Route searching device, method, and program

Abstract

According to one embodiment of the present invention, when searching, by using a machine learning model, for a route for traveling and visiting a plurality of station buildings which constitute a network by means of a plurality of vehicles, a movement target vehicle and station buildings which serve as the destinations of the movement target vehicle are first selected from among the plurality of vehicles and the plurality of station buildings after excluding events already selected in the past. Next, during a movement period in which the selected movement target vehicle moves to the station buildings, a station building in which a change in the operation state is anticipated is selected, and the degree of influence that is exerted on the network by the change in the operation state of said station building is calculated. Then, the model parameter of the machine learning model is updated by reflecting the calculated degree of influence in the reward of a corresponding route candidate to be searched for by the machine learning model so that the machine learning model selects, on the basis of the reward, a route candidate having a large reward.

Description

Route search device, method, and program

[0001] One aspect of the present invention relates to a route search device, method, and program used in the task of visiting communication buildings to restore power to communication equipment and restore network services, for example, when network services are stopped during a disaster.

[0002] For example, if a widespread power outage occurs due to a disaster such as an earthquake, and the power supply to the communication equipment in a communication building operated by a communication carrier runs out, users of that communication equipment will be unable to communicate. This situation is particularly undesirable if it occurs in a communication building that provides communication services to highly public buildings such as government offices, as it will seriously hinder disaster response operations.

[0003] To prepare for such an event, telecommunications carriers are considering measures to dispatch power supply vehicles to multiple telecommunications buildings in the affected area to restore power to each building.In this case, an important issue to be considered is how to determine the route for the power supply vehicles to minimize the time that telecommunications services are interrupted.

[0004] As a method for searching for a route, Non-Patent Documents 1 and 2 propose a method for solving a route generation problem using deep reinforcement learning. By using these technologies, for example, the problem of routing power supply vehicles can be treated as a traveling salesperson problem (TSP) or a vehicle routing problem (VRP), and a route can be generated that minimizes the travel distance of the power supply vehicle while taking into account the time it takes for the emergency power supply at the accommodation station to run out.

[0005] Wouter Kool et al. Attention, learn to solve routing problems! ICLR, 2018.Zhao Wang et al. A method of delivering fuel to telecommunication exchange buildings in disaster response. ICAIIC, 2021.

[0006] When searching for a route, it is necessary to consider the impact on the communication services provided by the network. The impact on the communication services is calculated based on the time required to visit each accommodating station, the order of visits, and the connection relationships between the accommodating stations in the network.

[0007] However, conventional route search methods do not take into account the impact on the network when a power supply vehicle travels along the generated route. Specifically, the operating status of the power supply station may change while the power supply vehicle is moving along the route, but the impact of this status change on the network is not quantitatively evaluated, making it impossible to search for a route that minimizes the impact. Furthermore, because the network topology, i.e., the connection relationships between the power supply stations, is not taken into account, it is impossible to search for a route with less impact.

[0008] The present invention has been made in light of the above circumstances, and aims to provide a technique that enables route searching that takes into account the degree of impact on the network.

[0009] In order to solve the above problem, one aspect of a route search device or route search method according to the present invention uses a machine learning model to search for a route for a plurality of vehicles to visit a plurality of stations constituting a network. First, a vehicle to be moved and its destination station are selected from the plurality of vehicles and the plurality of stations, after excluding previously selected events. Next, stations whose operating status is expected to change during the travel period in which the selected vehicle to be moved travels to the station are selected, and the degree of impact of the change in the operating status of the station on the network is calculated. The calculated degree of impact is then reflected in a reward for the corresponding route candidate searched by the machine learning model, and model parameters of the machine learning model are updated based on the reward so that the machine learning model selects a route candidate with a high reward.

[0010] According to one aspect of the present invention, it is possible to provide a technology that makes it possible to search for a route that will have the least impact on the network, taking into account the impact that changes in the operating status of a station will have on the network.

[0011] FIG. 1 is a diagram illustrating an example of a system to which a route search device according to an embodiment of the present invention is applied. FIG. 2 is a block diagram illustrating an example of a functional configuration of a route search device according to an embodiment of the present invention. FIG. 3 is a block diagram illustrating an example of a hardware configuration of a route search device according to an embodiment of the present invention. FIG. 4 is a block diagram illustrating an example of a software configuration of a route search device according to an embodiment of the present invention. FIG. 5 is a flowchart illustrating an example of a series of processing steps and processing contents including a route search process executed by a control unit of the route search device shown in FIG. 4. FIG. 6 is a flowchart illustrating an example of a processing step and processing contents of a route search process among the processing steps shown in FIG. 5. FIG. 7 is a flowchart illustrating an example of a processing step and processing contents of a reward calculation process among the route search processing steps shown in FIG. 6. FIG. 8 is a diagram illustrating an example of the operation of the route search processing unit shown in FIG. 4. FIG. 9 is a diagram illustrating an example of the operation of an action selection process among the route search process shown in FIG. 6. FIG. 10 is a diagram illustrating an example of the operation of a reward calculation process shown in FIG. 7. FIG. 11 is a diagram illustrating an example of the effect of a route search method according to an embodiment of the present invention.

[0012] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0013] [One Embodiment] (Configuration Example) (1) System FIG. 1 is a diagram showing an example of a system to which a route search device SV according to one embodiment of the present invention is applied.

[0014] In FIG. 1, BS1 to BSn are accommodation stations (hereinafter referred to as "stations") deployed by telecommunications carriers to configure a network, and these stations BS1 to BSn provide communication services to users who subscribe to the network.

[0015] In addition, MS1 to MSk are power supply vehicles that, in the event of a power outage in the commercial power supply used by stations BS1 to BSn due to an earthquake disaster, for example, will travel around to stations BS1 to BSn and charge the emergency power supplies of each station, thereby maintaining the functionality of the communication equipment equipped at the stations, i.e., communication services.

[0016] (2) Route Search Device SV (2-1) Overview of Functions Fig. 2 shows an example of functions provided in a route search device SV according to an embodiment of the present invention. The route search device SV has, as its main functions, a network information management function 100, a geographic information encoder 200, a current information encoder 300, a GAT (Graph Attention Network) encoder 400, and a route search function unit 500.

[0017] The network information management function 100 manages information relating to the configuration and status of the network (hereinafter referred to as network information). The network information includes, for example, topology information indicating the installation locations of the stations BS1 to BSn and their connection relationships, information indicating the operating status of each station BS1 to BSn, and information indicating the movement status of the power supply vehicles MS1 to MSk.

[0018] The geographic information encoder 200 receives and encodes the geographic information from the network information management function 100, and outputs its features. The input geographic information includes information indicating the installation locations of the stations BS1 to BSn and the travel times of the power supply vehicles MS1 to MSk when they travel between the stations BS1 to BSn.

[0019] The current information encoder 300 receives information representing the current status of the network (current information) from the network information management function 100, encodes it, and outputs its features. The current information includes, for example, information representing the operating status of the stations BS1 to BSn, the predicted time until the emergency power supplies of the stations BS1 to BSn stop, and the travel time required for the power supply vehicles MS1 to MSk to travel between the stations BS1 to BSn. The travel time also includes the time the power supply vehicles MS1 to MSk spend at the stations BS1 to BSn.

[0020] The GAT encoder 400 receives network topology information from the network information management function 100, encodes it, and outputs its features. The topology information is represented, for example, by a graph in which the stations BS1 to BSn are nodes and the connections between the stations BS1 to BSn are edges.

[0021] The route search function unit 500 inputs the feature quantities output from the geographic information encoder 200, current information encoder 300, and GAT encoder 400 into a machine learning model to search for the route that the power supply vehicles MS1 to MSk will take to travel between the stations BS1 to BSn.

[0022] More specifically, the system first selects the power supply vehicle to be moved and the station to which it will be moved, eliminating any previously selected events. Next, it calculates the impact on the network of changes in the status of stations BS1 to BSn during the selected power supply vehicle's movement, and reflects this impact in the reward for the machine learning model. The machine learning model then updates the model parameters based on the reward, allowing it to search for a route with the least impact.

[0023] (2-2) Hardware and Software Configuration The route search device SV is configured, for example, by a personal computer and is installed, for example, in a management center operated by a telecommunications carrier. Note that the route search device SV may also be configured, for example, by a server computer located on the Web or the cloud.

[0024] 3 and 4 are block diagrams showing an example of the hardware configuration and software configuration of a route search device SV according to an embodiment of the present invention.

[0025] The route search device SV includes a control unit 1 that uses a hardware processor such as a central processing unit (CPU). A storage unit having a program storage unit 2 and a data storage unit 3, and a communication interface unit (hereinafter referred to as a communication I / F unit) 4 are connected to the control unit 1 via a bus 5.

[0026] Under the control of the control unit 1, the communication I / F unit 4 uses a communication protocol defined by a network (not shown) including, for example, the Internet, to send and receive information data between the control terminals installed in the station buildings BS1 to BSn and the on-board terminals installed in the power supply vehicles MS1 to MSk.

[0027] The program storage unit 2 is configured by combining, for example, a non-volatile memory such as a hard disk drive (HDD) or a solid state drive (SSD) as a storage medium that can be written to and read at any time, and a non-volatile memory such as a read only memory (ROM), and stores application programs necessary for executing various control processes according to one embodiment of the present invention, in addition to middleware such as an operating system (OS).

[0028] The data storage unit 3 is, for example, a combination of a non-volatile memory such as an HDD or SSD as a storage medium that can be written to and read from at any time, and a volatile memory such as RAM (Random Access Memory), and the storage area includes a network information storage unit 31 and a route search model storage unit 32.

[0029] The network information storage unit 31 is used to store network information acquired from the stations BS1 to BSn and power supply vehicles MS1 to MSk. As mentioned above, the network information includes topology information that indicates the installation locations of the stations BS1 to BSn and the connection relationships between the stations BS1 to BSn, information that indicates the operating status of each station BS1 to BSn, and information that indicates the current movement status of the power supply vehicles MS1 to MSk.

[0030] The path search model storage unit 32 stores data related to a machine learning model that employs multi-agent deep reinforcement learning, including model parameters.

[0031] The control unit 1 includes, as processing units according to one embodiment of the present invention, a network information acquisition processing unit 11, a geographic information encoding processing unit 12, a GAT encoding processing unit 13, a current information encoding processing unit 14, a route search processing unit 15, and a route information output processing unit 16.

[0032] The processing units 11 to 16 are all realized by causing a hardware processor in the control unit 1 to execute an application program stored in the program storage unit 2. Note that some or all of the processing units 11 to 16 may be realized using hardware such as an LSI (Large Scale Integration) or an ASIC (Application Specific Integrated Circuit).

[0033] The network information acquisition processing unit 11 acquires topology information that indicates the installation locations and connection relationships of the stations BS1 to BSn, information that indicates the current operating status of each of the stations BS1 to BSn, and information that indicates the current movement status of the power supply vehicles MS1 to MSk via the communication I / F unit 4. The network information acquisition processing unit 11 then stores the acquired information in the network information storage unit 31 as network information.

[0034] The geographic information encoding processing unit 12 reads information representing the installation locations of the stations BS1 to BSn and the travel times of the power supply vehicles MS1 to MSk between the stations BS1 to BSn from the network information storage unit 31, encodes each of the read information, and outputs its feature values.

[0035] The GAT encoding processing unit 13 reads the topology information from the network information storage unit 31, encodes the read topology information, and outputs its feature quantities. The topology information is represented by a graph in which the stations BS1 to BSn are nodes and the connections between the stations BS1 to BSn are edges.

[0036] The current information encoding processing unit 14 reads current information that is expected to change while the power supply vehicles MS1 to MSk are in motion from the network information storage unit 31, encodes each piece of information that has been read, and outputs its feature quantities. The current information includes information that indicates the current operating status of the stations BS1 to BSn and information that indicates the travel time of the power supply vehicles MS1 to MSk between the stations BS1 to BSn. The information that indicates the operating status of the stations BS1 to BSn includes the predicted time until the emergency power supplies are shut down, and the travel time of the power supply vehicles MS1 to MSk includes the stay time at the stations BS1 to BSn.

[0037] The route search processing unit 15 uses, for example, a machine learning model that employs multi-agent deep reinforcement learning as a tool for route search. The route search processing unit 15 inputs the feature quantities output from the geographic information encoding processing unit 12, the GAT encoding processing unit 13, and the current information encoding processing unit 14 into the machine learning model stored as a route search model in the route search model storage unit 32, and uses this machine learning model to search for a route with the least impact on the network. An example of the route search processing by the route search processing unit 15 will be described in detail in the operation example.

[0038] The route information output processing unit 16 generates route information to be notified to the power supply vehicles MS1 to MSk based on the route search results obtained by the route search processing unit 15. The generated route information is then transmitted from the communication I / F unit 4 to the on-board terminals of the power supply vehicles MS1 to MSk that are the notification destinations.

[0039] (Example of Operation) Next, an example of operation of the route search device SV configured as above will be described.

[0040] For example, if a disaster occurs and causes a failure in station buildings BS1 to BSn, making it necessary for power supply vehicles MS1 to MSk to make a round visit, the route search device SV executes the process of searching for a route for the round visit as follows.

[0041] FIG. 5 is a flowchart showing an example of a series of processing steps and processing contents including a route search process executed by the control unit 1 of the route search device SV.

[0042] (1) Acquisition of Network Information In step S10, the control unit 1 of the route search device SV first acquires, under the control of the network information acquisition processing unit 11, topology information indicating the installation locations of the central offices BS1 to BSn and their connection relationships from, for example, a management database of a telecommunications carrier, via the communication I / F unit 4, and stores the acquired topology information in the network information storage unit 31. Note that the installation locations of the central offices BS1 to BSn and the topology information may also be acquired during normal times and stored in the network information storage unit 31.

[0043] During a fault, the network information acquisition processing unit 11 periodically acquires information indicating the operating status of each station BS1 to BSn via the communication I / F unit 4. The network information acquisition processing unit 11 also acquires information indicating the travel time, including the stay time, of the power supply vehicles MS1 to MSk at the stations BS1 to BSn. The network information acquisition processing unit 11 then stores the acquired information indicating the operating status and travel time in the network information storage unit 31.

[0044] (2) Encoding of Geographical Information In step S11, under the control of the geographical information encoding processing unit 12, the control unit 1 of the route search device SV reads information indicating the installation locations of the stations BS1 to BSn and the travel times of the power supply vehicles MS1 to MSk between the stations BS1 to BSn from the network information storage unit 31. Then, the control unit 1 encodes each piece of the read information and outputs its feature values.

[0045] (3) Encoding of Topology Information Subsequently, in step S12, the control unit 1 of the route search device SV reads the topology information from the network information storage unit 31 under the control of the GAT encoding processing unit 13. At this time, the topology information is represented by a graph in which the stations BS1 to BSn are nodes and the connections between the stations BS1 to BSn are edges so that the topology information can be handled by the GAT. The GAT encoding processing unit 13 encodes the read topology information and outputs its feature values.

[0046] (4) Encoding of Current Information Next, in step S13, the control unit 1 of the route search device SV, under the control of the current information encoding processing unit 14, reads from the network information storage unit 31 information representing the current operating status of the stations BS1 to BSn and information representing the travel time of the power supply vehicles MS1 to MSk between the stations BS1 to BSn. At this time, the information representing the operating status of the stations BS1 to BSn includes the predicted time until the emergency power supply is shut off. Furthermore, the travel time of the power supply vehicles MS1 to MSk includes the stay time at the stations BS1 to BSn. The current information encoding processing unit 14 then encodes each of the above read information and outputs its feature values.

[0047] (5) Route Search In step S14, the control unit 1 of the route search device SV executes a route search process under the control of the route search processing unit 15 as follows.

[0048] (5-1) Outline of Route Search Processing FIG. 8 is a diagram for explaining the outline of the operation of the route search processing executed by the route search processing unit 15. As shown in FIG.

[0049] The route search processing unit 15 inputs the feature amounts output from the geographic information encoding processing unit 12, the GAT encoding processing unit 13, and the current information encoding processing unit 14 to the route search model stored in the route search model storage unit 32. As a result, the route search model performs deep reinforcement learning in accordance with the model parameters and generates a route or a portion of a route.

[0050] In one embodiment of the route search method, when generating the above route, the route search model simultaneously searches for routes for multiple agents, i.e., multiple power supply vehicles MS1 to MSk. To do this, the route search model evaluates actions corresponding to multiple candidate routes for each power supply vehicle MS1 to MSk at each step. However, at this time, the model parameters are initialized with random numbers.

[0051] The route search processing unit 15 then selects the power supply vehicle corresponding to the action with the highest evaluation value and the station to which it will move, based on the evaluation results from the route search model. However, in this process, actions selected in the past are excluded. Next, the route search processing unit 15 calculates the impact on the network of changes in the operating status of stations BS1 to BSn that are expected to occur during the movement of the selected power supply vehicle, such as the shutdown of emergency power supplies. Then, for each of the actions, a reward is calculated that reflects the calculated impact. At this time, the reward is calculated so that the smaller the impact, the larger the value.

[0052] Based on the calculated reward value, the route search processing unit 15 updates the model parameter θ so that the route search model selects a route with a high reward. Thereafter, the route search processing unit 15 repeats the above update process until a route with an influence level equal to or less than a preset value is selected, or until a predetermined number of repetitions is reached. Through the above process, the route search model learns to generate routes with a low impact on the network.

[0053] Then, when it is determined in step S15 that the above determination conditions are satisfied, the route search processing unit 15 ends the route search process.

[0054] (5-2) Specific Example of Route Search Processing FIG. 6 is a flowchart showing an example of the processing procedure and processing content of the route search processing executed by the route search processing unit 15.

[0055] (5-2-1) Selection of power supply vehicles to be moved and their destinations In order to simultaneously search for routes for multiple power supply vehicles MS1 to MSk, the route search processing unit 15 first selects a power supply vehicle to be moved and a station building that will be the next destination (visit destination) for each step in step S141. However, in order to calculate the reward for each step, the route search processing unit 15 executes the following process to limit the time at which a power supply vehicle arrives at its next destination so that it is later than the arrival time of the last power supply vehicle to have moved.

[0056] FIG. 9 is a diagram for explaining an example of the operation for selecting a power supply vehicle MS1 to MSk and a station building as the next destination.

[0057] That is, in step S142, the route search processing unit 15 inputs feature quantities representing the network state, such as the operating status of the stations BS1 to BSn, the connection relationships between the stations BS1 to BSn, and the remaining time of the emergency power supply, into the route search model. The route search model calculates the probability of each action according to the input feature quantities and internal model parameters. Then, in step S144, the route search processing unit 15 selects the action with the highest calculated probability from among the above actions.

[0058] However, prior to selecting the above action, in step S143, the route search processing unit 15 restricts, for each step, the time at which a power supply vehicle arrives at its next destination so that it is later than the time at which the last power supply vehicle arrived at its destination. In other words, among the actions, actions that correspond to past events are excluded from the selection candidates.

[0059] For example, in the example of FIG. A In the step after power supply vehicle MS1 moves from station BS1 to station BS4, the event of power supply vehicle MS2 moving to station BS1 is excluded because it is a past event. A From the subsequent actions, for example, an action to move power supply vehicle MS1 from station BS4 to station BS2 is selected.

[0060] (5-2-2) Calculation of Reward When the power supply vehicle to be moved and an action corresponding to its destination are selected, the route search processing unit 15 then calculates the reward for the selected action in step S145 as follows.

[0061] FIG. 7 is a flowchart showing an example of the processing procedure and processing contents of the remuneration calculation processing, and FIG. 10 is a diagram for explaining an example of the operation.

[0062] That is, first, in step S1451, the route search processing unit 15 calculates the travel period of the power supply vehicle in the above action. For example, in the example of FIG. 10, the time t A The time t when the power supply vehicle MS1 moves to the next station BS2 A +t 4→2 The moving period is calculated as

[0063] Next, in step S1452, the route search processing unit 15 selects all stations whose emergency power supply is predicted to stop during the period in which the power supply vehicle is moving, based on information indicating the operating status of stations BS1 to BSn. For example, in the example of Fig. 10, station BS5 is selected as the station whose emergency power supply is predicted to stop. Note that if there is a station whose emergency power supply is predicted to be restored during the period in which the power supply vehicle is moving, for example, by receiving power from another power supply vehicle, this station is also selected.

[0064] Subsequently, in step S1453, the route search processing unit 15 calculates the network state during the above-mentioned moving period from the time t A The period from when the emergency power supply of the station BS5 stops (t A The state "045" at time t' to time t' when the emergency power supply is stopped and the time t when the power supply vehicle MS1 moves to the station BS2. A +t 4→2 ) period (t' to t A +t 4→2 ) and state "04" in the state "05." Then, for each of these states, the degree of impact that each state has on the network is calculated. The degree of impact is calculated as a value quantified based on, for example, the number of subscribers who will be unable to communicate, but is not limited to this.

[0065] Next, in step S1454, the route search processing unit 15 calculates a total influence for the entire period of the movement period based on the influence degrees calculated for each of the states. The total influence degree is calculated, for example, as a value obtained by integrating the influence degrees calculated for each of the states over time. Next, in step S1455, the route search processing unit 15 calculates a reward value that reflects the calculated total influence degrees. For example, the negative number of the total influence degree is set as the reward. In other words, the route search processing unit 15 calculates the reward so that the greater the calculated influence degree, the smaller the reward.

[0066] After the reward is calculated, the route search processing unit 15 then executes processing in step S146 to update the model parameter θ based on the reward value so that the route search model selects a route with a large reward.

[0067] (6) Transmission of Route Information Finally, in step S16, under the control of the route information output processing unit 16, the control unit 1 of the route search device SV generates route information to be notified to the power supply vehicles MS1 to MSk based on the route search results obtained by the route search processing unit 15, and transmits the generated route information to the on-board terminals of the corresponding power supply vehicles MS1 to MSk from the communication I / F unit 4. As the route information, for example, information in which the identification information of the destination station building is displayed on map data is used.

[0068] (Effects) As described above, in one embodiment, when searching for routes for the power supply vehicles MS1 to MSk to visit on a patrol using a route search model, first, for each step, the power supply vehicle to be moved and the station that will be its next destination are selected after excluding events selected in previous steps. Next, the degree of impact on the network of changes in the operating status of stations BS1 to BSn that occur during the period when the selected power supply vehicle is traveling to the destination station is calculated, and a reward that reflects the calculated degree of impact is calculated. Then, based on the calculated reward, the process of updating the model parameter θ is repeatedly performed so that the route search model selects a route with a high reward.

[0069] Therefore, according to the route search device SV of one embodiment, it is possible to search for a route that will have the least impact on the network, taking into account the impact that changes in the operating status of the stations BS1 to BSn have on the network.

[0070] FIG. 11 is a diagram showing an example of the effect of an embodiment of the present invention, which is compared through an experiment with the MA method described in Non-Patent Document 1 and the PPT method described in Non-Patent Document 2, which are conventional techniques.

[0071] In this experiment, it was assumed that a power outage occurred at 10 or 20 randomly selected stations in a network consisting of multiple stations in a certain prefecture. In this situation, a power supply vehicle traveled to the affected stations along routes generated by each method, and the overall impact was evaluated until power was restored to all stations.

[0072] As is clear from the results of this experiment, in both cases where a power outage occurred at 10 stations and where it occurred at 20 stations, the impact on the network was reduced compared to the conventional MA and PPT methods, and in particular, when there were 20 power outages, the impact was reduced by a total of 85.0%.

[0073] Furthermore, when calculating the degree of influence, the network state during the period in which the power supply vehicle is in motion is divided into a plurality of states, with the station where the emergency power supply has stopped as the boundary, a partial degree of influence is calculated for each of these states, and the calculated partial degrees of influence are integrated over time to calculate the total degree of influence for the entire period in which the power supply vehicle is in motion. This makes it possible to accurately calculate the degree of influence on the network during the period in which the power supply vehicle is in motion.

[0074] Furthermore, in one embodiment, the topology information is represented by a graph in which the stations BS1 to BSn are nodes and the connections between the stations BS1 to BSn are edges, and the feature quantities of the topology information are obtained by encoding this topology information using a GAT encoder. As a result, it is possible to perform a route search in which the connection relationships between the stations are appropriately reflected in the route search model.

[0075] [Other embodiments] The functional configuration of the route search device SV, the hardware and software configuration for realizing it, the series of processing steps and processing contents in the route search process, the type of neural network that constitutes the route search model and its operating algorithm, etc. can be modified and implemented in various ways within the scope that does not deviate from the gist of this invention.

[0076] Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in every respect. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, when implementing the present invention, specific configurations according to the embodiments may be appropriately adopted.

[0077] In short, this invention is not limited to the above-described embodiments, and in the implementation stage, the components can be modified and embodied without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0078] SV...Route search device BS1 to BSn...Branch MS1 to MSk...Power supply vehicle 100...Network information management function 200...Geographic information encoder 300...Current information encoder 400...GAT encoder 500...Route search function unit 1...Control unit 2...Program storage unit 3...Data storage unit 4...Communication I / F unit 5...Bus 11...Network information acquisition processing unit 12...Geographic information encoding processing unit 13...GAT encoding processing unit 14...Current information encoding processing unit 15...Route search processing unit 16...Route information output processing unit 31...Network information storage unit 32...Route search model storage unit

Claims

1. A route search device that uses a machine learning model to search for a route for a plurality of vehicles to make a tour of a plurality of stations constituting a network, the device comprising: a network information acquisition processing unit that acquires network information including information representing connection relationships and operating states of the plurality of stations and information representing a moving state of the vehicles; a route search processing unit that inputs features extracted from the network information into the machine learning model to search for a route for the plurality of vehicles to make a tour of the plurality of stations; and an output processing unit that generates and outputs route information based on a result of the route search, the route search processing unit comprising: a first processing unit that selects a vehicle to be moved and a station to be the destination of the move from among the plurality of vehicles and the plurality of stations, excluding events that have been selected in the past; a second processing unit that selects the station whose operating state is expected to change during the moving period in which the selected vehicle to be moved moves to the station; a third processing unit that calculates the degree of influence that a change in the operating state of the selected station has on the network, and reflects the calculated degree of influence in a reward for a corresponding route candidate searched by the machine learning model; and a fourth processing unit that updates model parameters of the machine learning model based on the reward so that the machine learning model selects a route candidate with a large reward.

2. A route search device according to claim 1, wherein said network information acquisition processing unit acquires, as information representing the configuration of said network, topology information represented by a graph in which a plurality of said offices are nodes and connections between the nodes are edges.

3. The route search device according to claim 1, wherein the second processing unit selects a station where an emergency power supply is stopped as the station where a change in the operating state is expected.

4. The route search device of claim 1, wherein the third processing unit divides the moving period into a first period until the operating status of the station changes and a second period after the operating status of the station changes, calculates a partial influence on the network for each of the first period and the second period, and combines the calculated partial influences in the time axis direction to calculate an overall influence for the entire moving period.

5. The route search device according to claim 1, wherein the third processing unit calculates the reward such that the greater the calculated degree of influence, the smaller the reward becomes.

6. A route search method executed by an information processing device, comprising: a step of acquiring network information including information representing the connection relationships and operating states of a plurality of stations constituting a network, and information representing the movement states of a plurality of vehicles; a step of inputting features extracted from the network information into a machine learning model to search for a route for a plurality of the vehicles to visit a plurality of the stations; and a step of generating and outputting route information based on a result of the route search, wherein the step of searching for the route comprises: a step of selecting a vehicle to be moved and a station to be the destination from among the plurality of the vehicles and the plurality of the stations, after excluding events that have been selected in the past; a step of selecting the station whose operating state is expected to change during the movement period in which the selected vehicle to be moved moves to the station; a step of calculating the degree of influence that the change in the operating state of the selected station has on the network, and reflecting the calculated degree of influence in a reward for the corresponding route candidate searched by the machine learning model; and a step of updating model parameters of the machine learning model based on the reward so that the machine learning model selects a route candidate with a large reward.

7. A program for causing a processor included in a route search device to execute at least one of the processes performed by the network information acquisition processing unit and the route search processing unit included in the route search device according to any one of claims 1 to 5.

Images