Measure specification program, measure specification method, and information processing apparatus

Abstract

To improve the accuracy of selecting measures.SOLUTION: An information processing apparatus executes filtering processing on a plurality of types of social measures to be examined on the basis of the characteristics of a prediction model. The information processing apparatus specifies social measures to be executed from the plurality of types of social measures to be examined on the basis a result of simulation of the one or more types of social measures on which the filtering processing is executed.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a program, a method, and an information processing device. [Background technology]

[0002] Policy analysis is required in a variety of situations, and social simulation, which is an example of simulation, is being used to enhance policy analysis. For example, when deciding on a special bus schedule for an event, policy analysis involves listing options that change various settings such as departure and arrival times and examining the pros and cons of each in order to find a schedule that will avoid the occurrence or expansion of congestion.

[0003] Social simulation is a model of social systems such as transportation systems in a computable form, and by using social simulation, it is possible to quantitatively evaluate policies and improve policy analysis. However, social simulation takes a long time to calculate, making it difficult to comprehensively evaluate policies and find the optimal ones for policy optimization.

[0004] In recent years, surrogate models have become known as an example of an alternative model to simulations, which are less accurate than simulations but require less computation time. Another well-known technique for generating surrogate models for complex social simulations is to use trained models that have been trained on a large amount of social data using neural networks. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2023-113928 [Patent Document 2] Patent Publication No. 2021-157456 [Patent Document 3] Japanese Patent Publication No. 2022-3451 [Patent Document 4] US Patent Application Publication No. 2018 / 0112498 Summary of the Invention [Problem to be solved by the invention]

[0006] However, it is difficult to find optimal measures using proxy models such as the surrogate model. For example, while a surrogate model can comprehensively evaluate measures in a short time, it is less accurate than a simulation and therefore does not necessarily find good measures.

[0007] In one aspect, an object of the present invention is to provide a measure specifying program, a measure specifying method, and an information processing device that can improve the accuracy of measure selection. [Means for solving the problem]

[0008] In the first proposal, the policy identification program causes a computer to perform a filtering process on multiple social policies under consideration based on the characteristics of a predictive model, and to identify social policies to be implemented from among the multiple social policies under consideration based on the results of a simulation of one or more social policies for which the filtering process has been performed. [Effects of the Invention]

[0009] According to one embodiment, it is possible to improve the accuracy of selecting measures. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating an information processing device that executes optimization of measures. [Figure 2] FIG. 2 is a diagram for explaining the simulation. [Figure 3] FIG. 3 is a diagram illustrating a prediction based on a simulation. [Figure 4]FIG. 4 is a diagram illustrating the surrogate model. [Figure 5] FIG. 5 is a diagram illustrating prediction using a surrogate model. [Figure 6] FIG. 6 is a diagram illustrating a policy optimization system using an information processing device according to the first embodiment. [Figure 7] FIG. 7 is a functional block diagram of the information processing apparatus according to the first embodiment. [Figure 8] FIG. 8 is a diagram illustrating the training of a surrogate model. [Figure 9] FIG. 9 is a diagram illustrating the prediction process. [Figure 10] FIG. 10 is a diagram illustrating pre-filtering using a surrogate model. [Figure 11] FIG. 11 is a diagram illustrating pre-filtering using a surrogate model. [Figure 12] FIG. 12 is a diagram for explaining the identification of measures by simulation. [Figure 13] FIG. 13 is a flowchart showing the flow of the policy optimization system. [Figure 14] FIG. 14 is a diagram illustrating a general example of extension of learning data used in a surrogate model. [Figure 15] FIG. 15 is a diagram illustrating a learning process 1 for improving the accuracy of the surrogate model. [Figure 16] FIG. 16 is a diagram illustrating a learning process 2 for improving the accuracy of the surrogate model. [Figure 17] FIG. 17 is a diagram illustrating an example using a digital twin. [Figure 18] FIG. 18 is a diagram illustrating an example of a hardware configuration. DETAILED DESCRIPTION OF THE INVENTION

[0011] The following describes in detail embodiments of the present invention, the program for identifying a policy, the method for identifying a policy, and the information processing device, with reference to the accompanying drawings. Note that the present invention is not limited to these embodiments. Furthermore, the embodiments can be combined as appropriate within a consistent range. [Example]

[0012] (Description of Information Processing Device 10) The information processing device 10 according to the first embodiment is an example of a computer that executes a policy optimization system that identifies an optimal policy from among multiple policies. In this embodiment, an example will be described in which, in a situation where a special bus operation schedule for an event is determined, an optimal policy (operation schedule) that can avoid the occurrence and expansion of congestion is identified from each operation schedule, which is a plurality of policies.

[0013] Fig. 1 is a diagram illustrating an information processing device that executes policy optimization. As shown in Fig. 1, the information processing device 10 evaluates policies 1 to N (N is a natural number) including road networks, traffic demands, and policy data using a surrogate model (hereinafter sometimes referred to as SM) or a social simulation (hereinafter sometimes referred to as simulation or SIM). Then, the information processing device 10 identifies at least one policy information item from the evaluation results.

[0014] The road network indicates the configuration of roads that are the target of the implementation of the measures, and is an example of a specified area in a certain road configuration. The traffic demand is an example of a predicted value of traffic demand in a specified area, for example, an Origin-Destination (OD) table showing the forecast value of travel demand between each point. The policy data is an example of a candidate policy to be applied to a specified area, for example, an OD table showing the operation schedule of special buses during events, which is a policy to alleviate traffic congestion in a specified area.

[0015] Here, we will explain the simulation and surrogate model of social policies that have been used for evaluating policies.

[0016] Figure 2 is a diagram explaining the simulation, and Figure 3 is a diagram explaining the predictions made by the simulation. As shown in Figure 2, the simulation takes as input the travel demand between bases (OD table) with the dimensions of departure point, arrival point, and time, and the road network, runs a simulation of the traffic volume on each link in the road network, and outputs the traffic volume on each link in each time period (OD table) as the simulation result. The road network is made up of nodes such as n1 and links that connect nodes such as e1.

[0017] Then, as shown in Figure 3, the simulation outputs evaluations of each measure, with the lowest rating being "--", the low rating being "-", the high rating being "+", and the highest rating being "++", and the highest rating, "Measure 10000", is identified.

[0018] On the other hand, Fig. 4 is a diagram explaining a surrogate model, and Fig. 5 is a diagram explaining predictions using the surrogate model. As shown in Fig. 4, the surrogate model is a trained neural network or the like, and outputs the traffic density of each link in response to inputs of parameters "road network, traffic demand" and policy variables "policy data (e.g., event traffic demand)."

[0019] As shown in Figure 5, the surrogate model then identifies "Measure 1" as the highest rating among the output evaluations of each measure, which are the lowest rating "--", the low rating "-", the high rating "+", and the highest rating "++".

[0020] Figure 5 shows the results of a simulation of the same measure (SIM results) and the judgment results using a surrogate model (SM results), but the judgment results for the optimal measure differ. Specifically, measure 1, which received the lowest evaluation in the simulation, was judged to be optimal by the surrogate model. This phenomenon is due to the low reliability of the surrogate model, which depends on the training data.

[0021] In other words, simulation can identify optimal policies, but considering all policies requires unrealistic calculation times. Surrogate models can reduce calculation times, but they are less accurate and may not necessarily select optimal policies. Therefore, it is difficult to say that either simulation or surrogate models are optimization methods. For policies such as social policies, which cannot afford to fail, it is difficult to adopt either simulation or surrogate models.

[0022] Therefore, the information processing device 10 according to the first embodiment achieves both high processing speed and high accuracy by performing pre-filtering that focuses on the characteristic of a surrogate model trained by supervised learning: "performance is low outside the distribution of training data, but performance is high within the distribution of training data." Specifically, the information processing device 10 focuses on the characteristic that, although the accuracy of a surrogate model is generally lower than that of a simulation, significant performance degradation occurs especially outside the distribution of training data. In FIG. 5, a general deterioration in accuracy is, for example, an error in determining a policy with the highest rating as high (result 10000; prediction cannot be made with sufficient detail), and a significant deterioration in performance is, for example, an error in determining a policy with the lowest rating as highest (result 1; meaningful prediction cannot be made).

[0023] FIG. 6 is a diagram illustrating a policy optimization system using an information processing device 10 according to a first embodiment. In the policy optimization system, the information processing device 10 first executes a filtering process for a plurality of social policies to be considered based on the characteristics of a prediction model. Then, the information processing device 10 identifies a social policy to be implemented from among the plurality of social policies to be considered based on the results of a simulation of one or more social policies for which the filtering process has been executed. The prediction model is a model that uses machine learning to replace simulation. The prediction model is, for example, a surrogate model.

[0024] That is, the information processing device 10 uses a machine learning model that replaces simulation to identify one or more social policies for which filtering processing has been performed from among multiple social policies under consideration. More specifically, the information processing device 10 uses a machine learning model trained to simulate simulation to identify one or more social policies for which filtering processing has been performed from among multiple social policies under consideration.

[0025] Next, the information processing device 10 identifies a social policy to be implemented from among the one or more social policies for which filtering processing has been performed based on the execution of the simulation. More specifically, the information processing device 10 executes filtering processing using the trained machine learning model, and identifies a social policy to be implemented based on a simulation of the one or more identified social policies.

[0026] As shown in Fig. 6, the information processing device 10 performs filtering processing on a plurality of measures under consideration, namely, Measure 1, Measure 2, ... Measure N, using a surrogate model or the like. Then, the information processing device 10 performs a simulation on one or more measures for which filtering processing has been performed, and identifies a social measure to be implemented from among the plurality of measures under consideration based on the results of the simulation. In Fig. 6, the one or more measures for which filtering processing has been performed are Measure 2 and Measure N. Then, the social measure to be implemented is Measure 2.

[0027] Furthermore, for example, the learning data refers to past data, the inside of the distribution of the learning data refers to an interpolation region that is within the range of the past trend, and the outside of the distribution of the learning data refers to an extrapolation region that is outside the past trend. The information processing device 10 evaluates measures that fall within the interpolation region using a surrogate model and simulation, and evaluates measures that fall within the extrapolation region using only simulation.

[0028] In this way, the information processing device 10 uses different evaluation methods for the interpolation domain and the extrapolation domain for each measure under consideration, thereby achieving both accuracy and calculation efficiency and enabling the optimal measure to be found in a short time.

[0029] (Functional configuration) 7 is a functional block diagram illustrating a functional configuration of the information processing device 10 according to the first embodiment. As illustrated in FIG.

[0030] The storage unit 12 is an example of a processing unit that stores various data and programs executed by the control unit 20, and is realized by, for example, a memory or a hard disk. The storage unit 12 stores the surrogate model 13 and simulation information 14. The storage unit 12 can also store learning data used to train the surrogate model 13.

[0031] The surrogate model 13 is a neural network that outputs the traffic density of each link in the road network in response to input of the parameters "road network, traffic demand" and the policy variable (event traffic demand). Note that the surrogate model 13 may be one that has been trained on a separate device.

[0032] The simulation information 14 is various information used to execute a social simulation. For example, the simulation information 14 is information such as mathematical formulas, parameters, and road network information. A traffic simulator, which is an example of a social simulation, is executed using this simulation information 14.

[0033] The control unit 20 is a processing unit that controls the entire information processing device 10, and is realized by, for example, a processor. The control unit 20 has a machine learning unit 30 and a prediction processing unit 40. The machine learning unit 30 and the prediction processing unit 40 are realized by electronic circuits included in the processor, processes executed by the processor, etc.

[0034] The machine learning unit 30 is a processing unit that generates the surrogate model 13. Specifically, the machine learning unit 30 performs training (learning) of the surrogate model 13 by supervised learning, and generates the surrogate model 13 that outputs the traffic density of each link in the road network.

[0035] Fig. 8 is a diagram illustrating the training of the surrogate model 13. As shown in Fig. 8, the machine learning unit 30 inputs learning data, in which the explanatory variables are "road network, traffic demand, and policy data" and the objective variable is "traffic density," into the surrogate model 13 and obtains the prediction results of the surrogate model 13. Then, the machine learning unit 30 generates the surrogate model 13 by updating the parameters of the surrogate model 13 (neural network) so as to reduce the error between the prediction results of the surrogate model 13 and the objective variable "traffic density."

[0036] To give a more specific example, the machine learning unit 30 generates a machine learning model trained to predict the calculation results of the simulator by training the relationship between the input conditions and output results of the simulator. For example, the machine learning unit 30 acquires information on "road network, traffic demand, and policy data," which are the input conditions of the traffic simulator, and "traffic density," which is the output result of the simulator. The machine learning unit 30 then generates the surrogate model 13 by performing machine learning with the "road network, traffic demand, and policy data" as the objective variables and the "traffic density" as the explanatory variable. In this way, the machine learning unit 30 generates the surrogate model 13, which is a machine learning model that replaces the simulation.

[0037] The prediction processing unit 40 has a filtering unit 41 and an identification unit 42, and is a processing unit that executes a policy optimization system that identifies an optimal policy from multiple policies. FIG. 9 is a diagram illustrating the prediction process. For example, as shown in FIG. 9, the prediction processing unit 40 performs pre-filtering using the surrogate model 13 on policies 1 to 10000 to be evaluated, thereby comprehensively evaluating the policies in a short calculation time and removing policies with clearly poor results. Thereafter, the prediction processing unit 40 runs a traffic simulator on policies 3 and 10000, which have not received poor evaluations in the surrogate model 13, to accurately evaluate the excellent policies and identify the optimal policy 10000.

[0038] The filtering unit 41 is a processing unit that executes filtering for a plurality of measures using the surrogate model 13. Specifically, the filtering unit 41 inputs a plurality of measures to the surrogate model 13, and determines whether each of the plurality of measures is in the interpolation region or the extrapolation region based on the output result of the surrogate model 13.

[0039] For example, the higher the uncertainty of the prediction result of the surrogate model, the more likely it is to be in the extrapolation region. Therefore, the filtering unit 41 measures the entropy as the uncertainty of the prediction result, and determines that a measure with high entropy is in the extrapolation region.

[0040] 10 and 11 are diagrams explaining pre-filtering by the surrogate model 13. In Fig. 10, we consider the problem of predicting the traffic density of each road on a 10-point scale (from non-congested to congested). In other words, the neural network-based surrogate model 13 solves the problem of determining which of 10 categories the traffic density will fall into when a measure is implemented.

[0041] As shown in FIG. 10, the filtering unit 41 inputs road networks, traffic demands, and policies including policy data to the surrogate model 13. The surrogate model 13 outputs a probability indicating which category the traffic density of each road belongs to. For example, when the input is x and the category is y (y∈{1, 2, , 10}) which is the congestion level, the output of the surrogate model 13 is Pit (y|x). Here, P it (y|x) indicates the probability that the congestion level of road i at time t is y when traffic demand is x.

[0042] Next, the filtering unit 41 inputs the probability value of each category into equation (1) to calculate the entropy. Then, the filtering unit 41 uses equation (2) to calculate the average entropy for all roads and all times to calculate the prediction uncertainty. Note that the inner sigma symbol in equation (2) indicates the sum for all roads, the outer sigma symbol indicates the sum for all times, T indicates the number of time increments (t∈{1,2,...,T}), and I indicates the number of roads (i∈{1,2,...,I}).

[0043]

number

number

[0044] Thereafter, the filtering unit 41 determines that the measure falls within the extrapolation domain if the uncertainty U is equal to or greater than the threshold δ, and determines that the measure falls within the interpolation domain if the uncertainty U is less than the threshold δ. The threshold δ is empirically determined depending on the problem.

[0045] 11, the filtering unit 41 inputs measures 1 to N into the surrogate model 13 and determines the uncertainty of each measure. Then, the filtering unit 41 determines that measures 1, 3, etc. fall within the interpolation region and decides to subject them to pre-filtering, and determines that measures 2, N, etc. fall within the extrapolation region and decides that they are not subject to pre-filtering but are subject to traffic simulation only.

[0046] Returning to FIG. 7, the identification unit 42 is a processing unit that identifies a policy to be implemented from among multiple policies under consideration, based on the results of one or more simulations (traffic simulators) filtered by the filtering unit 41. Specifically, for policies that fall within the interpolation region, the identification unit 42 narrows down the policies using the output of the surrogate model 13, and runs the traffic simulator only for the narrowed-down policies. On the other hand, for policies that fall within the extrapolation region, the identification unit 42 does not narrow down the policies using the output of the surrogate model 13, and runs the traffic simulator for all of them. Then, the identification unit 42 identifies the optimal policy (traffic policy) from the results of the traffic simulator.

[0047] 12 is a diagram illustrating the identification of measures by simulation. As shown in FIG. 12, the identification unit 42 determines whether measures 1 and 3, which are determined to be in the interpolation region by the surrogate model 13, are to be simulated based on the output result of the surrogate model 13. Explaining the above example, if, out of 10 levels of traffic density, measure 1 is estimated to have a traffic density of 7 and measure 3 is estimated to have a traffic density of 3, the identification unit 42 determines measure 3, which is below the threshold (traffic density 5), to be the target of simulation.

[0048] The identifying unit 42 then executes a traffic simulation for measure 3, which is highly evaluated by the surrogate model 13 among the measures determined to be in the interpolation region by the surrogate model 13, and for measures 2 and N, which are determined to be in the extrapolation region by the surrogate model 13. The identifying unit 42 then determines, from the results of the traffic simulations for measure 2, measure 3, and measure N, that measure 3, which has the lowest traffic density, in other words, the least predicted congestion, as the optimal solution. The identifying unit 42 displays measure 3, which has been determined to be the optimal solution, on a display or the like, or transmits it to a terminal designated by an administrator or the like.

[0049] (Processing flow) 13 is a flowchart showing the flow of the policy optimization system. Note that here, it is assumed that the surrogate model 13 has already been trained.

[0050] 13, the information processing device 10 determines whether to start processing. The information processing device 10 determines whether to start processing based on, for example, whether or not there is an instruction to start processing (S100). If the processing is not to be started (S100: No), the information processing device 10 receives an instruction to start processing. On the other hand, if the processing has been started (S100: Yes), the information processing device 10 acquires each policy that is the target of evaluation prediction. For example, the information processing device 10 acquires data related to the policy that is the target of evaluation prediction (S101).

[0051] Next, the information processing device 10 predicts the traffic density when each measure is executed via the surrogate model 13. For example, the information processing device 10 inputs data related to the measure to the surrogate model 13, and thereby acquires the prediction result (traffic density) output by the surrogate model 13 (S102).

[0052] Next, the information processing device 10 determines whether each measure is in an interpolation region or an extrapolation region using the prediction result (traffic density) of each measure. For example, the information processing device 10 determines whether the data related to the measure is in an interpolation region or an extrapolation region based on the prediction result (traffic density) of the surrogate model 13. More specifically, for example, the information processing device 10 determines whether the data is within or outside the range of the training data distribution of the surrogate model 13 based on the prediction probability of the prediction result (traffic density) (S103).

[0053] If the information processing device 10 determines that the region is an interpolation region (S103: interpolation region), it performs pre-filtering using the prediction result of the surrogate model for each measure determined to be an interpolation region (S104). Subsequently, the information processing device 10 obtains a final evaluation by simulation only for measures that have good prediction results from the surrogate model 13 (S105).

[0054] On the other hand, if the information processing device 10 determines that the area is an extrapolation area (S103: extrapolation area), it obtains a final evaluation by simulation for each measure determined to be in the extrapolation area without performing pre-filtering using the prediction results of the surrogate model (S106).

[0055] Then, the information processing device 10 identifies the optimum measure for the requirement such as low traffic density from the simulation results of each measure (S107).

[0056] (effect) As described above, the information processing device 10 can improve computational efficiency by pre-filtering candidate measures taking advantage of the characteristics of the surrogate model 13, and can also perform highly accurate predictions taking advantage of the characteristics of simulation, thereby making it possible to find the optimal measure in a short period of time.

[0057] Furthermore, the information processing device 10 makes use of the characteristics of the surrogate model 13 to determine the interpolation region and the extrapolation region, and therefore can make accurate determinations without relying on the type of measure or other information, thereby preventing the search time for the optimal measure from becoming longer and accuracy from decreasing depending on the type of measure, etc.

[0058] Furthermore, the information processing device 10 can achieve both high speed and high accuracy by evaluating measures in multiple stages, so that it can be applied with confidence to social measures that often have special circumstances where failure is not an option. [Example]

[0059] However, the determination of the interpolation region and the extrapolation region by the surrogate model 13 largely depends on the learning data of the surrogate model 13. If the extrapolation region is too wide, the range evaluated in the simulation will be wide, and the reduction rate of the calculation time may be small. Therefore, in Example 2, a technical measure for widening the interpolation region of the surrogate model 13 will be described.

[0060] First, we will explain commonly used training data extensions. Figure 14 is a diagram explaining a typical example of extension of training data used in a surrogate model 13. (1) in Figure 14 shows an example of a training dataset design. In general, it is possible to broaden the training target and widen the interpolation region by broadening the distribution of the training dataset. For example, training data for predictable scenarios such as Golden Week and New Year's holidays is easy to collect because a large amount of past data exists.

[0061] However, data on various policies often does not exist, making this method difficult to consider appropriate for social policies that often have special circumstances that cannot tolerate failure. For example, no matter how much past data (actual data) is researched, it will be impossible to understand the details of new train stations and shuttle bus operations for the Olympics, and since such past data cannot be collected, it cannot be said that the interpolation area can be expanded appropriately.

[0062] Furthermore, (2) in Figure 14 shows data augmentation using synthetic data. When there is a limit to how much the collection of past data such as (1) in Figure 14 can broaden the tail of the dataset's distribution, it is possible to augment the data using artificially created data (synthetic data). For example, traffic demands and measures that do not exist in the actual data are input into a traffic simulator to calculate the traffic state (traffic density). Then, synthetic data is generated with the input "traffic demands and measures" as explanatory variables (features) and the output traffic state as the correct label.

[0063] However, as explained in the first embodiment, generating data through simulation requires high calculation costs. In other words, generating a sufficient amount of synthetic data takes a huge amount of time, which is not realistic.

[0064] (Solution 1) Therefore, the information processing device 10 according to the second embodiment generates synthetic data using a theoretical model whose calculation cost is lower than that of a simulation model, and uses the synthetic data as training data for the surrogate model 13.

[0065] 15 is a diagram illustrating learning process 1 for improving the accuracy of a surrogate model. As shown in FIG. 15, the machine learning unit 30 of the information processing device 10 uses a theoretical model to construct learning data, which is extended data (S201), and then uses the learning data to perform pre-learning of the surrogate model 13 (S202).

[0066] Then, the machine learning unit 30 constructs learning data using actual data, which is past data prepared in advance (S203), and uses the learning data to fine-tune the surrogate model 13 learned with the extended data (S204). After that, the information processing device 10 uses the fine-tuned surrogate model 13 to execute the policy optimization system shown in Fig. 13 (S205).

[0067] For example, when reproducing traffic, a theoretical model of traffic flow is used. Specifically, the machine learning unit 30 calculates the equilibrium flow under a predetermined traffic demand using a Static-Traffic-Assignment model. Then, the machine learning unit 30 generates training data (theoretical values) using the input traffic demand as an explanatory variable (feature) and the equilibrium flow as correct answer data. The machine learning unit 30 performs pre-training using the theoretical values, and then trains using actual traffic data as the difference between them. In this way, the information processing device 10 can quickly calculate the equilibrium flow using the theoretical model and improve accuracy in areas where actual traffic data is insufficient.

[0068] (Solution 2) For example, the information processing device 10 according to the second embodiment focuses on the rules of the target domain of the surrogate model 13 and constructs a model that follows not only the learning data but also the rules. Fig. 16 is a diagram illustrating a learning process 2 for improving the accuracy of the surrogate model.

[0069] 16, the machine learning unit 30 sets a special error function that evaluates not only the degree of agreement with the training data but also the degree of agreement with the rules of the target domain (S301), constructs training data using actual data that is past data prepared in advance (S302), and uses the training data to train the surrogate model 13 using the special error function and the training data (S303). Thereafter, the information processing device 10 uses the trained surrogate model 13 to execute the policy optimization system shown in FIG. 13 (S304).

[0070] For example, when reproducing traffic, attention is paid to the laws related to traffic flow. Specifically, the machine learning unit 30 focuses on whether each intersection satisfies the flow conservation law, and designs an error function that increases the penalty the more the output result of the surrogate model 13 deviates from the training data and the more it deviates from the flow conservation law. The flow conservation law is a law that states that the sum of the inflow and outflow traffic at each intersection is zero. In this way, the machine learning unit 30 can build a model that also obeys the flow conservation law by training the surrogate model 13 using this error function. As a result, the information processing device 10 can align behavior in areas where actual data is insufficient with experience, thereby achieving high accuracy.

[0071] By implementing the above-described solution 1 or solution 2 for widening the interpolation region, the range to which pre-filtering using the surrogate model 13 can be applied is widened, and the calculation efficiency can be further improved compared to the first embodiment. [Example]

[0072] Although the embodiments of the present invention have been described above, the present invention may be embodied in various different forms other than the above-described embodiments.

[0073] (Numbers, etc.) The numerical values, data, etc. used in the above embodiments are merely examples and can be changed as desired. Furthermore, the process flow described in each flowchart can also be changed as appropriate within a consistent range.

[0074] (OD table) In the above embodiment, an example was explained in which an OD table with dimensions of "departure point," "arrival point," and "time" was used as the transportation demand, and an OD table with dimensions of "intersection," "number of vehicles passing through the intersection," and "time" as the simulation results (traffic conditions), but this is not limited to this. For example, various data formats can be adopted, such as a table format, any data format, a three-dimensional graphical format, or an XML (Extensible Markup Language) format, instead of a matrix format like an OD table.

[0075] (Digital Twin) Furthermore, the information processing device 10 can also execute a simulation using a digital twin that virtually reproduces the traffic flow at the same time in the real world, as an example of a traffic simulator that executes a traffic simulation.

[0076] FIG. 17 is a diagram illustrating an example using a digital twin. As shown in FIG. 17, the information processing device 10 reproduces real-world data such as traffic congestion information, weather information, and event information for a relevant area on a digital twin. More specifically, the information processing device 10, for example, synchronizes time with the real space in a virtual space to create a digital twin that virtually reproduces traffic flow at the same time in the real world. The information processing device 10 then performs a simulation of each measure and, using the results of each simulation, can identify optimization measures using a method similar to that of Example 1. As a result, the information processing device 10 can identify measures that are highly reproducible, highly reliable, and highly feasible.

[0077] (system) The information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings may be changed arbitrarily unless otherwise specified.

[0078] Furthermore, the specific form of distribution or integration of the components of each device is not limited to that shown in the figure. For example, the machine learning unit 30 and the prediction processing unit 40 may be integrated. That is, all or some of the components may be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions of each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware using wired logic.

[0079] Furthermore, all or any part of the processing functions performed by each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware using wired logic.

[0080] (Hardware) Fig. 18 is a diagram illustrating an example of a hardware configuration. As shown in Fig. 18, an information processing device 10 includes a communication device 10a, an HDD (Hard Disk Drive) 10b, a memory 10c, and a processor 10d. The components shown in Fig. 18 are connected to each other via a bus or the like.

[0081] The communication device 10a is a network interface card or the like, and communicates with other devices. The HDD 10b stores programs and DBs that operate the functions shown in FIG.

[0082] The processor 10d reads out a program that executes the same processes as the respective processing units shown in FIG. 7 from the HDD 10b or the like and loads it into the memory 10c, thereby operating a process that executes the respective functions described in FIG. 7 or the like. For example, this process executes the same functions as the respective processing units of the information processing device 10. Specifically, the processor 10d reads out a program having the same functions as the machine learning unit 30, the prediction processing unit 40, or the like from the HDD 10b or the like. Then, the processor 10d executes a process that executes the same processes as the machine learning unit 30, the prediction processing unit 40, or the like.

[0083] In this way, the information processing device 10 operates as an information processing device that executes policy optimization by reading and executing a program. The information processing device 10 can also realize functions similar to those of the above-described embodiment by reading the program from a recording medium using a medium reading device and executing the read program. Note that the program in this other embodiment is not limited to being executed by the information processing device 10. For example, the above-described embodiment may also be applied in the same way to cases where another computer or server executes the program, or where these execute the program in cooperation with each other.

[0084] This program may be distributed via a network such as the Internet. Alternatively, this program may be recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, a magneto-optical disk (MO), or a digital versatile disk (DVD), and may be read out from the recording medium and executed by a computer. [Explanation of symbols]

[0085] 10. Information processing equipment 11 Communications Department 12 Storage section 13 Surrogate Model 14 Simulation Information 20 Control Unit 30 Machine Learning Department 40 Prediction processing unit 41 Filtering section 42 Specific part

Claims

1. On the computer, Based on the characteristics of the predictive model, a filtering process is performed for multiple social policies under consideration, Identifying a social policy to be implemented from among the plurality of social policies to be considered based on a result of the simulation of one or more social policies for which the filtering process has been executed. A program that specifies the measures to be taken to execute the process.

2. the predictive model is a model that uses machine learning to replace simulation, The filtering process uses a machine learning model that replaces the simulation to identify one or more social policies on which the filtering process has been performed from among the plurality of social policies under consideration; the identifying process identifies the social measures to be implemented based on the results of the simulation for the one or more social measures for which the filtering process has been executed. The policy specification program according to claim 1 .

3. Using the prediction model trained by machine learning, from among the social policies under consideration, obtain policies related to the learning data used to train the prediction model; Evaluating measures related to the acquired learning data using the prediction model; Conducting a simulation of the social policy based on the evaluation results of the prediction model. The policy specification program according to claim 1 .

4. Analyzing the distribution of learning data of the prediction model created from past data; determining whether the social policy to be evaluated is within or outside the range of the training data distribution of the prediction model; If the prediction result is within the range of the learning data distribution, after evaluating the social measures using the prediction model, re-evaluating only the measures that have a high prediction result using the prediction model by simulating the social measures; If the distribution of the learning data is outside the range, the social policy is evaluated by simulating the social policy. The policy specification program according to claim 1 .

5. generating theoretically calculated augmented data using the theoretical model; Performing machine learning of a surrogate model using the augmented data as training data; causing the computer to execute a process of fine-tuning the surrogate model that has been trained using the augmented data, using training data generated from past data; The filtering process includes: performing the filtering process on the plurality of social policies under consideration based on the characteristics of the surrogate model after the fine tuning; The policy specification program according to claim 1 .

6. the plurality of social policies are policy data for preventing the occurrence and expansion of congestion on links between nodes in a road network in a predetermined area; The theoretical model is a theoretical model for calculating an equilibrium flow under a predetermined traffic demand. The measure specifying program according to claim 5.

7. causing the computer to execute a process of performing machine learning of a surrogate model using training data generated from past data and an error function that evaluates the degree of agreement between the training data and the rules of the target domain; The filtering process includes: performing the filtering process on the plurality of social policies under consideration based on the characteristics of the surrogate model trained using the training data and the error function; The policy specification program according to claim 1 .

8. the plurality of social policies are policy data for preventing the occurrence and expansion of congestion on links between nodes in a road network in a predetermined area; The error function is an error function that satisfies the degree of agreement with the training data and the flow conservation law regarding traffic outflow to intersections in the road network. The measure specifying program according to claim 7.

9. The process of performing the filtering process includes: Acquire data on multiple social policies that are candidates for application in a specified area, performing the filtering process on the acquired plurality of pieces of social policy data using the prediction model trained by machine learning; The identifying process includes: Using a simulator, predict the state of the predetermined area, which indicates the result of simulating the one or more social policy data on which the filtering process has been performed; Identifying social policy data to be applied to the predetermined area based on the predicted state of the predetermined area. The policy specification program according to claim 1 .

10. the predictive model is a machine learning model that replaces the simulation of a traffic simulator, The social policy data is data indicating transportation policies to be applied to the predetermined area and policies to reduce congestion in the predetermined area, The simulation of the social policy data uses the traffic simulator to predict the degree of congestion in the predetermined area; Identifying social policy data to be applied to the predetermined area based on the calculated congestion degree. The measure specifying program according to claim 9.

11. The prediction process includes: Calculate the degree of congestion in the specified area based on the results of a digital twin simulation that recreates real-world traffic flow at the same time in a virtual space. The measure specifying program according to claim 9.

12. The computer Based on the characteristics of the predictive model, a filtering process is performed for multiple social policies under consideration, Identifying a social policy to be implemented from among the plurality of social policies to be considered based on a result of the simulation of one or more social policies for which the filtering process has been executed. The method of identifying the measures to be taken.

13. Based on the characteristics of the predictive model, a filtering process is performed for multiple social policies under consideration, Identifying a social policy to be implemented from among the plurality of social policies to be considered based on a result of the simulation of one or more social policies for which the filtering process has been executed. An information processing device having a control unit.

Images