Training device, estimation device, and trained model

Abstract

To make it possible to estimate fluctuation parameters in offset following.SOLUTION: A device according to an aspect of the present disclosure is a training device including a memory that stores an unlearned model, and a processor that executes training of the unlearned model. The unlearned model is a model whose input node is data included in a signal control command for remotely controlling a traffic signal controller, and whose output node is a fluctuation parameter in offset following executed by the traffic signal controller. The processor executes the training using a set of data included in the signal control command and data included in signal control execution information representing control details executed by the traffic signal controller as teacher data.SELECTED DRAWING: Figure 7

Description

Technical Field

[0001] The present disclosure relates to a training device, an estimation device, and a learned model.

Background Art

[0002] Patent Document 1 describes a roadside relay device that provides signal information to a vehicle. Based on a signal control command transmitted by a central device for remotely controlling a traffic signal controller and operation state information transmitted by the traffic signal controller to the central device, the roadside relay device of Patent Document 1 generates signal information for a vehicle. Further, the roadside relay device transmits the generated signal information for the vehicle to a communication device capable of wireless communication with the vehicle.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] When an offset change is instructed by a signal control command, the traffic signal controller performs offset following. In this case, the traffic signal controller gradually changes the cycle length or the like within the range of the follow-up width specified by the command in the follow-up direction specified by the signal control command. However, the calculation method of the variation parameter such as the cycle length to be varied in offset following is not defined by police specifications or the like and depends on the manufacturer's own concept. Therefore, when providing signal information to a vehicle, it is difficult to grasp accurate signal information during offset following.

[0005] In view of such conventional problems, an object of the present disclosure is to enable estimation of variation parameters in offset following.

Means for Solving the Problems

[0006] An apparatus according to one aspect of the present disclosure is a training apparatus comprising a memory for storing an untrained model and a processor for training the untrained model, wherein the untrained model is a model in which data included in a signal control command for remotely controlling a traffic signal controller is used as input nodes and fluctuating parameters in offset tracking performed by the traffic signal controller are used as output nodes, and the processor performs the training using a set of data included in the signal control command and data included in signal control execution information representing the control content performed by the traffic signal controller as training data.

[0007] A model according to one aspect of the present disclosure is a trained model for causing a processing unit to function to output variable parameters in offset tracking performed by a traffic signal controller based on data included in a signal control command for remotely controlling the traffic signal controller, and comprises a neural network having an input layer including a plurality of input nodes, a hidden layer including a plurality of intermediate nodes, and an output layer including a plurality of output nodes, wherein the weight and bias values ​​between each node included in the neural network are values ​​trained using a set of data including the data included in the signal control command as input data and data included in signal control execution information representing the control content performed by the traffic signal controller as training data. [Effects of the Invention]

[0008] According to this disclosure, it is possible to estimate the variation parameters in offset tracking. [Brief explanation of the drawing]

[0009] [Figure 1] Figure 1 is a block diagram showing an example of an information provision system. [Figure 2] Figure 2 shows an example of a signal control command format. [Figure 3] Figure 3 shows an example of the format of signal control execution information. [Figure 4]Figure 4 shows an example of the format for signal operation status information. [Figure 5] Figure 5 shows an example of a timetable generated by the connection adapter. [Figure 6A] Figure 6A shows the data structure of signal information. [Figure 6B] Figure 6B is an explanatory diagram showing the data values ​​and data contents stored in the header and data sections of the signal information. [Figure 7] Figure 7 is a block diagram showing an example of the model configuration. [Figure 8] Figure 8 is a block diagram showing an example configuration of a training device for an untrained model. [Figure 9] Figure 9 is a block diagram showing an example configuration of a parameter estimation device. [Modes for carrying out the invention]

[0010] <Summary of the embodiments of this disclosure> The embodiments of this disclosure are outlined below. (1) The training device according to this embodiment is a training device comprising a memory for storing an unlearned model and a processor for training the unlearned model, wherein the unlearned model is a model in which data included in a signal control command for remotely controlling a traffic signal controller is used as an input node and variable parameters in offset tracking performed by the traffic signal controller are used as output nodes, and the processor performs the training using a set of data included in the signal control command and data included in signal control execution information representing the control content performed by the traffic signal controller as training data.

[0011] According to the training device of this embodiment, the above training generates a trained model that outputs variable parameters from the data included in the signal control command. Therefore, by inputting the data included in the signal control command into the learned model, the variable parameters in offset tracking are output. Accordingly, it is possible to estimate the variable parameters in offset tracking, which may differ depending on the manufacturer, in terms of the calculation method.

[0012] (2) In the training device of the present embodiment, the input node may include at least one of, for example, cycle length, split, presence / absence of offset change, follow direction specification, follow width specification, offset tracking monitoring, offset reference time, and offset value. The reason is that at least one of the above-mentioned information is required when the traffic signal controller performs offset tracking regardless of the algorithm adopted by the traffic signal controller.

[0013] (3) In the training device of the present embodiment, the output node may include, for example, cycle length, and the duration of the variable ramp, or the durations of a plurality of ramps including the variable ramp. The reason is that the parameters (variable parameters) that the traffic signal controller variably changes step by step in offset tracking usually include cycle length and the number of seconds of the variable ramp.

[0014] (4) In the training device of the present embodiment, the processor may execute the training on the condition that the traffic signal controller is in the remote operation state. In this way, it is possible to prevent the training of the model based on the signal control execution information of the traffic signal controller that is not in the remote control state, and execute the training based on normal teacher data.

[0015] (5) In the training device of the present embodiment, the processor may execute the training on the condition that the traffic signal controller is in the offset synchronization state. In this way, it is possible to prevent the training of the model based on the signal control execution information of the traffic signal controller that is not in the offset tracking state, and execute the training based on normal teacher data.

[0016] (6) In the training device of this embodiment, the processor may perform the training on the condition that there is no abnormality in the operation of the traffic signal controller. In this way, it is possible to prevent training models based on signal control execution information from traffic signal controllers that are presumed to be malfunctioning, and instead perform training based on normal training data.

[0017] (7) The model according to this embodiment is a trained model for causing a processing unit to function to output variable parameters in offset tracking performed by a traffic signal controller based on data included in a signal control command for remotely controlling the traffic signal controller, and is composed of a neural network having an input layer including a plurality of input nodes, a hidden layer including a plurality of intermediate nodes, and an output layer including a plurality of output nodes, wherein the weight and bias values ​​between each node included in the neural network are values ​​trained using a set of data including the data included in the signal control command as input data and data included in signal control execution information representing the control content performed by the traffic signal controller as training data.

[0018] The model in this embodiment is a trained model in which the weight and bias values ​​have been adjusted through the above training to output variable parameters from the data included in the signal control command. Therefore, by inputting the data included in the signal control command into the trained model, the variable parameters in offset tracking are output. Consequently, it is possible to estimate the variable parameters in offset tracking, which may differ depending on the manufacturer's calculation method.

[0019] (8) An estimation device according to one aspect of this embodiment is an estimation device comprising a memory for storing a trained model and a processor for estimating fluctuation parameters in offset tracking performed by a traffic signal controller, wherein the trained model is a model trained by the training device described above, and the processor inputs data included in a signal control command for remotely controlling the traffic signal controller into the trained model and uses the output data obtained as the fluctuation parameters.

[0020] According to the estimation device of this embodiment, the processor inputs data included in the signal control command for remotely controlling a traffic signal controller into a trained model, and the output data obtained is used as the variable parameters. Therefore, the variable parameters in offset tracking, whose calculation method may differ depending on the manufacturer, can be easily estimated from the signal control command.

[0021] (9) An estimation device according to another embodiment of this embodiment is an estimation device comprising a memory for storing a data conversion tool and a processor for estimating fluctuation parameters in offset tracking performed by a traffic signal controller, wherein the data conversion tool is a conversion tool capable of performing data conversion equivalent to or approximating a trained model, the trained model is a model trained by the training device described above, and the processor inputs data included in a signal control command for remotely controlling the traffic signal controller into the data conversion tool and uses the output data obtained as the fluctuation parameters.

[0022] According to the estimation device of this embodiment, the processor inputs data included in the signal control command for remotely controlling a traffic signal controller into a data conversion tool, and the output data obtained is used as the variable parameter. Therefore, the variable parameter in offset tracking, whose calculation method may differ depending on the manufacturer, can be easily estimated from the signal control command.

[0023] This disclosure can be implemented not only as a system and apparatus having the characteristic configuration described above, but also as a program for causing a computer to execute such characteristic configuration. Furthermore, this disclosure can be implemented as a semiconductor integrated circuit that implements part or all of the system and apparatus.

[0024] <Details of the embodiments of this disclosure> The embodiments of this disclosure will be described in detail below with reference to the drawings. At least some of the embodiments described below may be combined in any way.

[0025] [Example of an information provision system configuration] Figure 1 is a block diagram showing an example of the information provision system 100. As shown in Figure 1, the information provision system 100 of this embodiment is a system that provides signal information representing the color status of traffic signal lights 5 to vehicles 6 traveling on a road, and comprises a central unit 1, a traffic signal controller 2, a connection adapter 3, and a communication device 4.

[0026] Central device 1 is a server owned by the operator responsible for traffic control. Based on traffic information collected from vehicle detectors and probe vehicles (neither of which are shown), central device 1 performs traffic-sensitive control (centralized control) for multiple intersections. Traffic-sensitive control includes "system control," which controls the operation of signal groups belonging to a system section, and "wide-area control (area control)," which extends system control to the road network.

[0027] Therefore, the central unit 1 can communicate with multiple traffic signal controllers within its jurisdiction, including the traffic signal controller 2 shown in Figure 1. The central unit 1 employs a "table control method" as the method for remotely controlling the traffic signal controller 2. The table control method is a system in which the central unit 1 transmits a "signal control command" (see Figure 2) containing signal control parameters such as cycle, split, and offset to the traffic signal controller 2. In this method, the traffic signal controller 2 determines a "timetable" (see Figure 5) that represents the duration of each stage of the signal lights 5 based on the received signal control command.

[0028] Therefore, when the central unit 1 performs traffic-sensitive control, it generates a signal control command for each intersection and transmits this signal control command to the traffic signal controller 2, which controls the operation of the signal lights 5. The traffic signal controller 2 switches the color state of the signal lights 5 (green light, red light, yellow light, and right-turn arrow, etc.) according to a timetable determined by the unit based on the signal control command. The timetable is generated for each cycle.

[0029] In the table control method, the traffic signal controller 2 transmits "signal control execution information" (see Figure 3), which represents the control content performed in the previous cycle, to the central device 1 at the start of the current cycle. The information stored in the signal control execution information includes the cycle start time of the previous cycle (start time of the first step), the number of seconds for each step executed, the content of the offset executed, the type of sensor execution representing the type of terminal sensor control executed, and the sensing information of the vehicle sensor (traffic volume and occupancy time).

[0030] The traffic signal controller 2 can also transmit "signal operation status information" (see Figure 4), which indicates the operating status of the current cycle, to the central device 1. The signal operation status information includes, for example, execution stage information including the current display status and stage number, and operation status for notifying of malfunctions in the machine (such as timer malfunctions or CPU malfunctions). The execution stage information is always transmitted to the central device 1 at the timing of the stage progression.

[0031] The connection adapter 3 has the following functions 1 to 3. In this respect, it is similar to the roadside relay device described in Patent Document 1. The internal configuration of the connection adapter 3 will be described later. Function 1: A function to relay and intercept communications between the central unit 1 and the traffic signal controller 2. Function 2: A function that generates "signal information" (see Figure 6) for vehicles from signal control commands received from the central unit 1 and signal operation status information received from the traffic signal controller 2. Function 3: Communication function with communication device 4 that provides signal information wirelessly.

[0032] Communication device 4 is a communication device for wirelessly providing signal information to vehicle 6, and is, for example, an ITS (Intelligent Transport Systems) radio or a cellular communication base station. If the communication device 4 is an ITS radio, the signal information is transmitted wirelessly to the ITS-compatible vehicle 6. If the communication device 4 is a base station, the signal information is transferred from the base station to the cloud server of the core network. The cloud server transfers the signal information to the same base station or another base station, and that base station wirelessly transmits the signal information to the mobile terminal installed in the vehicle 6.

[0033] [Signal control command, signal control execution information, and signal operation status information] Figure 2 shows an example of a signal control command format. Figure 3 shows an example of the format of signal control execution information. Figure 4 shows an example of the format for signal operation status information. These formats are defined in the "U-type Traffic Signal Controller U-type Communication Application Standard" published by the Japan Traffic Management System Association (UTMS Association). Therefore, the data content included in each format is defined in the said standard document.

[0034] [Timetable] Figure 5 shows an example of a timetable created by the connection adapter 3. In the example shown in Figure 5, one cycle consists of the following eight steps. Entrance R1 is an entrance that extends in a first direction (e.g., east-west) and is equipped with a pedestrian signal. Entrance R2 is an entrance that extends in a second direction (e.g., north-south) that intersects with the first direction and is not equipped with a pedestrian signal.

[0035] 1PG: Both the vehicle and pedestrian signals on the entrance ramp R1 are green. 1PF: The vehicle traffic light on entrance ramp R1 is green and the pedestrian traffic light is flashing blue. 1PR: The vehicle traffic light on entrance ramp R1 is green and the pedestrian traffic light is red. 1Y: The vehicle traffic light on entrance ramp R1 is yellow and the pedestrian traffic light is red. 1AR: Both inflow channels R1 and 2 are red (all red) 2G: The vehicle signal light for entrance ramp R2 is blue. 2Y: Vehicle traffic light on entrance ramp R2 is yellow. 2R: The vehicle signal light for entrance ramp R2 is red.

[0036] In the example shown in Figure 5, of the eight stages, 1PG and 2G are variable stages that can change due to the execution of offset tracking and terminal-sensitive control. Therefore, unlike the other stages, the duration of 1PG and 2G is defined as a time range rather than a fixed length. In the example shown in Figure 5, the duration range for 1PG is 30 to 50 seconds, and the duration range for 2G is 20 to 40 seconds. When the traffic signal controller 2 is performing offset tracking or terminal-sensitive control, it determines the durations of 1PG and 2G according to the control results.

[0037] [Signal information for vehicles] Figure 6 shows an example of a signal information format for vehicles. Specifically, Figure 6A shows the data structure of the signal information. Figure 6B is an explanatory diagram showing the data values ​​and data contents stored in the header and data sections of the signal information. The signal information in Figure 6B is the signal information for inflow channel R1 in the timetable in Figure 5.

[0038] As shown in Figure 6A, the signal information for vehicles has a data structure that includes a header section, a data section, and a footer section. The header section contains an identifier indicating that it is signal information, the size of the signal information, and the number of light colors to be provided (3 in the example diagram). The footer section stores the CRC (Cyclic Redundancy Check) value, etc. The data section stores the planned display time (in seconds in the example diagram) for the number of light colors (1) to (3) defined in the header section.

[0039] In Figure 6B, the correspondence between the data values ​​(codes) for light colors (1) to (3) and the actual signal light colors is as follows: The signal light color (1) with code "01" = green light The signal light color (2) with code "02" = yellow light The signal light color (3) with code "03" = red light

[0040] As shown in Figure 6B, in the inlet passage R1, the shortest time for the display of light color (1) (= green light) is 40 seconds, and the longest time is 70 seconds. The minimum guaranteed time for the light color (1) stores the sum of the shortest times when the same display is shown as a vehicle light in the stages from the currently executing stage (1PG) onward. In the example in Figure 6B, 10 seconds is stored as the sum of the shortest times for 1PF and 1PR.

[0041] In entrance lane R1, the minimum and maximum durations for the display of light color (2) (= yellow signal) are both 5 seconds. In entrance ramp R1, the minimum and maximum display times for light color (3) (=red light) are both 55 seconds. Thus, the display time for light colors where the minimum and maximum times coincide is fixed. Note that the display time may be expressed in units of 100 milliseconds or 10 milliseconds, and the format itself is not limited to the format shown in Figure 6.

[0042] [Internal configuration of the connection adapter] Returning to Figure 1, the connection adapter 3, which is a type of roadside relay device, comprises a first communication unit 31, a second communication unit 32, a third communication unit 33, a control unit 34, a storage unit 35, and a synchronization processing unit 36. The first communication unit 31 is a communication board to which a communication line 41 conforming to the communication standard (e.g., UD type transmission method) adopted by the central device 1 can be connected. Other relay devices such as routers may be interposed in the communication path between the central device 1 and the first communication unit 31.

[0043] The second communication unit 32 is a communication board to which a communication line 42 conforming to the communication standard (e.g., U-type transmission method) adopted by the traffic signal controller 2 can be connected. Other relay devices such as routers may be interposed in the communication path between the traffic signal controller 2 and the second communication unit 32. The third communication unit 33 is a communication board to which a communication line 43 conforming to the communication standard adopted by the communication device 4 (for example, Ethernet; "Ethernet" is a registered trademark) is connected. Other relay devices such as routers may be interposed in the communication path between the communication device 4 and the third communication unit 33.

[0044] The control unit 34 is an arithmetic processing unit that includes one or more CPUs (Central Processing Units) and one or more RAMs (Random Access Memory). The control unit 34 may also include an integrated circuit such as an FPGA (Field-Programmable Gate Array). The control unit 34 reads the computer program 37 stored in the memory unit 35 into the main memory (RAM) and performs predetermined information processing according to the program 37. Predetermined information processing includes relaying communication frames and generating signal information.

[0045] The storage unit 35 consists of auxiliary storage devices including non-volatile memory such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive). The storage unit 35 may include flash ROM (Read Only Memory), USB (Universal Serial Bus) memory, or an SD card.

[0046] The synchronization processing unit 36 ​​is a processing unit that synchronizes the time with other communication nodes, such as the traffic signal controller 2, using a predetermined synchronization method. The control unit 34 determines the timing of turning on or off each traffic light to be included in the signal information, according to the local time generated by the synchronization processing unit 36. The synchronization method of the synchronization processing unit 36 ​​may employ, for example, a synchronization method based on the output of a GNSS (Global Navigation Satellite System) receiver, or a synchronization method using communication frames such as NTP (Network Time Protocol) and PTP (Precision Time Protocol).

[0047] The control unit 34 can emulate control operations similar to those of the traffic signal controller 2 based on signal control commands and signal operation status information. For example, the control unit 34 can create a timetable (see Figure 5) based on the signal control parameters (cycle, split, and offset, etc.) included in the signal control command. The control unit 34 can also perform the same type of terminal-sensitive control as the traffic signal controller 2 based on the type of terminal-sensitive control specified in the signal control command.

[0048] The control unit 34 generates signal information for the vehicle (see Figure 6) based on the created timetable. Details of the "signal information generation process" by the control unit 34 will be described later. When the control unit 34 generates signal information for the vehicle, it generates a communication frame containing the generated signal information and outputs the generated communication frame to the third communication unit 33. The third communication unit 33 transmits the communication frame containing the input signal information to the communication device 4.

[0049] [Signal information generation process] The signal information generation process performed by the control unit 34 includes the following processes 1 to 4. The control unit 34 executes the following processes 1 to 4 in a sufficiently short predetermined calculation cycle (e.g., 100 milliseconds) to update the signal information in near real-time. When the control unit 34 outputs signal information to the third communication unit 33, it outputs the latest (most recent) signal information at the time of output.

[0050] Process 1: Create a timetable The control unit 34 creates a timetable to be applied to the next cycle from the signal control command (see Figure 2) and signal operation status information (see Figure 4). The timetable includes the number of steps included in one cycle, the duration of each step (e.g., in seconds), and the cycle start time Ts (see Figure 5).

[0051] Specifically, the control unit 34 determines multiple stages to be included in the timetable and the duration of each stage, based on the stage number included in the signal control command, the reference values ​​for splits 1 to 6, the positive fluctuation value, the negative fluctuation value, the cycle length, and constants representing the relationship between each stage and each indication.

[0052] Process 2: Calculation of the duration of the variable steps The control unit 34 determines from the sensor permission included in the signal control command whether the traffic signal controller 2 is performing terminal sensor control (control that changes the variable level based on sensing information from vehicle sensors). If the control unit 34 is performing terminal sensing control, it performs terminal sensing control of the type specified by the sensing permission included in the signal control command (for example, dilemma sensing, bus sensing, or gap sensing) and reflects this in the duration of the variable step.

[0053] Process 3: Correction of the duration of the steps If the control unit 34 detects a slight difference of a predetermined value or more between the duration of the stage it calculated (first time) and the execution time of the stage calculated from the signal operation status information (second time), it corrects the first time based on the timing of receiving the execution stage information transmitted when the stage changes. Specifically, it adjusts the duration of each stage included in the timetable to match the second time.

[0054] Process 4: Creation of signal information The control unit 34 creates signal information in a predetermined format (see, for example, Figure 6) based on the calculated duration for each stage. The signal information includes the color of the light in each inflow path, and the planned display time for each light color from the current time onward (remaining seconds from the current time).

[0055] [Challenges and solutions when tracking offsets] Offset tracking refers to a control method that does not immediately respond to the offset value specified in the signal control command, but rather extends or shortens the cycle length to gradually approach that offset value. Regarding offset following, the standard specifies that, taking into consideration the impact on traffic flow, the maximum following width should be, for example, 12.5% ​​or 25% of the cycle length, and that following should be completed within 4 cycles from the start of following. However, the traffic signal controller 2 performs offset following independently, provided that it follows the following width and following direction instructed by the signal control command.

[0056] In other words, the method for calculating the parameters to be varied during offset tracking (e.g., cycle length and the number of seconds for the variable step; hereinafter referred to as "variable parameters") is outside the scope of the standard and therefore relies on the manufacturer-specific algorithm of the traffic signal controller 2. Therefore, when the traffic signal controller 2 is performing offset tracking, the control unit 34 of the connection adapter 3 cannot calculate the fluctuating parameters determined by the controller 2 using its own algorithm, and thus cannot reproduce the timetable before the start of the cycle. Consequently, there is a problem in that the control unit 34 of the connection adapter 3 cannot provide accurate signal information.

[0057] In this embodiment, to solve the above problems, a trained model ML capable of outputting variable parameters from a signal control command is generated by training an untrained model MU using a dataset of signal control commands and signal control execution information, which is the control performance of the traffic signal controller 2, as training data, with the signal control command as input data and the variable parameters as output data. Specifically, the computer program 37 in the memory unit 35 includes a learning program 37A that causes the control unit 34 to perform the above-mentioned training.

[0058] In other words, the machine learning function of the control unit 34 is a process that generates a model ML capable of calculating variable parameters by training the model MU. Hereafter, the untrained model MU may be referred to as "untrained model MU" or "model MU," and the trained model ML may be referred to as "trained model ML" or "model ML." Furthermore, the collective term for both the untrained model MU and the trained model ML will be abbreviated as "model MU,ML."

[0059] Models MU and ML are defined, for example, as neural networks in which at least one hidden layer (intermediate layer) is interposed between the input layer and the output layer. The input data for models MU and ML consists of information included in signal control commands. The training data (dataset) for the untrained model MU consists of data included in signal control commands and data included in signal control execution information that represents the control content of the traffic signal controller 2. The information included in the signal operation status information is used as conditional data for determining whether training is necessary. In addition, some of the information included in the signal control execution information (for example, "abnormal information") is also used as conditional data for determining whether training is necessary.

[0060] [Examples of model configuration and training] Figure 7 is a block diagram showing example configurations for the Model MU and ML. As shown in Figure 7, the model MU,UL is a neural network having an input layer LI containing multiple input nodes, a hidden layer LH containing multiple intermediate nodes, and an output layer LO containing multiple output nodes. The connections between nodes can be fully connected or partially connected. Also, although the example figure shows one hidden layer LH, there may be two or more layers.

[0061] The input node of the input layer LI includes information belonging to the signal control command, such as cycle length, split, offset change status, tracking direction specification, tracking width specification, offset tracking monitoring, offset reference time, and offset value. The reason is that this information is necessary for the traffic signal controller 2 to perform offset tracking, regardless of the algorithm it employs. Therefore, it is preferable to use all of the above information as input nodes, but it is also acceptable to use at least one of the above information.

[0062] The output node of the output layer LO includes, as an example of the variable parameters in offset tracking, the cycle length Ci (i=1,2...), the duration of 1PG Gi1 (i=1,2...), and the duration of 2G Gi2 (i=1,2...). The discrete variable i is the cycle count, with the first cycle immediately following the start of offset tracking being considered the first cycle (i=1). For example, if the number of cycles required to complete offset tracking is specified as four or fewer, the maximum value N of the discrete variable i should be set to 4.

[0063] The training data TD in Figure 7 is the ground truth data for the output node of the output layer LO, and includes the cycle length TCi (i=1,2...), the duration of 1PG TGi1 (i=1,2...), and the duration of 2G TGi2 (i=1,2...). The value of the cycle length TCi is obtained from the planned cycle length in the signal control execution information (see Figure 3). The values ​​of the duration TGi1 and TGi2 are obtained from the execution stage n (n=1,2...24) in the signal control execution information.

[0064] The control unit 34 of the connection adapter 3 trains the model MU using the training data TD so that the weights and biases between nodes are optimized. Through this training, the weights and biases between nodes in the untrained model MU converge to obtain appropriate output data (variable parameters), and a trained model ML is generated.

[0065] Once the generation of the trained model ML is complete, the control unit 34 inputs data such as cycle length, split, tracking direction specification, and tracking width specification extracted from the signal control command into the input layer LI of the model ML, causing the model ML to output the cycle length Ci, the duration of 1PG Gi1, and the duration of 2G Gi2. The control unit 34 uses the cycle length Ci, the duration of 1PG Gi1, and the duration of 2G Gi2 output by the model ML to calculate the timetable for the next cycle (see Figure 5).

[0066] [Conditions for conducting the training] As shown in Figure 7, the control unit 34 of the connection adapter 3 determines whether the traffic signal controller 2 is under remote control (step ST11) and whether it is undergoing offset synchronization (step ST12). The information necessary for these determinations is included in the "operation status 2" of the signal operation status information. The control unit 34 then enables data input to the input layer LI only if each of the above determination results is positive.

[0067] This prevents the training of the model MU based on signal control execution information from traffic signal controllers 2 that are not under remote control, and from traffic signal controllers 2 that are not performing offset tracking. Therefore, training based on normal training data can be performed.

[0068] Furthermore, the control unit 34 of the connection adapter 3 determines whether or not the traffic signal controller 2 is functioning correctly (step ST13). The information necessary for this determination is included in the "abnormal information" (identification information such as timer abnormality or clock abnormality) of the signal control execution information. The control unit 34 then enables data input to the input layer LI only if the above determination result is positive.

[0069] This prevents the training of the model MU based on signal control execution information from the traffic signal controller 2, which is presumed to be malfunctioning. Consequently, training can be performed based on normal training data.

[0070] [Training device for untrained models] Figure 8 is a block diagram showing an example configuration of the training device 34T for the untrained model MU. Specifically, Figure 8 is a block diagram relating to the training function of the control unit 34 of the connection adapter 3. Therefore, the training device 34T is physically the same arithmetic processing unit as the control unit 34, and includes memory 34M and processor 34P.

[0071] The memory 34M includes at least one RAM, and the processor 34P includes at least one integrated circuit such as a CPU or FPGA. The memory 34M stores the untrained model MU shown in Figure 7. Processor 34P trains the untrained model MU using the data contained in the signal control command and the data contained in the signal control execution information as training data. The training includes the following processes T1 to T6.

[0072] Process T1: Data Reset Processor 34P resets the weight and bias values ​​between each node in the model MU to their initial values.

[0073] Process T2: Input and output of training data Processor 34P inputs the input data extracted from the signal control command (cycle length, split, offset change status, tracking direction specification, and tracking width specification, etc.) as training data into the input layer LI of the model MU, causing the model MU to calculate output data (Ci, Gi1, Gi2).

[0074] Process T3: Loss calculation Processor 34P compares the output data (Ci, Gi1, Gi2) with the training data (TCi, TGi1, TGi2) extracted from the signal control execution information and calculates the loss (e.g., mean squared error) of both.

[0075] Process T4: Searching for weights and biases Processor 34P repeats processes T2 and T3 using methods such as gradient descent until it obtains the node weight and bias values ​​that minimize the loss.

[0076] Process T5: Generation of trained ML model The processor 34P records the weight and bias values ​​between nodes that minimize loss in memory 34M, and the model MU with these weight and bias values ​​is designated as the trained model ML.

[0077] Process T6: Determination of training execution conditions In process T2, the processor 34P determines whether the training execution conditions are met, based on condition data (e.g., whether remote operation is in progress) included in at least one of the signal control execution information and signal operation state information. The processor 34P allows data input to the input layer LI only if the execution conditions are met.

[0078] [Device for estimating variable parameters] Figure 9 is a block diagram showing an example configuration of the parameter estimation device 34E. Specifically, Figure 9 is a block diagram relating to the parameter estimation function of the control unit 34 of the connection adapter 3. Therefore, the estimation device 34E is physically the same arithmetic processing unit as the control unit 34, and includes a memory 34M and a processor 34P.

[0079] The memory 34M includes at least one RAM, and the processor 34P includes at least one integrated circuit such as a CPU or FPGA. The trained model ML shown in Figure 7 is stored in the memory 34M. Processor 34P uses a trained model ML to calculate the variable parameter Yk from the input data Xj obtained from the signal control command. Specifically, the input data Xj is input to the input layer LI of model ML, and the data output by model ML is used as the variable parameter Yk.

[0080] As shown in Figure 9, the memory 34M can also store the data conversion tool DT that converts the input data Xj into the variable parameter Yk. The data transformation tool DT is a transformation tool designed to perform data transformations equivalent to or approximating those of the trained model ML, and which can take the form of, for example, a determinant or a reference table. In this case, the processor 34P inputs the input data Xj to the data transformation tool DT, and the output result of tool DT is used as the variable parameter Yk.

[0081] Note that "equivalent or approximate data transformation" does not mean requiring the output to be exactly the same as the solution obtained with the trained model ML. Rather, it refers to a data transformation that can output the variable parameter Yk with an error of a predetermined value (100ms) or less compared to the output value of the trained model ML.

[0082] [Other variations] The embodiments described above (including variations) are illustrative in all respects and not restrictive. The scope of rights of this disclosure is indicated by the claims and is intended to include all modifications within the meaning and scope of the equivalents of the claims.

[0083] In the above embodiment, instead of variable steps G1i and G2i as the variable parameters during offset tracking, the duration of multiple steps including variable steps, such as the blue time of inflow channel R1 and the blue time of inflow channel R2, may be used. Furthermore, the unit of the variable parameter may be something other than seconds, such as a ratio (%) to the cycle length.

[0084] In the above-described embodiment, the training device 34T of the model MU (Figure 8) may be a arithmetic processing unit mounted on an information processing device other than the connection adapter 3, such as a relay device or server (hereinafter referred to as "external device"), rather than the arithmetic processing unit (control unit 34) of the connection adapter 3. In this case, if the trained model ML generated by the external training device 34T is stored in the memory 34M of the connection adapter 3, the control unit 34 of the connection adapter 3 can be operated as a variable parameter estimation device 34E.

[0085] Similarly, in the above-described embodiment, the parameter estimation device 34E (Figure 9) may be a arithmetic processing device mounted on an external device, rather than the arithmetic processing device (control unit 34) of the connection adapter 3. In this case, if the estimation device 34E of the external device is configured to send the calculated fluctuation parameters to the connection adapter 3, the connection adapter 3 can use the received fluctuation parameters to create a time table. [Explanation of Symbols]

[0086] 1 central unit 2. Traffic signal controller 3. Connection adapter (roadside relay device) 4. Communication equipment 6 vehicles 31. First Communications Department 32 Second Communications Department 33 Third Communications Department 34 Control Unit (Arithmetic Processing Unit) 34T Training Device (Computational Processing Unit) 34E Estimation device (arithmetic processing unit) 34M memory 34P processor 35 Storage section 36 Synchronization Processing Unit 37 Computer Programs 37A Learning Program 41 Communication lines 42 Communication lines 43 Communication lines 100 Information Provision System MU (Untrained Model) ML pre-trained models (pre-trained models) DT Data Conversion Tool

Claims

1. Memory for storing untrained models, A training device comprising a processor that performs training on the aforementioned untrained model, The aforementioned untrained model, This model uses data included in signal control commands for remotely controlling a traffic signal controller as the input node, and the fluctuating parameters in the offset tracking performed by the traffic signal controller as the output node. The aforementioned processor, A training device that performs the training using a set of data, consisting of data included in the signal control command and data included in the signal control execution information representing the control content executed by the traffic signal controller, as training data.

2. The aforementioned input node includes: The training device according to claim 1, comprising at least one of the following: cycle length, split, offset change status, tracking direction specification, tracking width specification, offset tracking monitoring, offset reference time, and offset value.

3. The output node includes: The training apparatus according to claim 1 or claim 2, comprising cycle length and duration of a variable step, or duration of a plurality of steps including the variable step.

4. The aforementioned processor, The training device according to claim 1 or claim 2, wherein the training is performed on the condition that the traffic signal controller is operating remotely.

5. The aforementioned processor, The training device according to claim 1 or claim 2, wherein the training is performed on the condition that the traffic signal controller is in offset synchronization.

6. The aforementioned processor, The training device according to claim 1 or claim 2, wherein the training is performed on the condition that the operation of the traffic signal controller is normal.

7. A trained model for causing a processing unit to function to output variable parameters in offset tracking performed by a traffic signal controller based on data included in a signal control command for remotely controlling the traffic signal controller, An input layer containing multiple input nodes, A hidden layer containing multiple intermediate nodes, It consists of a neural network having an output layer containing multiple output nodes, The weight and bias values ​​between each node in the aforementioned neural network are: A trained model is a set of data that has been trained using as training data the data included in the signal control command as input data and the data included in the signal control execution information representing the control content performed by the traffic signal controller as training data.

8. A memory in which the trained model described in Claim 7 is stored, A processor for estimating the variable parameters in the offset tracking performed by the traffic signal controller, The aforementioned processor, An estimation device that inputs data included in a signal control command for remotely controlling the traffic signal controller into a trained model stored in the memory, and uses the output data obtained as the variable parameter.

9. Memory for storing data conversion tools, An estimation device comprising a processor for estimating fluctuation parameters in offset tracking performed by a traffic signal controller, The aforementioned data conversion tool is A transformation tool capable of performing data transformations equivalent to or approximating those of a pre-trained model. The aforementioned trained model is A model trained by the training device described in claim 1, The aforementioned processor, An estimation device that inputs data included in a signal control command for remotely controlling the traffic signal controller into the data conversion tool, and uses the output data obtained as the variable parameter.

Images