JP6785741B2 - Optimizer, traffic signal control system, parameter search device, optimization method, and program - Google Patents

Description

The present invention relates to an optimization device, a traffic signal control system, a parameter search device, an optimization method, and a program, and particularly an optimization device for optimizing machine learning and simulation parameters, a traffic signal control system, and a parameter search device. , Optimization methods, and programs.

In recent years, the importance of machine learning and simulation has increased. As an example of a technique using machine learning or simulation, there is a technique of moving a large number of cars on a simulation to reproduce urban traffic (Non-Patent Document 1). The performance of machine learning varies depending on its hyperparameters. In addition, the output of the simulation also fluctuates depending on the parameters. Here, hyperparameters or parameters are collectively referred to as parameters.

The parameters need to be optimized to the appropriate values. Optimization is performed so that the pre-specified index is the best. Acquiring the value of the index is called evaluation here. With the sophistication of machine learning and simulation in recent years, the time required for one evaluation is increasing. Therefore, in order to reduce the time required for optimizing the parameters, a technique for reducing the number of evaluations has been proposed (Non-Patent Document 2).

Krajzewicz, D., Brockfeld, E., Mikat, J., Ringel, J., Rossel, C., Tuchscheerer, W., Wagner, P., and Wosler, R .: Simulation of modern Traffic Lights Control Systems using the open source Traffic Simulation SUMO, Proceedings of the 3rd Industrial Simulation Conference 2005, pp. 299-302 (2005). Shahriari, B., Swersky, K., Wang, Z., Adams, RP and Freitas, de N .: Taking the human out of the loop: A review of bayesian optimization, Proceedings of the IEEE, Vol. 104, No. 1, pp. 148-175 (2016). Papageorgiou, M., Diakaki, C., Dinopoulou, V., Kotsialos, A. and Wang, Y .: Review of road traffic control strategies, Proceedings of the IEEE, Vol. 91, No. 12, pp. 2043-2067 (2003).

When performing the above optimization, the space that the parameter can take may be divided into several subspaces. In that case, it is necessary to limit the space that the parameter can take to a certain subspace, and then select the optimum parameter from the subspace. Hereinafter, this subspace is simply referred to as a parameter space.

Traffic signal control is an example of the above optimization. In traffic signal control, a method is used in which a plan for switching signal light colors is created for one cycle, and signal control is performed according to the repetition of the plan. The plan is uniquely determined by specifying parameters. In the process of optimizing this parameter, simulation is used to execute the evaluation (Non-Patent Document 1).

Here, in traffic signal control, it is known that system control of a plurality of signals is effective in alleviating traffic congestion (Non-Patent Document 3). In this case, in order to determine the range for system control, all the signals to be controlled belong to any one subarea. By determining the sub-area to which each signal belongs, the parameter space in which the parameter can be taken is uniquely determined. Therefore, it is necessary to determine the parameter space by determining the sub-area to which each signal belongs, and then determine the optimum parameter in the parameter space.

The present invention has been made in view of the above points, and an object of the present invention is to provide an optimization device, an optimization method, and a program capable of optimizing parameters with a small number of evaluations.

Another object of the present invention is to provide a traffic signal control system capable of optimizing signal parameters with a small number of traffic simulations.

Another object of the present invention is to provide a parameter search device capable of obtaining optimized parameters.

The optimization device according to the present invention is an optimization device that optimizes parameters used when calculating with evaluation data as an input, and is a result of the calculation based on the evaluation data and the parameters. An evaluation unit that calculates an index for evaluating the above, a plurality of parameter spaces that divide the space that the parameter can take, an optimization unit that optimizes the parameters, processing by the evaluation unit, and the optimization unit. A model for predicting an index for the parameter space based on the parameter and the index, including an output unit that outputs an optimized parameter obtained by repeating the process. Is trained, and for each of the plurality of parameter spaces, the first index which is an index for the parameter space is predicted based on the first model, and the predicted first index is used. With a parameter space optimization unit that selects the parameter space to be evaluated next by the evaluation unit based on the value calculated based on the value, and a model for predicting the index for the parameter based on the parameter and the index. A second model is learned, and among the parameters included in the parameter space selected by the parameter space optimization unit, each of the parameters that have not been evaluated in the past is based on the second model. A parameter optimization unit that predicts a second index, which is an index for the parameter, and selects a parameter to be evaluated next by the evaluation unit based on a value calculated based on the predicted second index. Consists of including.

Further, the optimization method according to the present invention is an optimization method for optimizing the parameters when the evaluation data is input and output using the parameters, and the evaluation unit uses the evaluation data and the parameters. Based on, the step of calculating the index for evaluating the output, the step of optimizing the plurality of parameter spaces in which the optimizer divides the space that the parameter can take, and the step of optimizing the parameter, and the output section The step of optimizing by the optimizing unit includes parameter space optimization, including a step of outputting an optimized parameter obtained by repeating the processing by the evaluation unit and the processing by the optimizing unit. The unit learns a first model, which is a model for predicting an index for the parameter space, based on the parameter and the index, and for each of the plurality of the parameter spaces, based on the first model. A step of predicting a first index, which is an index for the parameter space, and selecting a parameter space to be evaluated next by the evaluation unit based on a value calculated based on the predicted first index, and a parameter. The optimization unit learns a second model, which is a model for predicting an index for the parameter based on the parameter and the index, and is included in the parameter space selected by the parameter space optimization unit. Of the parameters, for each of the parameters that have not been evaluated in the past, a second index that is an index for the parameter is predicted based on the second model, and calculation is performed based on the predicted second index. A step of selecting a parameter to be evaluated next by the evaluation unit based on the value to be evaluated is included.

According to the optimization device and the optimization method according to the present invention, the evaluation unit calculates an index for evaluating the output based on the evaluation data and the parameter, and the optimization unit determines the space that the parameter can take. The divided parameter space and the parameters are optimized, and the output unit outputs the optimized parameters obtained by repeating the processing by the evaluation unit and the processing by the optimization unit.

Then, in the processing by the optimization unit, the parameter space optimization unit learns the first model, which is a model for predicting the index for the parameter space based on the parameter and the index, and for each of the plurality of parameter spaces. , Predict the first index, which is an index for the parameter space, based on the first model, and select the parameter space to be evaluated next by the evaluation unit based on the value calculated based on the predicted first index. Then, the parameter optimization unit learns a second model, which is a model for predicting an index for the parameter based on the parameter and the index, and the parameter included in the parameter space selected by the parameter space optimization unit. Of these, for each of the parameters that have not been evaluated in the past, the second index, which is an index for the parameter, is predicted based on the second model, and based on the value calculated based on the predicted second index. Then, the evaluation department selects the parameter to be evaluated next.

In this way, based on the model for predicting the index for the parameter space, the first index, which is the index for the parameter space, is predicted, and based on the value calculated based on the predicted first index, the next The parameter space to be evaluated is selected, and among the parameters included in the selected parameter space, each of the parameters that have not been evaluated in the past is an index for the parameter based on a model for predicting the index for the parameter. Predict the second index, select the parameter to be evaluated next based on the value calculated based on the predicted second index, and output the output based on the evaluation data and the selected parameter. By repeating the calculation of the index to be evaluated, the parameters can be optimized with a small number of evaluations.

Further, in the optimizing device according to the present invention, the parameter space optimizing unit predicts the evaluation of the parameter space for each of the plurality of the parameter spaces by using the first model, and the parameter space is evaluated. The first acquisition function with the prediction of evaluation as a variable is calculated, the parameter space in which the value of the first acquisition function is maximized is selected as the parameter space to be evaluated next by the evaluation unit, and the parameter optimization is performed. Using the second model, the evaluation unit is next for each of the parameters included in the parameter space selected by the parameter space optimization unit that has not been evaluated in the past. The parameter that predicts the evaluation for the parameter space to be evaluated, calculates the second acquisition function with the prediction of the evaluation for the parameter that has not been evaluated in the past as a variable, and maximizes the value of the second acquisition function. Can be selected as the parameter to be evaluated next by the evaluation unit.

Further, in the optimization device according to the present invention, the first model and the second model can be considered to be a probabilistic model using a Gaussian process.

Further, in the optimization device according to the present invention, the parameter space optimization unit includes the parameter space to which the parameter used for evaluation by the evaluation unit belongs, the number of times of evaluation using the parameter belonging to the parameter space, and the number of times. The first model was learned using the index obtained by the evaluation unit, and the parameter optimization unit was obtained by the parameter used when the evaluation unit performed the evaluation and the evaluation unit. The second model can be trained using the index.

The traffic signal control system according to the present invention includes a control device that controls a plurality of traffic signals, and an optimization device that optimizes signal parameters used by the control device by inputting evaluation data necessary for calculating traffic conditions. A traffic signal control system including, the control device uses an input unit that receives an input of a situation and a signal parameter obtained by the optimization device with the traffic situation as an input, and uses the plurality of traffic. The optimization device includes a control unit that controls a traffic light, and based on the evaluation data and the signal parameter, the optimization device calculates the traffic condition using the signal parameter with the evaluation data as an input. An evaluation unit that calculates an index for evaluating the calculated traffic condition, a plurality of parameter spaces that divide the space that the signal parameter can take, an optimization unit that optimizes the signal parameter, and the evaluation unit. The optimization unit includes an output unit that outputs an optimized parameter obtained by repeating the processing by the optimization unit and the optimization unit, based on the signal parameter and the index. A first model, which is a model for predicting an index for a parameter space, is learned, and for each of the plurality of the parameter spaces, a first index, which is an index for the parameter space, is predicted based on the first model. Then, based on the predicted value calculated based on the first index, the parameter space optimization unit selects the parameter space to be evaluated next by the evaluation unit, and based on the signal parameter and the index. A second model, which is a model for predicting an index for the signal parameter, is learned, and among the signal parameters included in the parameter space selected by the parameter space optimization unit, the above-mentioned signal parameters that have not been evaluated in the past. For each of the signal parameters, the second index, which is an index for the signal parameter, is predicted based on the second model, and the evaluation unit is based on the value calculated based on the predicted second index. Is configured to include a parameter optimization unit that selects the signal parameters to be evaluated next.

According to the traffic signal control system according to the present invention, the first index, which is an index for the parameter space, is predicted based on the model for predicting the index for the parameter space, and the calculation is performed based on the predicted first index. Based on the values to be evaluated, the parameter space to be evaluated next is selected, and among the signal parameters contained in the selected parameter space, for each of the signal parameters that have not been evaluated in the past, the index for the parameter is predicted. The second index, which is an index for the signal parameter, is predicted based on the model of, and the signal parameter to be evaluated next is selected based on the value calculated based on the predicted second index. By repeating the calculation of the traffic condition based on the selected signal parameter, the signal parameter can be optimized with a small number of traffic simulations.

The parameter search device according to the present invention has the parameters based on a first model for predicting a first index which is an index for the parameter space for each of a plurality of parameter spaces that divide the space that the parameter can take. Included in the parameter space optimization unit that predicts the first index, which is an index for space, and selects the parameter space based on the predicted first index, and the parameter space selected by the parameter space optimization unit. For each of the parameters, the second index, which is an index for the parameter, is predicted based on the second model for predicting the second index, which is an index for the parameter, and based on the second index. A parameter optimization unit for selecting parameters and an output unit for outputting the parameters selected by the parameter optimization unit are included.

According to the parameter search device according to the present invention, for each of a plurality of parameter spaces that divide the space that the parameters can take, based on a first model for predicting a first index that is an index for the parameter space. The first index, which is an index for the parameter space, is predicted, the parameter space is selected based on the predicted first index, and each of the parameters contained in the selected parameter space is the index for the parameter. Predicting a second index, which is an index for a parameter, based on a second model for predicting the second index, selecting a parameter based on the second index, and outputting the selected parameter. Allows the optimized parameters to be obtained.

The program according to the present invention is a program for functioning as each part of the above-mentioned optimization device.

According to the optimization device, the optimization method, and the program of the present invention, the first index, which is an index for the parameter space, is predicted and predicted based on the model for predicting the index for the parameter space. Based on the values calculated based on the metric, select the parameter space to evaluate next, and for each of the parameters contained in the selected parameter space, the metric for the parameter based on the model for predicting the metric for the parameter. The second index is predicted, the parameter to be evaluated next is selected based on the value calculated based on the predicted second index, and the evaluation data and the selected parameter are used. Since the calculation of the index for evaluating the output is repeated, the parameter space can be optimized and the parameters in the parameter space can be optimized with a small number of evaluations.

According to the traffic signal control system of the present invention, the first index, which is an index for the parameter space, is predicted based on the model for predicting the index for the parameter space, and the calculation is performed based on the predicted first index. To select the parameter space to be evaluated next based on the values, and to predict the index for the parameter for each of the signal parameters included in the selected parameter space that have not been evaluated in the past. The second index, which is an index for the signal parameter, is predicted based on the model, and the signal parameter to be evaluated next is selected based on the value calculated based on the predicted second index. By repeating the calculation of the traffic condition based on the selected signal parameter, the signal parameter can be optimized with a small number of traffic simulations.

According to the parameter search device of the present invention, for each of a plurality of parameter spaces that divide the space that the parameter can take, the parameter space is based on a first model for predicting a first index that is an index for the parameter space. The first index, which is an index for, is predicted, the parameter space is selected based on the predicted first index, and for each of the parameters contained in the selected parameter space, the second index, which is an index for the parameter, is selected. By predicting the first index, which is an index for the parameter, based on the second model for predicting the index, selecting the parameter based on the second index, and outputting the selected parameter. Optimized parameters can be obtained.

It is a block diagram which shows the structure of the traffic signal control system which concerns on embodiment of this invention. It is a figure which shows a part example of the information stored in the parameter space storage part which concerns on embodiment of this invention. It is a figure which shows a part example of the information stored in the parameter storage part which concerns on embodiment of this invention. It is a figure which shows the relationship between the number of searches in the example of the experimental result which concerns on embodiment of this invention, and the congestion loss time for each parameter space. It is a flowchart which shows the optimization processing routine in the optimization apparatus which concerns on embodiment of this invention.

<Structure of Traffic Signal Control System According to Embodiment of the Present Invention>
Hereinafter, embodiments of the present invention will be described with reference to the drawings.

In the present embodiment, a case where the optimization device according to the present embodiment is applied to the optimization of the signal parameters s using the traffic simulation in the traffic signal control will be described.

In the present embodiment, the traffic signal control is performed by the control device. In the traffic signal control, a plan for switching the color of the signal light is created for one cycle, and the signal is controlled according to the repetition of the plan. This plan is uniquely determined by specifying the signal parameter s. The process of optimizing the signal parameter s is performed by the optimization device according to the present embodiment.

In the present embodiment, in order to determine the range for system control of a plurality of signals, all the signals to be controlled are assigned to any one subarea based on the subarea. By determining the sub-area to which each signal belongs, the parameter space i that can be taken by the signal parameter s is uniquely determined.

FIG. 1 is a block diagram showing a configuration of a traffic signal control system according to an embodiment of the present invention.

The traffic signal control system 1 according to the present embodiment includes an optimization device 10, a control device 50, and a plurality of traffic signals (not shown).

The optimization device 10 according to the present embodiment is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing an optimization processing routine described later, and is functionally as shown below. It is configured in.

As shown in FIG. 1, the optimization device 10 according to the embodiment of the present invention includes an optimization unit 100, an evaluation data storage unit 200, an evaluation unit 300, and an output unit 400. ..

The optimization unit 100 optimizes the parameter space i obtained by dividing the space that the signal parameter s can take and the signal parameter s.

Specifically, the optimization unit 100 includes a parameter space storage unit 110, a parameter space optimization unit 120, a parameter storage unit 130, and a parameter optimization unit 140.

The parameter space storage unit 110 stores data in the evaluation performed in the past.

Specifically, the parameter space storage unit 110 is a feature vector of the parameter space i selected in the traffic simulations t and t times.

, The number of times the parameter space i was selected by the tth time

And the number of times you chose some parameter space

Vector with

, Value calculated from the set of indicators in the traffic simulation performed by selecting the parameter space i up to the tth time.

Is. Feature vector of parameter space i

Can be anything. As an example, there is a binary vector expressing whether or not two signals belong to the same parameter space. Value calculated from a set of indicators

May be a value calculated by any calculation. As an example of calculation, there are a value obtained by taking the minimum value of a set of indicators, a value obtained by rearranging a set of indicators, and the like. Feature vector of parameter space i at t = 1, 2, ...

Set of X, vector

Set of Z,

The set of is represented as Y. An example of a part of the information stored in FIG. 2 is shown.

Further, the parameter space storage unit 110 includes the parameter space i and its feature vector.

Correspondence table of is also stored in advance.

Further, the parameter space storage unit 110 includes the parameter space i from the evaluation unit 300.

When we get, the parameter space i and its feature vector

From the correspondence table of, the feature vector of the parameter space it _{+ 1}

To get. Then, the parameter space storage unit 110

To

Update as and from there

To get.

Then, the parameter space storage unit 110

To X,

Was obtained from the evaluation unit 300 in Z.

To Y, respectively.

The parameter space optimization unit 120 is an index for the parameter space based on the parameter space i and its feature vector X, the vector Z whose elements are the number of times each parameter space is selected and the number of times some parameter space is selected, and the index Y. The first model, which is a model for predicting, is trained, and for each of a plurality of parameter spaces, the first index, which is an index for the parameter space, is predicted and predicted based on the first model. The evaluation unit 300 selects the parameter space to be evaluated next based on the value calculated based on the index of.

Specifically, the parameter space optimization unit 120 includes a parameter space model learning unit 122 and a parameter space selection unit 124.

The parameter space model learning unit 122 learns the first model based on X, Z, and Y stored in the parameter space storage unit 110.

First, the parameter space model learning unit 122 starts with the parameter space storage unit 110 to X, Z, and Y, the parameter space i, and its feature vector.

Get the correspondence table of.

Then, the parameter space model learning unit 122 learns the first model based on X, Z, and Y. As an example of the first model, there is a Gaussian process which is a stochastic model (Reference 1).

[Reference 1] Rasmussen, C.E. and Williams, C.K.I .: Gaussian processes for machine learning, MIT Press (2006).

By using regression by Gaussian process, it is possible to infer an unknown index y as a probability distribution in the form of a normal distribution for inputs x and z for which traffic simulation is not performed. Here, the kernel function of the Gaussian process is separated into a kernel related to x and a kernel related to z as shown in the following equation (1).

Also, the kernel for x can be anything. As an example, there is a Gaussian kernel represented by the following equation (2) (Non-Patent Document 2).

Here, θ is a parameter that takes a real number. As an example of θ, a point-estimated value is used as the value that maximizes the peripheral likelihood of the Gaussian process (Reference 1).

Also, any kernel related to z may be used. As an example, a kernel represented by the following equation (3) is used.

here,

Is the inner product, a, b, and

Is a parameter that takes a real number. As an example of a and b, a = 100 and b = 1 can be set. Also,

As an example, a point-estimated value can be used as the value that maximizes the peripheral likelihood of the Gaussian process (Reference 1).

Then, the parameter space model learning unit 122 transmits the learned first model to the parameter space selection unit 124.

The parameter space selection unit 124 selects the parameter space it _{+ 1 for} which the traffic simulation is to be performed next, based on the first model received from the parameter space model learning unit 122.

Specifically, first, the parameter space selection unit 124 performs Gaussian process regression, and then represents the degree to which the parameter space should be traffic-simulated.

, Parameters for all parameter spaces

Calculate against.

here,

Is the addition of the change in value due to the selection of the parameter space i in the future to zi _{and t} . here,

As an example, there is the following equation (4) which is z _{i, t} when the parameter space i is selected in the following traffic simulation.

Here, T means transpose.

Further, as an example of the acquisition function, there is an upper confidence bound represented by the following formula (5) (Non-Patent Document 2).

here,

,as well as

Are the mean and variance regressed in the Gaussian process, respectively, and β _{t + 1} is a parameter. For example

Can be.

Then, the parameter space it _{+ 1} that maximizes the acquisition function represented by the following equation (6) is selected and transmitted to the parameter selection unit 144.

Like the parameter space storage unit 110, the parameter storage unit 130 stores the results of traffic simulations performed in the past, reads data according to a request, and transmits the corresponding data to the parameter model learning unit 142.

Data to be stored in the parameter storage unit 130 is the number of traffic simulation t, the signal parameter selected in the t-th traffic simulation s _t, the t-th traffic simulation index l _t. The sets of _st and l _t at t = 1, 2, ... Are represented as S and L, respectively. An example of a part of the information stored in FIG. 3 is shown.

Further, the parameter storage unit 130 is a space of the parameter space i and a signal parameter space that can be taken in the parameter space i.

The correspondence table of is also stored.

Further, when the parameter storage unit 130 acquires _{st + 1} and l _{t + 1} from the evaluation unit 300, the parameter storage unit 130 adds _{st + 1} and l _{t + 1} to S and L, respectively.

The parameter optimization unit 140 learns a second model, which is a model for predicting the index l for the parameter, based on the signal parameter s and the index, and enters the parameter space i selected by the parameter space optimization unit 120. For each of the included signal parameters, the evaluation unit 300 predicts the second index, which is an index for the signal parameter, based on the second model, and the evaluation unit 300 predicts the value calculated based on the predicted second index. Then select the parameters to evaluate.

Specifically, the parameter optimization unit 140 includes a parameter model learning unit 142 and a parameter selection unit 144.

The parameter model learning unit 142 learns the second model from S and L received from the parameter storage unit 130.

Specifically, first, the parameter model learning unit 142 is a space of signal parameters that can be taken from the parameter storage unit 130 to S, L, the parameter space i, and the parameter space i.

Get the correspondence table of.

Next, the parameter model learning unit 142 learns the second model from S and L. Here, too, there is a Gaussian process as an example of the second model, and an unknown index l is inferred as a probability distribution in the form of a normal distribution for the input of the signal parameter s for which the traffic simulation is not performed. The kernel for s may be anything, but as an example, the Gaussian kernel represented by the above equation (2) is used. Here, it is usually modeled by one Gaussian process, but if there are a plurality of indexes, it can be modeled by a plurality of Gaussian processes.

Further, when learning signal parameters in different parameter spaces in a Gaussian process, it is possible to supplement the parameter deficiency by using a method such as arc-GP (Reference 2).

[Reference 2] Swersky, K, Duvenaud, D, Snoek, J, Hutter, F, and Osborne, M .: Raiders of the Lost Architecture: Kernels for Bayesian Optimization in Conditional Parameter Spaces, Proceedings of NIPS workshop on Bayesian Optimization in theory and practice (2013).

Then, the parameter space model learning unit 122 describes the learned second model, the parameter space i, and the space of the signal parameters that can be taken in the parameter space i.

The correspondence table of is transmitted to the parameter selection unit 144.

The parameter selection unit 144 selects the signal parameter s _{t + 1 for} which the traffic simulation is performed next, based on the second model received from the parameter model learning unit 142 and the parameter space it _{+ 1} for which the traffic simulation is performed next.

Specifically, the parameter selection unit 144 performs Gaussian process regression, and obtains the acquisition function α (s) in the signal parameter s from the signal parameters belonging to the parameter space it _{+ 1} acquired from the parameter space selection unit 124 in the past. Calculate for the unevaluated parameter s. Here, too, the upper confidence bound can be used as an example of the acquisition function (Non-Patent Document 2).

Then, the parameter selection unit 144 selects the signal parameter _{st + 1} that maximizes the acquisition function represented by the following equation (7). Here, any method may be used for maximization.

After that, the parameter selection unit 144 transmits the selected signal parameter s _{t + 1} and the parameter space it _{+ 1} to the evaluation unit 300.

The evaluation data storage unit 200 stores evaluation data, which is data necessary for performing a traffic simulation.

Here, if the evaluation data is data necessary for performing traffic simulation, for example, the shape of the road, the speed limit of each road, the number of vehicles, the approach time of each vehicle to the traffic simulation section, and those The route of the vehicle, the start time and end time of the traffic simulation, etc. can be used.

Based on the evaluation data and the signal parameters, the evaluation unit 300 calculates the traffic condition using the evaluation data as an input and the signal parameters, and calculates an index for evaluating the calculated traffic condition.

Specifically, first evaluation unit 300, a parameter space _{i t + 1} and the signal parameters _{s t + 1} from the parameter selection unit 144, acquires the evaluation data from the evaluation data storage unit 200.

Next, the evaluation unit 300 performs a traffic simulation for calculating the traffic condition using the evaluation data acquired from the evaluation data storage unit 200 and the signal parameter _{st + 1} acquired from the parameter selection unit 144, and sets an index. calculated, the calculated index, and outputs as an indication _{y t + 1} for the parameter space _{i t + 1,} and outputs as an indication _{l t + 1} for the signal parameters _{s t + 1.}

As an example of the index, there is a traffic jam loss time per vehicle. Congestion loss time is the total time that is delayed from a certain standard time due to congestion.

The evaluation section 300 sends a combination of the index _{y t + 1} parameter space _{i t + 1} and in the parameter space _{i t + 1} in the parameter space memory 110, the combined signal parameters _{s t + 1} and the index _{l t + 1} in the parameter storage unit 130.

Then, the evaluation unit 300 determines whether or not the number of times t of the current traffic simulation is performed exceeds the maximum number of times (for example, 1000 times) for repeating the predetermined traffic simulation. If t exceeds the maximum number of times, the output unit 400 is instructed to output the optimum signal parameter. On the other hand, if it does not exceed, the optimization unit 100 is instructed to perform the process again.

Instead of determining the maximum number of repetitions, it may be configured to repeat until the index converges.

The output unit 400 outputs the optimum signal parameters to the signal control control device 50.

Specifically, first, the output unit 400 acquires the signal parameters and indexes stored in the parameter storage unit 130 that have been subjected to the traffic simulation so far.

Then, the output unit 400 outputs the signal parameter s that minimizes the index l to the control device 50.

The control device 50 according to the present embodiment is composed of a computer including a CPU, a RAM, and a ROM storing a program for controlling a plurality of traffic signals, and is functionally configured as shown below. ing.

As shown in FIG. 1, the control device 50 according to the embodiment of the present invention includes an input unit 500 and a control unit 510.

The input unit 500 acquires the signal parameter s from the output unit 400.

Further, the input unit 500 receives an input of a traffic condition in an area including a plurality of signal devices.

The control unit 510 acquires the signal parameters s and the traffic condition from the input unit 500.

Then, the control unit 510 controls a plurality of signal devices by inputting the traffic condition and using the signal parameters s.

Specifically, the control unit 510 issues commands to each of the plurality of signal devices, such as switching, maintaining, and blinking the signal light color, based on the signal parameter s.

<< Experimental Results >>
Next, the experimental results of the optimization device according to the present invention are shown. In this experiment, nine traffic signals were controlled, and the parameter spaces i were set to nine types (series 1 to series 9). Moreover, the maximum number of times was set to 1000 times. In addition, the congestion loss time was adopted as an index.

FIG. 4 is a graph showing the relationship between the number of searches (that is, the number of traffic simulations) in the experimental results and the congestion loss time for each parameter space i. As shown in FIG. 4, immediately after the start, each parameter space i is randomly selected, and the congestion loss time for the signal parameter in the parameter space in each parameter space i is calculated.

However, a parameter space with a small number of searches and a large congestion loss time (for example, series 7) is learned as a parameter space in which a good index cannot be obtained, and the search is terminated or the number of searches is reduced.

On the other hand, by learning that the parameter space (series 3) having the shortest congestion loss time is optimal, it can be seen that the parameter space is optimized with a small number of searches.

Then, the parameters in the parameter space can be optimized with a small number of searches.

<Operation of the optimizing device according to the embodiment of the present invention>
FIG. 5 is a flowchart showing an optimization processing routine according to the embodiment of the present invention.

When the evaluation data is input to the evaluation unit 300, the optimization device 10 executes the optimization processing routine shown in FIG.

First, in step S100, the evaluation unit 300 acquires evaluation data from the evaluation data storage unit 200.

Next, in step S110, t = 1 is initialized.

In step S120, the parameter space model learning unit 122 receives the feature vector of the parameter space i from the parameter space storage unit 110.

Set X, vector

Set Z,

Set Y, parameter space i and its feature vector

Get the correspondence table with.

In step S130, the parameter space model learning unit 122 learns the first model based on X, Z, and Y.

In step S140, the parameter space selection unit 124 selects the parameter space it _{+ 1 for} which the traffic simulation is to be performed next, based on the first model received from the parameter space model learning unit 122.

In step S150, the parameter model learning unit 142, the set S of _{s t} from the parameter storage unit 130, a set of _{l t} L, the parameter space i, the spatial signal parameters which can be taken at the time of the parameter space i

Get the correspondence table of.

In step S160, the parameter model learning unit 142 learns the second model from S and L.

In step S170, the parameter selection unit 144 sets the signal parameter s _{t + 1 for} which the traffic simulation is performed next based on the second model received from the parameter model learning unit 142 and the parameter space it _{+ 1} for which the traffic simulation is performed next. select.

In step S180, the evaluation unit 300 performs a traffic simulation for calculating the traffic condition using the evaluation data acquired from the evaluation data storage unit 200 and the signal parameter _{st + 1} acquired from the parameter selection unit 144, and performs an index. was calculated, the calculated index, and outputs as an indication _{y t + 1} for the parameter space _{i t + 1,} and outputs as an indication _{l t + 1} for the signal parameters _{s t + 1.}

In step S190, the evaluation unit 300 determines whether or not the number of times t of the current traffic simulation is performed exceeds the maximum number of times the predetermined traffic simulation is repeated.

When t does not exceed the maximum number of times (NO in step S190), t = t + 1 is set in step S200, and the processes of steps S120 to S180 are repeated.

On the other hand, when t exceeds the maximum number of times (YES in step S190), in step S210, the output unit 400 outputs the signal parameter s at which the index l is the minimum to the control device 50, and performs the optimization processing routine. finish.

As described above, according to the optimization device according to the present embodiment, the first index, which is an index for the parameter space, is predicted and predicted based on the model for predicting the index for the parameter space. Select the parameter space to be evaluated next based on the value calculated based on the index of, and for each of the parameters included in the selected parameter space that have not been evaluated in the past, the index for the parameter. Predict the second index, which is an index for the parameter, based on the model for predicting, select the parameter to be evaluated next based on the value calculated based on the predicted second index, and use it for evaluation. By repeating the calculation of the index for evaluating the output based on the data and the selected parameters, the parameters can be optimized with a small number of evaluations.

Further, according to the traffic signal control system according to the present embodiment, the first index, which is an index for the parameter space, is predicted based on the model for predicting the index for the parameter space, and the predicted first index is used. Based on the value calculated based on, the parameter space to be evaluated next is selected, and among the signal parameters included in the selected parameter space, the index for the parameter is set for each of the signal parameters that have not been evaluated in the past. The second index, which is an index for the signal parameter, is predicted based on the model for prediction, and the signal parameter to be evaluated next is selected and evaluated based on the value calculated based on the predicted second index. By repeating the calculation of the traffic condition based on the data and the selected signal parameter, the signal parameter can be optimized with a small number of traffic simulations.

The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

In the present embodiment, the case where the traffic simulation is selected as the evaluation and the signal parameter is selected as the parameter has been described, but the present invention is not limited to this.

For example, as another embodiment, it can be applied to guide the crowd using a guide. As the correspondence in this example, for example, the evaluation can be performed by performing a human flow simulation, the parameter space can be the location of the guide, and the parameter can be the guidance method.

In addition, as another embodiment, it can be applied to the optimization of hyperparameters of machine learning. As the correspondence in this example, for example, the evaluation can be the training of the machine learning model, the parameter space can be the pipeline of machine learning, and the parameters can be hyperparameters.

Further, although described as an embodiment in which the program is pre-installed in the specification of the present application, it is also possible to provide the program by storing it in a computer-readable recording medium.

10 Optimization device 50 Control device 100 Optimization unit 110 Parameter space storage unit 120 Parameter space optimization unit 122 Parameter space model learning unit 124 Parameter space selection unit 130 Parameter storage unit 140 Parameter optimization unit 142 Parameter model learning unit 144 Parameter selection Unit 200 Evaluation data storage unit 300 Evaluation unit 400 Output unit 500 Input unit 510 Control unit

Claims (8)

It is an optimization device that optimizes the parameters used when calculating with evaluation data as input.
An evaluation unit that calculates an index for evaluating the result of the calculation based on the evaluation data and the parameters.
A parameter space that divides the space that the parameter can take, an optimization unit that optimizes the parameter, and
An output unit that outputs optimized parameters obtained by repeating the processing by the evaluation unit and the processing by the optimization unit.
Including
The optimization unit
A first model, which is a model for predicting an index for the parameter space, is learned based on the parameter and the index, and for each of the plurality of parameter spaces, the parameter space is based on the first model. A parameter space optimization unit that predicts a first index, which is an index for, and selects a parameter space to be evaluated next by the evaluation unit based on a value calculated based on the predicted first index.
A second model, which is a model for predicting an index for the parameter, is learned based on the parameter and the index, and among the parameters included in the parameter space selected by the parameter space optimization unit, the past For each of the parameters that have not been evaluated, a second index, which is an index for the parameter, is predicted based on the second model, and based on a value calculated based on the predicted second index. The parameter optimization unit, which selects the parameter to be evaluated next by the evaluation unit,
Optimizer including. The parameter space optimization unit uses the first model to predict the evaluation of the parameter space for each of the plurality of parameter spaces, and the first acquisition using the prediction of the evaluation of the parameter space as a variable. The function is calculated, and the parameter space in which the value of the first acquisition function is maximized is selected as the parameter space to be evaluated next by the evaluation unit.
The parameter optimization unit
Using the second model, among the parameters included in the parameter space selected by the parameter space optimization unit, each of the parameters that have not been evaluated in the past is next evaluated by the evaluation unit. The evaluation for the parameter space is predicted, the second acquisition function with the prediction of the evaluation for the parameter that has not been evaluated in the past as a variable is calculated, and the parameter having the maximum value of the second acquisition function is described. The optimization device according to claim 1, which is selected by the evaluation unit as a parameter to be evaluated next. The optimization device according to claim 1 or 2, wherein the first model and the second model are stochastic models using a Gaussian process. The parameter space optimization unit uses the parameter space to which the parameter used for evaluation by the evaluation unit belongs, the number of times of evaluation using the parameter belonging to the parameter space, and the index obtained by the evaluation unit. After learning the first model,
Any of claims 1 to 3, wherein the parameter optimization unit learns the second model using the parameters used when the evaluation unit performs the evaluation and the index obtained by the evaluation unit. The optimization device according to item 1. A traffic signal control system including a control device that controls a plurality of traffic signals and an optimization device that optimizes signal parameters used by the control device by inputting evaluation data necessary for calculating traffic conditions. ,
The control device
An input section that accepts status input and
A control unit that controls the plurality of traffic signals by using the traffic conditions as an input and the signal parameters obtained by the optimization device.
Including
The optimization device
Based on the evaluation data and the signal parameter, the evaluation unit calculates the traffic condition using the signal parameter with the evaluation data as an input, and calculates an index for evaluating the calculated traffic condition. ,
A parameter space that divides the space that the signal parameter can take, an optimization unit that optimizes the signal parameter, and
An output unit that outputs optimized parameters obtained by repeating the processing by the evaluation unit and the processing by the optimization unit.
Including
The optimization unit
A first model, which is a model for predicting an index for the parameter space, is learned based on the signal parameter and the index, and for each of the plurality of the parameter spaces, the parameter is based on the first model. A parameter space optimization unit that predicts the first index, which is an index for space, and selects the parameter space to be evaluated next by the evaluation unit based on the value calculated based on the predicted first index. ,
A second model, which is a model for predicting an index for the signal parameter, is learned based on the signal parameter and the index, and the signal parameter included in the parameter space selected by the parameter space optimization unit. Of these, for each of the signal parameters that have not been evaluated in the past, a second index, which is an index for the signal parameter, is predicted based on the second model, and calculation is performed based on the predicted second index. A parameter optimization unit that selects the signal parameter to be evaluated next by the evaluation unit based on the value to be evaluated.
Traffic signal control system including. It is an optimization method that optimizes the parameters when the evaluation data is input and output using the parameters.
A step in which the evaluation unit calculates an index for evaluating the output based on the evaluation data and the parameters.
The optimizing unit divides the space that the parameter can take, the parameter space, and the step of optimizing the parameter.
A step in which the output unit outputs an optimized parameter obtained by repeating the process by the evaluation unit and the process by the optimization unit.
Including
The steps that the optimization unit optimizes are
The parameter space optimization unit learns a first model that is a model for predicting an index for the parameter space based on the parameter and the index, and for each of the plurality of the parameter spaces, the first model is described. The first index, which is an index for the parameter space, is predicted based on the model, and the parameter space to be evaluated next by the evaluation unit is selected based on the value calculated based on the predicted first index. Steps and
The parameter optimization unit learns a second model, which is a model for predicting an index for the parameter based on the parameter and the index, and is included in the parameter space selected by the parameter space optimization unit. Of the parameters, for each of the parameters that have not been evaluated in the past, a second index that is an index for the parameter is predicted based on the second model, and based on the predicted second index. Based on the calculated value, the step of selecting the parameter to be evaluated next by the evaluation unit, and
Optimization methods including. A first index that is an index for the parameter space based on a first model for predicting a first index that is an index for the parameter space for each of a plurality of parameter spaces that divide the space that the parameter can take. And a parameter space optimization unit that selects a parameter space based on the predicted first index.
For each of the parameters included in the parameter space selected by the parameter space optimization unit, the index for the parameter is based on the second model for predicting the second index which is the index for the parameter. A parameter optimization unit that predicts two indexes and selects parameters based on the second index.
An output unit that outputs the parameters selected by the parameter optimization unit, and
Parameter search device including. A program for causing a computer to function as each part of the optimization device according to any one of claims 1 to 4.