furthermore, the edges can also have branches that represent switching points / switches in the route network.

Each of these nodes, edges and branches can have properties or attributes, such as for the nodes (stops) maximum capacity of passengers to be carried, average capacity of passengers to be carried, a handling frequency of transport vehicles, maximum number of transport vehicles to be processed at the same time and for connecting routes one Transport vehicle density, a maximum speed and the like and, for branches, information that is dependent on the direction of travel (in the case of switches in a rail network, it is not possible to select the direction in every direction of travel).

To create a route network for any city map of an urban area, it is proposed to train existing knowledge in the form of route networks for existing urban areas in a data-based model and to use the expert knowledge contained therein for the creation of further route networks. This approach provides for using a GAN (Generative Adversarial Network) model that has a data-based generator model and a data-based discriminator model.

Such a GAN model is exemplified in 2 shown. In 2 is a schematic of the structure of the GAN model for training the data-based generator model 11th and the data-based discriminator model 12th shown.

The generator model 11th can be constructed as a neural network, in particular as a CNN (convolutional neural network). The data-based discriminator model 12th is with the generator model 11th coupled and is also constructed as a neural network, in particular as a CNN or DNN (Deep Neural Network).

To train the GAN model, the generator model receives 11th as the input variable a random vector with a predetermined number of random input values, possibly in the order of 100. Depending on the type of GAN model, unconditioned or conditioned GAN, further specification information can be added to the generator model 11th to be provided. The generator model 11th creates a route network layout from this in the form of an output of a graph, which is a discriminator model 12th is made available.

The discriminator model 12th is designed to alternately use the generator model 11th received route network layout S _gen and an existing route network layout S _pre . The discriminator model 12th is trained to distinguish the provided route network layouts S _gen , S _pre by assigning a label to them, that indicates whether the relevant route network layout S _gen , S _pre a by the generator model 11th generated route network layout S _{gen or a real existing route network layout S pre} , or one that already exists for a city map.

The data format for the route network layout is selected in such a way that it describes a graph showing the nodes, edges and branches of the edges. Attributes are assigned to each element of the graph. The branches can be represented as nodes that have a corresponding attribute that indicates that it is not a stop, but a branch. The nodes that represent stops (and / or branches) can be assigned stop properties, such as their geographical position in the city map, handling capacity of people and transport vehicles and the like. The edges, which represent connecting routes, can be assigned route properties, such as their geographical extent in the city map, the transport capacity (in transport vehicles / hour), the transport duration (average speed). For example, the route network layout can be stored as one or more adjacency lists and / or one or more adjacency matrices, it being possible to refer to the corresponding features of a node / an edge by means of an auxiliary structure.

During the workout, the generator model 11th and the discriminator model 12th trained together with an adversarial criterion. The discriminator model 11th is trained so that there are real route network layouts S _pre generated by the generator model from generated route network layouts S _gen 11th have been modeled. For this purpose, a set of real route network layouts S _{pre is} selected, and a comparable number is selected by the generator model 11th generated route network layouts, S _gen . Then the discriminator model 12th trained to convert these route network layouts into their correct categories e.g. B. to be classified using a stochastic gradient descent method.

Furthermore, the generator model 11th are trained to create route network layouts that suit the discriminator model 12th make it difficult, the generated route network layout S _gen than by the generator model 11th recognize generated route network layouts. For this purpose, a gradient signal is used in a manner known per se, which is obtained by back-propagating the error signal of the discriminator model 12th into the generator model 11th receives.

In practice, the training can be carried out iteratively, with between improving the discriminator model 12th and the generator model 11th is alternated. Thereby the discriminator model 12th , in particular to avoid overfitting, is usually not trained to convergence. At least initially it is for the discriminator model 12th very easy to differentiate between the generated and existing route network layouts. However, during the training of the entire GAN model, the generator model becomes 11th better to create route network layouts in such a way that they always worse with the discriminator model 12th must be distinguished from existing route network layouts. Thus, synthetic route network layouts can be generated by random vectors and these by the generator model 11th be propagated.

With an unconditioned GAN model, the route network layout can be trained independently of a city map. A method for creating a route network layout for an existing city map can be carried out here, as is shown in the flowchart of FIG 3 is shown.

The procedure of 3 uses the discriminator model 12th that has been trained in the manner previously described.

In step S1 a route network layout is generated for the existing city map using conventional methods.

In step S2 can use the trained discriminator model to generate the route network layout 12th to be checked. Since the discriminator model 12th has been trained to recognize features of existing route network layouts, a step S1 The generated route network layout can be checked to determine whether it can be distinguished from the route network layouts that already exist in practice, in particular that have proven themselves.

In step S3 it is checked whether the generated route network layout from the discriminator model 11th as is sufficiently similar to existing route network layouts. If this is the case (alternative: yes), the procedure in step S4 signal or output the suitability of the route network layout generated last, otherwise (alternative: no) the procedure is followed by step S1 continued and a new route network layout was generated.

The method described above can be carried out based on an unconditioned GAN model. The GAN model does not take into account the city map for which the route network layout is in step S1 has been created, but only evaluates the characteristics of the generated route network layout in and of itself.

The trained GAN model also enables a fully automated creation of a route network layout. This is done in conjunction with 2 described GAN model expanded to a conditioned GAN model. By using a conditional GAN, the generator model 11th trained in a comparable manner, but in addition to the random vector z, default information V is also supplied.

The generator model 11th and the discriminator model 12th in each case the specification information V, which has at least one city map, is provided, which allows a more targeted coordination of a by the generator model 11th generated route network layouts to the conditions. The generator model is used for this 11th trained to provide a generated route network layout depending on the random vector z and the specification information Vgen, which has a local reference to the city map of the specification information in question.

The existing route network layouts are then also provided with the corresponding default information Vgen.

• - Stadtinformationen, wie beispielsweise Bevölkerungszahl, Bevölkerungsdichte, Zahl der Pendler und dergleichen,
• - eine Satellitenbildinformation, insbesondere in Verbindung mit extrahierten quantifizierbaren Merkmalen, wie beispielsweise Reflektivität und dergleichen,
• - einer Topologie-Information der durch den Stadtplan abgedeckten Stadtfläche, die ein Höhenprofilm des Untergrunds angibt,
• - bestehende weitere öffentliche oder private alternative Transportmöglichkeiten nebst Fahrplaninformationen sowie
• - Vorgaben zu Flächennutzungen und dergleichen.

In addition to a city map, which can be provided in the form of image information or in some other way, the specification information Vgen can contain further specifications that can affect a configuration of a route network layout. This can include the following information:

• - City information, such as population size, population density, number of commuters and the like,
• - Satellite image information, in particular in connection with extracted quantifiable features, such as reflectivity and the like,
• - topology information of the urban area covered by the city map, which indicates a height profile of the subsurface,
• - existing other public or private alternative transport options along with timetable information as well as
• - Requirements for land use and the like.

• In dem Generatormodell 11 können zusätzlich Randbedingungen vorgegebenen werden, wie z.B. Länge des gesamten Streckennetzes, Prozentsatz von Einwohnern, die mit einem bestimmten Serviceniveau bedient werden sollen, wie beispielsweise einer maximalen Laufdistanz und einer maximalen Wartezeit auf verfügbare Transportfahrzeuge und dergleichen.

The generator model 11th can be trained in a corresponding way to generate a route network layout with the default information provided. The discriminator model 12th can be trained in a corresponding way to evaluate the generated and existing route network layout in connection with the default information. In this way, a complete route network layout for a transport system can be created for city maps and other default information, which is based on the experience of already existing route networks for other cities:

• In the generator model 11th In addition, boundary conditions can be specified, such as the length of the entire route network, percentage of residents who are to be served with a certain service level, such as a maximum walking distance and a maximum waiting time for available transport vehicles and the like.

Furthermore, fixed locations for stops can be specified as boundary conditions. Fixed locations for stops can be used, for example, with the help of a placeholder value that indicates whether a location of a stop should be fixed or can be freely selected. Furthermore, network-free zones in the city map can be specified through which the network should not be routed.

Boundary conditions for creating the route network layout can be taken into account by using the discriminator model 12th Solutions that do not meet the given boundary conditions are identified as negative examples. Otherwise, this can also be done iteratively by generating multiple times or by post-processing on the basis of the generator model 11th provided results are provided.

In 4th is a flowchart illustrating an automatic method for creating a route network layout.

In step S11 becomes the generator model after training the conditional GAN model 11th provided.

In step S12 a random vector and the specification information, which includes a city map, are fed to the generator model.

In step S13 is obtained as the output of the generator model 11th a route network layout for the given city map.

If the route network layout has been created using one of the methods described above, a corresponding route network can be created in an existing geographical area.

Claims (10)

Computer-implemented method for creating a route network layout for a transport system based on a geographical map, in particular based on a predetermined one City map of an urban area, with the following steps: - Providing (S11) a _{GAN model trained on existing route network layouts (S pre} ) with a data-based generator model (11) which is trained to produce route network layouts (S _gen ) depending on a To generate random vector (Z), and a data-based discriminator model (12) which is trained to distinguish generated and existing route network layouts (S _gen , S _pre ) from one another; - Using (S2, S3, S12) the GAN model to determine a route network layout for an existing geographical map. Procedure according to Claim 1 , wherein the route network layout is created by generating a possible route network layout based on a city map and checking the generated route network layout using the trained discriminator model (12), with a possible route network layout as a suitable route network layout for the City map is provided when the discriminator model (12) recognizes the possible route network layout as an existing route network layout. Procedure according to Claim 1 , the GAN model trained on existing route network layouts being provided with a generator model (11) which is trained to generate route network layouts depending on a random vector (Z) and default information (V _gen ) which comprises at least one city map , to generate, and a discriminator model (12) which is designed to distinguish generated and existing route network layouts (S _gen , S _pre ) from one another as a function of respectively assigned default information (V _gen); wherein the route network layout is generated with the aid of the trained generator model (11) as a function of default information (V _gen ) for the specified city map. Procedure according to Claim 3 , wherein the specification information (V _gen ) further comprises one or more of the following specification information: - city information, such as population, population density and / or number of commuters, - satellite image information, in particular in connection with extracted quantifiable features, - topology information from the City area covered by the city map, which indicates a height profile of the subsurface, - existing other public or private alternative transport options along with timetable information and - permitted areas that can be used for the route network. Method according to one of the Claims 1 until 4th , additional boundary conditions being specified in the generator model (11) in order to determine a route network layout for the geographical map as a function of the boundary conditions. Procedure according to Claim 5 , wherein fixed locations for stops are specified as a boundary condition, in particular with the help of a placeholder value that indicates whether a location of a stop should be defined or can be freely selected. Procedure according to Claim 5 or 6th , wherein a boundary condition for creating the route network layout is taken into account in that the discriminator model (12) identifies route network layouts that do not meet the specified boundary condition as negative examples that are used to train the generator model (11). Device, in particular a data processing device, for creating a route network layout for a transport system based on a geographical map, in particular based on a predetermined city map of an urban area, the device being designed to: - Provide a trained on existing route network layout (S _pre ) GAN model with a data-based generator model (11) that is trained to generate route network layouts (S _gen ) depending on a random vector (Z), and a data-based discriminator model (12) that is trained to generate and existing route networks - to distinguish between layouts (S _gen , S _pre); - Using the GAN model to determine a route network layout for a geographic map. Computer program with program code means which is set up to implement a method according to one of the Claims 1 until 7th execute when the computer program is executed on a computing unit. Machine-readable storage medium with a computer program stored thereon Claim 9 .

Images