Experimental planning device, experimental planning method, and experimental planning system

Abstract

To attain decentralization of a plurality of candidate points.SOLUTION: There is provided an experimental design device which, on the basis of a cost output as a result of inputting an input design parameter concerning an experimental device to an inference model which, when the input design parameter is input thereto, evaluates the experimental device, modifies the input design parameter by performing error back-propagation to the input design parameter, and, on the basis of the modified input design parameter, determines a design parameter candidate. The experimental design device: accepts a first input design parameter; calculates a first distance cost concerning a distance between the first input design parameter and the design parameter candidate; adds the first distance cost and a first cost output from the inference model as a result of inputting the first input design parameter to the inference model; on the basis of a result from the addition, performs the error back-propagation on the first input design parameter, thereby modifying the first input design parameter; and, on the basis of the modified first input design parameter, determines a first design parameter candidate.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to an experimental planning device, an experimental planning method, and an experimental planning system for planning an experiment using an experimental device. [Background technology]

[0002] Inference models using neural networks are being actively developed, and it is believed that by using these inference models to carry out experimental planning, it will be possible to design complex products at low cost. In this experimental planning application, the candidate experimental points must satisfy the following requirements (1) to (3).

[0003] (1) The candidate points to be output should not be just one solution to be finally adopted, but multiple candidate solutions to be experimented with should be output. (2) There must be significant experimental differences between the candidate points. If we conduct experiments under almost the same conditions or at points that give the same results, we will not be able to obtain meaningful data. (3) The properties of the candidate points can be intuitively controlled.

[0004] Furthermore, there are the following Patent Documents 1 and 2 as technologies using neural networks. Patent Document 1 discloses a neural network having an information element between one or more nodes of the neural network. The information element has a first input section having one or more first input nodes, a first output section having one or more first output nodes, and a plurality of first hidden nodes provided between the first input section and the first output section, with weighting coefficients set for connections on the input side and output side, the first input data received by the first input section and the first output data output by the first output section in response to the first input data match, and the weighting coefficient includes a value based on transmission information unrelated to learning of the neural network.

[0005] The following Patent Document 2 discloses a distributed deep learning system that performs distributed deep learning at high speed. In this distributed deep learning system, each learning node calculates the gradient of a loss function from the output result of inputting learning data to a neural network to be learned, packets the values ​​of multiple gradient components, and transmits them to a computing interconnect device. The computing interconnect device acquires the values ​​of multiple gradient components stored in the packets transmitted from each learning node, performs a calculation process in parallel for each of the multiple gradient component values ​​using the values ​​of the gradient components for the same configuration parameter of the neural network as input, and packets the calculation results and transmits them to each learning node. Each learning node updates the configuration parameter of the neural network based on the values ​​stored in the packets transmitted from the computing interconnect device. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2020-052813 A [Patent Document 2] JP 2019-168895 A Summary of the Invention [Problem to be solved by the invention]

[0007] Gradient descent adjusts the input to achieve a desired output by partial differentiation of the output of the inference model. However, in gradient descent, the output solutions are concentrated in the vicinity of a local optimum solution, and only a small number of candidate solutions with small differences and little experimental meaning are obtained. Therefore, gradient descent does not satisfy the above (1) and (2).

[0008] In addition, Bayesian optimization is a method to obtain multiple candidate solutions that are experimentally meaningful by increasing the cost using a kernel function so that it is easier to select points far from each candidate solution output. However, in the case of Bayesian optimization, it is difficult to control the optimization of the neural network inference model.

[0009] An object of the present invention is to disperse a plurality of candidate points. [Means for solving the problem]

[0010] An experimental planning apparatus according to one aspect of the invention disclosed in the present application comprises a processor for executing a program and a storage device for storing the program, and is capable of communicating with an experimental apparatus, wherein the processor performs an initial correction process for inputting initial design parameters in a design parameter space relating to the experimental apparatus as input parameters into an inference model for evaluating the experimental apparatus, calculating an initial evaluation cost as the evaluation result, and repeatedly correcting the initial design parameters by gradient descent based on the initial evaluation cost until a first condition relating to the number of corrections is satisfied; an initial candidate point addition process for adding the corrected initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether the number of candidate points in the candidate point output list has exceeded a predetermined number; and, if the number of candidate points in the candidate point output list does not exceed the predetermined number, setting the candidate point last added to the candidate point output list as a starting point, and calculating the initial design parameter space obtained by gradient descent from the starting point. a distance cost node addition process for adding a distance cost node that calculates a distance cost at a new position on the interval; a correction process for inputting a design parameter that is the candidate point last added to the candidate point output list as the input parameter into the inference model to calculate an evaluation cost, which is an evaluation result, and calculating a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes added so far, and repeatedly correcting the design parameter by gradient descent of the sum of the evaluation cost and each distance cost until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold is satisfied; a candidate point addition process for adding the corrected design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; and an output process for outputting the candidate point to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number. Effect of the Invention

[0011] According to the representative embodiment of the present invention, it is possible to distribute a plurality of candidate points. Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is an explanatory diagram illustrating an example of the system configuration of an experiment planning system. [Diagram 2] FIG. 2 is a block diagram showing an example of the hardware configuration of a computer. [Diagram 3] FIG. 3 is a block diagram of an example of a functional graph structure of the experimental planning apparatus according to the first embodiment. [Figure 4] FIG. 4 is a flowchart of an example of a procedure for generating candidate points by gradient descent according to the first embodiment. [Diagram 5] FIG. 5 is a diagram illustrating an example of generating candidate points by gradient descent according to the first embodiment. [Figure 6] FIG. 6 is a graph showing the progress of gradient descent according to the first embodiment. [Figure 7] FIG. 7 is an explanatory diagram showing an example of modification of design parameters. [Figure 8] FIG. 8 is a diagram of a specific example of the structure of the distance cost node according to the first embodiment. [Figure 9] FIG. 9 is an explanatory diagram showing an example of adding a distance cost node. [Figure 10] FIG. 10 is a diagram of an example of adding a distance cost node according to the second embodiment. [Figure 11]FIG. 11 is a diagram illustrating an example of generating candidate points by gradient descent according to the third embodiment. [Figure 12] FIG. 12 is an explanatory diagram of a first visualization example of a gradient using distance cost nodes according to the third embodiment. [Figure 13] FIG. 13 is an explanatory diagram of a second visualization example of a gradient using distance cost nodes according to the third embodiment. [Figure 14] FIG. 14 is a diagram illustrating an example of a distance cost node according to the third embodiment. [Figure 15] FIG. 15 is a diagram illustrating an example of a grid search according to the fourth embodiment. [Figure 16] FIG. 16 is a flowchart of an example of a procedure for generating candidate points by gradient descent according to the fourth embodiment. [Figure 17] FIG. 17 is a diagram illustrating an example of a distance cost node according to the fifth embodiment. [Figure 18] FIG. 18 is an explanatory diagram of a visualization example of a gradient using a distance cost node according to the fifth embodiment. [Figure 19] FIG. 19 is a diagram of an example of a graph structure according to the fifth embodiment. [Figure 20] FIG. 20 is an explanatory diagram illustrating an example of a cost adjustment layer. [Figure 21] FIG. 21 is an explanatory diagram showing an example of a graph showing a cost input / output relationship. [Figure 22] FIG. 22 is an explanatory diagram showing another example of the cost adjustment layer. [Figure 23] FIG. 23 is an explanatory diagram showing another example of a graph showing a cost input / output relationship. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] <Experimental Planning System> 1 is an explanatory diagram showing an example of the system configuration of an experimental planning system. The experimental planning system 100 includes an experimental planning device 101, a learning device 102, and an experimental device 103. The experimental planning device 101, the learning device 102, and the experimental device 103 can communicate with each other via a network 104 such as the Internet, a LAN (Local Area Network), or a WAN (Wide Area Network).

[0014] The experimental planning system 100 uses an inference model created by a learning device 102 to sample candidate points within a range in which design parameters can be set by an experimental planning device 101 and that are spaced sufficiently apart from each other, and provides the experimental device with, for example, experimental conditions that suppress manufacturing variations in the experimental device 103 as optimal design parameters.

[0015] The learning device creates an inference model using the design parameters and objective variables of the experimental device as a learning data set. The experimental device is, for example, a semiconductor manufacturing device. The experimental device performs an experiment with new design parameters obtained from the experimental planning device and obtains experimental results. The experimental device then outputs the new design parameters and the experimental results to the learning device. The learning device adds the new design parameters and the experimental results to the learning data set and re-learns the inference model.

[0016] <Hardware configuration of computer (experimental planning device and learning device)> FIG. 2 is a block diagram showing an example of a hardware configuration of a computer. The computer 200 includes a processor 201, a storage device 202, an input device 203, an output device 204, and a communication interface (communication IF) 205. The processor 201, the storage device 202, the input device 203, the output device 204, and the communication IF 205 are connected by a bus 206. The processor 201 controls the computer 200. The storage device 202 serves as a working area for the processor 201. The storage device 202 is a non-transient or temporary recording medium that stores various programs and data. Examples of the storage device 202 include a ROM (Read Only Memory), a RAM (Random Access Memory), a HDD (Hard Disk Drive), and a flash memory. The input device 203 inputs data. Examples of the input device 203 include a keyboard, a mouse, a touch panel, a numeric keypad, a scanner, and a microphone. The output device 204 outputs data. The output device 204 may be, for example, a display, a printer, or a speaker. The communication IF 205 is connected to the network 104 and transmits and receives data.

[0017] Hereinafter, examples of experimental planning using the above-mentioned experimental planning system will be described for each embodiment. EXAMPLES

[0018] Example 1 shows a candidate point generation example 1 by gradient descent. In the experimental planning device 101, each node constituting a graph structure described later is specifically realized by, for example, making the processor 201 execute a program stored in the storage device 202 shown in FIG.

[0019] 3 is a block diagram showing an example of a functional graph structure of the experimental planning device 101 according to the first embodiment. The input node 300 accepts an input of a design parameter x[i]. The design parameter x[i] is a set of design parameters xi1, xi2, ..., xij, ..., xin (n is an integer), and is an input vector to an inference model 301 and a distance cost node group 302. i is an iteration indicating the number of a candidate point Pi to be searched for by the experimental planning device 101. The initial value is i=0.

[0020] The design parameter x[i] is the initial input vector to the experimental setup 103 or a modified input vector by gradient descent of the design parameter x[i-1]. j is a number that identifies the elements xi1, xi2, ..., xij, ..., xin in the design parameter x[i] (sometimes referred to as the axis j).

[0021] The inference model 301 is a neural network that infers an output related to the experimental device 103, and may be regression or discrimination, and outputs an evaluation value that evaluates the inference result as a cost C1i. The cost C1i is, for example, an index value (the higher the value, the worse the performance; for example, failure rate or price) that indicates the poorness of performance when the design parameter x[i] is applied to the experimental device 103. The experimental planning device 101 will modify the design parameter x[i] so as to improve the performance.

[0022] Also, the experimental planning device 101 may have an untrained neural network, obtain learning parameters (weights, biases, hyperparameters) constituting the inference model 301 from the learning device 102, and set them in the untrained neural network. Note that although there is one inference model 301 in Fig. 3, there may be multiple inference models 301. This makes it possible to generate design parameters x[i] that can be commonly applied to multiple inference models 301.

[0023] The distance cost node group 302 is a set of zero or more distance cost nodes 302-1, 302-2, ..., 302-i. The distance cost node 302-i calculates a distance cost C2i for each axis j at a new position in the design parameter space that is separated by a separation distance thi from the design parameter x[i]. By separating by the separation distance thi, the output solution is prevented from concentrating in the vicinity of the local optimum solution. The distance cost C2i is a set of distance costs C2ij calculated for the design parameter xij.

[0024] The cost addition node 303 adds up the cost C1i from the inference model 301 and the distance cost C21 to C2i from the distance cost node 302-1 to the distance cost node 302-i when the design parameter x[i] is input to the inference model 301 and the distance cost node 302-1 to the distance cost node 302-i. The cost addition node 303 outputs the sum, cost C3i (=C1i+C21+...+C2i), to the loss function output node 304.

[0025] The loss function output node 304 inputs the cost C3i, differentiates the cost C3i, and performs gradient descent to repeat correction by error backpropagation to the design parameter x[i]. The differential value of the cost C3i is set as ΔC3i. Specifically, for example, the loss function output node 304 repeats gradient descent using the cost C3i obtained from the corrected design parameter x[i] and correction of the corrected design parameter x[i] until the differential value ΔC3i becomes equal to or less than a predetermined threshold value (a termination criterion value in FIG. 6 described later).

[0026] Fig. 4 is a flowchart showing an example of a candidate point generation process by gradient descent according to the first embodiment. Fig. 5 is an explanatory diagram showing an example of candidate point generation by gradient descent according to the first embodiment. Fig. 6 is a graph showing the progress of gradient descent according to the first embodiment. In Fig. 5, for the sake of simplicity, the design parameter x[i] is set to two (n=2) design parameters xi1 and xi2.

[0027] The experimental planning device 101 executes input initialization (step S401). Specifically, for example, as shown in FIG. 5A, the design parameter x[i] that is input-initialized in iteration i is expressed as xis. In FIG. 5A, since i=0, the experimental planning device 101 sets the design parameters xis=x0s={x01, x02}. At i=0, no distance cost node 302-i is connected.

[0028] Next, the experimental planning device 101 judges whether the candidate point output condition is satisfied (step S402). The candidate point output condition is that ΔC3i becomes equal to or smaller than the termination judgment reference value of the gradient descent shown in Fig. 6 by the gradient descent by differentiation of the cost C3i, or the number of gradient descents in the iteration i reaches a predetermined number (75 times in the example of Fig. 6). The design parameter x[i] when the candidate point output condition is satisfied by the gradient descent by differentiation of the cost C3i is denoted as the design parameter xie.

[0029] If ΔC3i becomes less than or equal to the gradient descent termination criterion value (step S402: Yes), it means that cost minimization through gradient descent has been achieved, and the experimental planning device 101 determines the position of the design parameter xie obtained by that modification as the candidate point Pi and adds it to the candidate point output list (step S404).

[0030] Even if ΔC3i is not equal to or less than the gradient descent termination criterion value, if the number of executions of step S403 reaches or exceeds a predetermined number, the number of executions of step S403 is initialized, and the experimental planning device 101 determines the position of the design parameter x[i] obtained by the correction at that time as the candidate point Pi (step S404). If the candidate point output condition is not satisfied (step S402: No), the experimental planning device 101 repeatedly executes step S403.

[0031] If step S402: No, the experimental planning device 101 executes cost minimization by gradient descent based on the cost calculation in the inference model 301 and the distance cost nodes 302-1 to 302-i and the addition by the cost addition node 303 (step S403). Specifically, for example, the experimental planning device 101 inputs the design parameter xis to the inference model 301 and the distance cost nodes 302-1 to 302-i, and obtains the cost C1i as an output from the inference model 301, and obtains the distance costs C21 to C2i from the distance cost node 302-1 to the distance cost node 302-i. Then, the experimental planning device 101 adds the cost C1i and the distance costs C21 to C2i by the cost addition node 303, and outputs the cost C3i as the addition result to the loss function output node 304.

[0032] When i=0, no distance cost node 302-i is connected, so the experimental planning device 101 inputs the design parameter x0s to the inference model 301, obtains the output cost C10 from the inference model 301 (step S403-0), and outputs the sum, i.e., the cost C30 (=C10), to the loss function output node 304.

[0033] Then, when the cost C3i is input from the cost addition node 303, the loss function output node 304 executes cost minimization by gradient descent. Specifically, for example, the loss function output node 304 differentiates the cost C3i to calculate ΔC3i as a loss, and corrects the design parameter x[i].

[0034] Then, the experimental planning device 101 judges whether ΔC3i from the loss function output node 304 satisfies the candidate point output condition (step S402). If ΔC3i satisfies the candidate point output condition (step S402: Yes), the experimental planning device 101 adds the design parameter xie at that time to the candidate point output list as the candidate point Pi (step S404). When i=0, as shown in FIG. 5A, the experimental planning device 101 adds the design parameter x0e at that time to the candidate point output list as the candidate point P0 (step S404-0).

[0035] Then, the experimental planning device 101 increments the iteration i (step S405) and determines whether the number of candidate points registered in the candidate point output list has exceeded a predetermined number N (N is an integer greater than or equal to 1), i.e., whether i has exceeded N (step S406).

[0036] If the number of candidate points does not exceed the predetermined number N (step S406: No), the experimental planning device 101 sets the candidate point P(i-1) newly added in step S404 as the design parameter xis, adds a distance cost node 302-i for the design parameter xis to the distance cost node group 302, generates a ReLU field Ai (step S407), and proceeds to step S402.

[0037] 5B, the experimental planning device 101 sets the candidate point P0 in (A) as the design parameter x1s and generates the ReLU field A1 (step S407-1). Then, the process proceeds to step S402.

[0038] Returning to step S402, the experimental planning device 101 executes a determination as to whether the design parameter x1s satisfies the candidate point output condition (step S402), and if it does not satisfy the condition (step S402: No), executes cost minimization by gradient descent (step S403-1) as shown in Fig. 5B and Fig. 6. At i=1, the experimental planning device 101 inputs the design parameter x1s to the inference model 301 and the distance cost node 302-1, and obtains the cost C11 and the distance cost C21, which are output, from each of the inference model 301 and the distance cost node 302-1.

[0039] At i=1, the experimental planning device 101 uses the cost addition node 303 to output the cost C31 (=C11+C21), which is the addition result, to the loss function output node 304. Then, when the loss function output node 304 receives the cost C31 from the cost addition node 303, it executes cost minimization by gradient descent. Specifically, for example, the loss function output node 304 differentiates the cost C31 to calculate ΔC31 as a loss function, and modifies the design parameter x[1].

[0040] Then, the experimental planning device 101 judges whether ΔC3i from the loss function output node 304 satisfies the candidate point output condition (step S402). For i=1, ΔC31 becomes equal to or smaller than the gradient descent termination criterion value before the number of gradient descents reaches a predetermined number, as shown in Fig. 6 (step S402: Yes). Therefore, the experimental planning device 101 adds the design parameter x1e at that time as candidate point P1 to the candidate point output list, as shown in Fig. 5(B) (step S404-1).

[0041] Then, the experimental planning device 101 increments the iteration i (step S405), and determines whether or not the number of candidate points registered in the candidate point output list exceeds a predetermined number N, that is, whether or not i exceeds N (step S406).

[0042] If the number of candidate points does not exceed the predetermined number N (step S406: No), the experimental planning device 101 sets the candidate point P1 newly added in step S404 as the design parameter x2s, adds the distance cost node 302-2 for the design parameter x2s to the distance cost node group 302, generates a ReLU field A2 (step S407-2), and proceeds to step S402.

[0043] Returning to step S402, the experimental planning device 101 executes a judgment as to whether the design parameter x2s satisfies the candidate point output condition (step S402), and if it does not satisfy it (step S402: No), executes cost minimization by gradient descent (step S403-2) as shown in Fig. 5C and Fig. 6. For i=2, the design parameter x2s is input to the inference model 301, the distance cost node 302-1, and the distance cost node 302-2, and the cost C12 and the distance costs C21 and C22, which are output, are obtained from each of the inference model 301, the distance cost node 302-1, and the distance cost node 302-2.

[0044] Then, in the experimental planning device 101, since the distance cost nodes 302-1 and 302-2 are connected at i=2, the cost addition node 303 outputs the cost C32 (C12+C21+C22) which is the addition result to the loss function output node 304. Then, when the cost C31 from the cost addition node 303 is input, the loss function output node 304 executes cost minimization by gradient descent. Specifically, for example, the loss function output node 304 differentiates the cost C31 to calculate ΔC32 as a loss function, and corrects the design parameter x[2].

[0045] 7 is an explanatory diagram showing an example of correction of the design parameter x[2]. The correction amount of the design parameter x[2] is determined so as to avoid the ReLU fields A1 and A2 by combining the correction power due to the gradient of the inference model 301, the correction power from the candidate point P0, and the correction power from the candidate point P1.

[0046] Returning to Fig. 4, the experimental planning device 101 judges whether or not ΔC3i satisfies the candidate point output condition (step S402). For i=2, as shown in Fig. 6, ΔC32 becomes equal to or smaller than the gradient descent termination criterion value before the gradient descent count reaches a predetermined count (step S402: Yes). Therefore, as shown in Fig. 5(C), the experimental planning device 101 adds the design parameter x2e at that time to the candidate point output list as candidate point P2 (step S404-2).

[0047] Then, the experimental planning device 101 increments the iteration i (step S405), and determines whether or not the number of candidate points registered in the candidate point output list exceeds a predetermined number N, that is, whether or not i exceeds N (step S406).

[0048] If the number of candidate points exceeds the predetermined number N (step S406: Yes), the experimental planning device 101 ends the candidate point generation process by gradient descent.

[0049] After that, the experimental planning device 101 transmits a candidate point output list including the candidate points P0 to P2 to the experimental device 103. The experimental device 103 performs an experiment using each of the candidate points P0 to P2 as a new design parameter x[i]. The experimental device 103 transmits the new design parameter x[i] and the experimental result to the learning device 102. The learning device 102 adds the new design parameter x[i] and the experimental result to a learning dataset, and re-trains the inference model 301.

[0050] <Specific example of distance cost node structure> 8 is an explanatory diagram showing a specific example of the structure of the distance cost node 302-i according to the first embodiment. In the distance cost node 302-i, (xi1, xi2, ..., xin) is input as the design parameter x[i] to the input node 300. When there is no need to distinguish between xi1, xi2, ..., xin, they are simply written as design parameters xi.

[0051] A design parameter x[i] is a vector in a design parameter space 500 spanned by n axes. The input node 300 accepts the input of the design parameter x[i] and outputs it to an inference model 301 and a distance cost node 302-i. The inference model 301 outputs a cost C1i, which is an inference result, to a cost addition node 303.

[0052] The distance cost node 302-i has a distance node 810 and a cost node 820. The distance node 810 is a node that calculates the distance between the design parameter xis of the iteration i, which is the candidate point P(ik) (hereinafter, may be written as Pa) generated in one or more previous iterations (ik), and the design parameter x[i] after gradient descent started from the design parameter xis, and executes a first subtraction process 811 and an absolute value process 812. Here, k is an integer between 1 and i. The first subtraction process 811 is a process that subtracts the design parameter x[i] of the input node 300 from the candidate point Pa (a1, a2, ..., an) and outputs a first subtraction result 8110.

[0053] The absolute value process 812 is a process that converts the first subtraction result 8110 of the first subtraction process 811 into an absolute value vector 8120. That is, the distance node 810 obtains the absolute value vector 8120 of the difference between the design parameter x[i] and the candidate point Pa, that is, the distance between the design parameter x[i] and the candidate point Pa.

[0054] The cost node 820 executes a second subtraction process 821, a ReLU activation process 822, and a weighting process 823. The second subtraction process 821 is a process for taking the difference between the separation distance th and the element of each axis j of the absolute value vector 8120 of the absolute value process 812, and outputting a second subtraction result 8210. The separation distance th is a hyperparameter that can be arbitrarily set for each axis j. This allows a set of candidate points Pi separated by the separation distance th or more to be obtained. The output at this point is positive for an axis j with insufficient separation, and negative for an axis j with sufficient separation.

[0055] The ReLU activation process 822 inputs the second subtraction result 8210 from the second subtraction process 821 into a ReLU function to activate it. Specifically, for example, the ReLU activation process 822 generates an output with a gradient of 0 when a negative value (a value of axis j with sufficient separation) is input, and a gradient of 1 when a positive value (a value of axis j with insufficient separation) is input. In other words, the ReLU activation process 822 generates a gradient for axis j with insufficient separation distance th, and feeds it back to the input to move it away from the candidate point Pa.

[0056] The weighting process 823 is a process that weights each of the output values ​​from the ReLU activation process 822 by appropriate combination and outputs the weighted values ​​to the cost addition node 303. Specifically, for example, the weighting process 823 may extract the minimum output value from the set of output values ​​from the ReLU activation process 822, or may calculate the sum of the set of output values. The weighting process 823 can adjust the priority of separation by changing the weights used at the time of combination.

[0057] Here, if the output from the cost addition node 303 is F(x), the loss function output node 304 modifies the design parameter x[i] according to the following formulas (1) and (2). Here, η is a learning rate (0<η≦1), "←" indicates substitution, and L(F(x)) is the output of the loss function output node 304. A loss function is generally evaluated based on the error from a target value, but if each node is designed so that F(x)>0, the target value may always be 0. In other words, the experimental planning device 101 performs an operation to simply minimize the function F(x).

[0058] d{L(F(x)-0)} / dx=dL / dF×dF / dx...(1) x[i]←x[i]-η×dL / dF×dF / dx (2)

[0059] In the error backpropagation by the above formulas (1) and (2), the contribution of the distance cost node 303 to the input correction, that is, the learning rate η, can be changed for each axis j by changing W used in the weighting process 823 for each axis j as shown in the following formula (3). Note that sign(x) is a function that returns -1 when x is negative, 1 when x is positive, and 0 when x is 0. C1i is the output value of the inference model and is not affected by the weight W.

[0060] dF(x) / dxj=dC1i / dxj+dC2i / dxj =dC1i / dxj-sign(xj-aj)×Wj (However, thj<|xj-aj|)····(3)

[0061] Fig. 9 is an explanatory diagram showing an example of adding a distance cost node 302-i. Fig. 9 shows an example of adding a distance cost node 302-2 to which a candidate point P1 is set when a new candidate point P1 is output to the candidate point output list in step S404 shown in Fig. 4. By changing the graph structure in this way, the distance cost node 302-i is connected every time a candidate point P(i-1) is added.

[0062] When the distance cost node 302-1 is connected and the distance cost node 302-2 is not connected, that is, when the candidate point is only P0, the distance cost node 302-1 calculates the distance cost C21 by inputting the design parameter x[1] and the candidate point P0. When the distance cost node 302-2 is connected, the distance cost node 302-1 calculates the distance cost C21 by inputting the design parameter x[2] and the candidate point P0.

[0063] In this way, the design parameter x[i] input to the i-1th distance cost nodes 302-(i-1) changes every time a distance cost node 302-i is added. Therefore, the distance cost C2k calculated by the kth (k=1 to i)th distance cost node 302-k based on the input of the design parameter x[i] and the candidate point P(k-1) is represented as C2k(i, k-1).

[0064] In the above example, the distance cost C21 of the former is C21(1,0), and the distance cost C21 of the latter is C21(2,0). The distance cost C21(1,0) and the distance cost C21(2,0) are different distance costs.

[0065] In this way, according to the first embodiment, the degree of separation can be intuitively determined by the separation distance th, so that a discretized group of candidate points Pi that is easy to use experimentally can be obtained. Furthermore, the priority of the separation can be determined by the weight W in the weighting process 823. A minimum value or sum may simply be taken, in which case the number of hyperparameters is small and it is easy to use. Furthermore, since the calculation of the distance cost is only a simple addition, subtraction, and integration, the gradient disappearance does not occur in the distance cost node 302-i, and the convergence of the gradient descent is not adversely affected. Furthermore, if the separation is sufficient, the optimization of the inference model 301 is not affected. EXAMPLES

[0066] Next, a second embodiment will be described. In the second embodiment, differences from the first embodiment will be mainly described, and therefore the description of the common parts with the first embodiment will be omitted. In the first embodiment, when a candidate point P(i-1) is newly output, a distance cost node 302-i is newly added. In the second embodiment, the experimental planning device 101 connects a plurality of distance cost nodes 302-1, 302-2, ..., 302-i in advance in a graph structure. Each value xaj of each candidate point Pa of the distance cost node 302-i is set to a value inf that cannot be taken as the candidate point Pa. That is, inf is a value that causes the distance cost node 302-i to output 0 when it is input to the distance cost node 302-i.

[0067] Then, when a new candidate point Pa is output by step S404 shown in FIG. 4, the experimental planning device 101 selects an unselected distance cost node 302-i, and replaces the value inf for each axis j of the candidate point Pa of the selected distance cost node 302-i with the value aj of the newly output candidate point Pa.

[0068] Fig. 10 is an explanatory diagram showing an example of adding distance cost nodes 302-i according to the second embodiment. For the sake of simplicity, the total number of prepared distance cost nodes 302-i is set to 3 in Fig. 10, but it may be 2 or 4 or more. (A) shows the initial state of the graph structure, (B) shows the graph structure after the first candidate point P0 is obtained, and (C) shows the graph structure after the second candidate point P1 is obtained.

[0069] The value inf of the candidate point Pa set in the unselected distance cost node 302-i is a value that outputs a gradient of 0 when input to the ReLU function of the distance cost node 302-i (ReLU(th-|inf-x|)=0).

[0070] (A) In the initial state of the graph structure, the experimental planning device 101 repeats cost minimization by gradient descent, that is, calculation of ΔC30 and modification of the design parameter x[0], until the candidate point output condition is satisfied (step S402: No), as shown in Fig. 4 and Fig. 5(A). At this time, the inference model 301 outputs the cost C10 to the cost addition node 303, and the distance cost nodes 302-1 to 302-3 output 0 as the distance cost.

[0071] When the candidate point output condition is satisfied (step S402: Yes), the experimental planning device 101 adds the candidate point P0 to the candidate point output list (step S404-0) as shown in Fig. 4 and Fig. 5 (A). As a result, as shown in (B), the value inf input to the distance cost node 302-1 is replaced with the value x0j of the candidate point P0. Therefore, when the design parameter x[1] is input, the distance cost node 302-1 outputs the distance cost C21(1,0).

[0072] (B) In the graph structure after acquiring the first candidate point P0, the experimental planning device 101 repeats the cost minimization by gradient descent, that is, the calculation of ΔC31 and the modification of the design parameter x[1], until the candidate point output condition is satisfied (step S402: No), as shown in Figures 4 and 5(A). At this time, the inference model 301 outputs the cost C11 to the cost addition node 303, the distance cost node 302-1 outputs the distance cost C21(1,0), and the distance cost nodes 302-2 and 302-3 output 0 as the distance cost.

[0073] When the candidate point output condition is satisfied (step S402: Yes), the experimental planning device 101 adds the candidate point P1 to the candidate point output list (step S404-1) as shown in Fig. 4 and Fig. 5(A). As a result, as shown in (C), the value inf input to the distance cost node 302-2 is replaced with the value x1j of the candidate point P1. Therefore, when the design parameter x[2] is input, the distance cost node 302-2 outputs the distance cost C22(2,1).

[0074] (C) In the graph structure after acquiring the second candidate point P1, the experimental planning device 101 repeats the cost minimization by gradient descent, that is, the calculation of ΔC32 and the modification of the design parameter x[2], until the candidate point output condition is satisfied (step S402: No), as shown in Figures 4 and 5(A). At this time, the inference model 301 outputs the cost C12 to the cost addition node 303, the distance cost node 302-1 outputs the distance cost C21(2,1), the distance cost node 302-2 outputs the distance cost C22(2,1), and the distance cost node 302-3 outputs 0 as the distance cost.

[0075] If the candidate point output condition is satisfied (step S402: Yes), the experimental planning device 101 adds the candidate point P2 to the candidate point output list (step S404-2), as shown in FIG. 4 and FIG. 5(A).

[0076] In this way, by connecting multiple distance cost nodes 302-i to which the value inf of candidate point Pa is set between the input node 300 and the cost addition node 303 before generating candidate point Pa by gradient descent, it is possible to execute calculations at the distance cost nodes 302-1 to 302-i simply by rewriting the value inf of candidate point Pa each time a candidate point Pa is added. In this way, the graph structure is fixed from the start of execution, resulting in a static graph structure and improving throughput. EXAMPLES

[0077] Next, a description will be given of Example 3. In Example 3, differences from Examples 1 and 2 will be mainly described, and therefore descriptions of common parts with Examples 1 and 2 will be omitted.

[0078] Fig. 11 is an explanatory diagram showing an example of generating candidate points by gradient descent according to the third embodiment. Fig. 12 is an explanatory diagram showing a first visualization example of a gradient using a distance cost node according to the third embodiment. Fig. 13 is an explanatory diagram showing a second visualization example of a gradient using a distance cost node according to the third embodiment. Fig. 12 shows the ReLU field B1 in (B) of Fig. 11, and Fig. 13 shows the ReLU fields B1 and B2 in (B) of Fig. 11.

[0079] 5 is that the shape of the ReLU field is different. Since the calculation of the distance cost node 302-i is different from that of the first embodiment, the ReLU field Bi has a quadrangular pyramid shape (so-called pyramid shape) with the candidate point P(i-1) (design parameter xis) as the apex. That is, unlike the ReLU field Ai, in the ReLU field Bi, the range of the values ​​xi1 and xi2 of each axis j of the design parameter xi of the ReLU field Bi is limited.

[0080] 13, the overlapping region OLB of the ReLU fields B1 and B2 is a flat surface, and therefore the multi-peak problem is alleviated compared to the cross-shaped overlapping region OLA of the ReLU fields A1 and A2 in the embodiment 1. Next, a specific example of the distance cost node 302-i that generates the above-mentioned ReLU fields B1 and B2 will be described.

[0081] 14 is an explanatory diagram showing an example of the distance cost node 302-i according to the third embodiment. Here, the distance cost node 302-1 is taken as an example for explanation. Among the distance cost nodes according to the third embodiment, the distance node 810 has the same configuration as that of the first embodiment. The cost node 820 according to the third embodiment executes a 1 / th multiplication process 1421 and a minimum value selection process 1422 between a second subtraction process 821 and a ReLU activation process 822, and does not execute a weighting process 823.

[0082] The 1 / th multiplication process 1421 is a process that multiplies the second subtraction result 8210 from the second subtraction process 821 by the reciprocal vector (1 / th1, ..., 1 / thj, ..., 1 / thn) of the separation distance th, and outputs the multiplication result 14210. This limits the range of each axis j of the ReLU field Bi to 1 / thj.

[0083] The minimum value selection process 1422 is a process for selecting a minimum value from the multiplication result 14210 of the 1 / th multiplication process 1421. Therefore, in the ReLU activation process 822, the minimum value selected by the minimum value selection process 1422 is input to the ReLU function for activation. The design parameter x[i] receives a correction force only in the axis j direction of the minimum value selected by the minimum value selection process 1422 of the ReLU field Bi. In other words, as is clear from a comparison of Fig. 5 and Fig. 11, the area that becomes unselectable by the ReLU field is small. Therefore, the variety of solutions that can be selected as candidate points is excellent.

[0084] In this way, the design parameter x[i] is subjected to gradient descent of the axis j whose separation distance has been minimized in the minimum value selection process 1422, that is, the design parameter x[i] is separated from the candidate point Pa, which is the vertex of the ReLU field Bi, by the axis j that can be separated by the shortest distance. In this way, since the axis j of interest is always handled, it is intuitive and easy to manage.

[0085] In addition, in the overlapping region OLB of the ReLU fields B1 and B2, the gradients cancel each other out, making the field simpler, and the problem of complexity caused by multiple cost peaks due to the loss function L is mitigated. This is particularly effective in the case of grid search. EXAMPLES

[0086] Next, Example 4 will be described. In Example 4, differences from Examples 1 to 3 will be mainly described, and explanations of common parts with Examples 1 to 3 will be omitted. Example 4 is an example in which the experimental planning device 101 divides the design parameter space 500 into a grid and quantizes the design parameter x[i] into lattice points to generate candidate points.

[0087] Fig. 15 is an explanatory diagram showing an example of a grid search according to the embodiment 4. Fig. 16 is a flowchart showing an example of a candidate point generation process procedure by gradient descent according to the embodiment 4. The difference from Fig. 4 is that quantization (step S1600) is added between step S402: Yes and step S404.

[0088] In FIG. 15, the design parameter space 500 is divided into a grid. An area surrounded by vertical and horizontal grid lines glv and glh is called a grid. In FIG. 15, grids G1 to G4 are shown as an example. An intersection point between a vertical grid line glv and a horizontal grid line glh is called a lattice point. The interval between adjacent lattice points is d. Note that d may take a different value for each axis j. If the design parameter x[i] is not located on a lattice point, it is quantized to the nearest lattice point by quantization.

[0089] Fig. 15 shows the state of Fig. 11(B). Here, for the sake of simplicity, the bottom surface of the ReLU field B1 is a square, and the length of the side is twice the separation distance th (2 × th). Note that the separation distance thj is a different distance for each axis j, but here, for the sake of simplicity, th1 = th2 = th.

[0090] In Fig. 16, after step S402: Yes, the experimental planning device 101 quantizes the design parameter x[i] that satisfies the candidate point output condition (step S1600). Specifically, for example, in Fig. 15, the experimental planning device 101 moves the design parameter x[i] to the lattice point with the shortest distance (which may be Euclidean distance or Manhattan distance). The quantized design parameter x[i] is designated as Q(x[i]).

[0091] For example, if x[1]=(0.1,1.3) is the convergence value of the gradient descent that satisfies the candidate point output condition, Q(x[1])=(0,1) is added as a candidate point (step S404). Also, in FIG. 15, the ReLU field B1 is generated around the lattice point where Q(x[1]) is located (step S407).

[0092] When performing gradient descent in the direction of axis j=1, Q(x[i]) on a lattice point moves to the neighboring lattice point to the right due to gradient descent. In other words, if 2×th>d is satisfied, Q(x[i]) on a lattice point within ReLU field B1 moves to a lattice point outside ReLU field B1. Therefore, Q(x[i]) does not remain at the same lattice point due to gradient descent, but instead gradient descents to a different lattice point.

[0093] In addition, in the gradient descent of the design parameter x[2], the design parameter x[2] does not converge within the range of the ReLU field B1, so when the design parameter x[2] satisfies the candidate point output condition (step S402: Yes), the quantized design parameter Q(x[2]) automatically converges to a lattice point outside the range of the ReLU field B1 (step S1600). This makes it unnecessary to determine whether Q(x[i]) fall on the same lattice point.

[0094] In this manner, in the fourth embodiment, by quantizing the design parameter x[i] in the gridded design parameter space 500, it is possible to prevent the design parameter Q(x[i]) from concentrating on the same lattice point, and it is possible to efficiently discretize the design parameter Q(x[i]). EXAMPLES

[0095] Next, a fifth embodiment will be described. In the fifth embodiment, differences from the first to fourth embodiments will be mainly described, and therefore descriptions of common parts with the first to fourth embodiments will be omitted. The fifth embodiment is an example in which the Euclidean distance is used as the cost.

[0096] For example, if the design parameter xi1 is energy and the design parameter xi2 is position, the number of dimensions of the design parameter xi1 is different from the number of dimensions of the design parameter xi2. Therefore, it is qualitatively dangerous to take the sum or difference between the design parameters xi1 and xi2, and it is better to separate them by the pyramidal ReLU fields B1 and B2 as shown in the third embodiment.

[0097] On the other hand, when the dimensionality of the design parameters xi1 and xi2 is the same, for example, as in the case of energy, the Euclidean distance can be used. In this case, the experimental planning device 101 can search for more candidate points in the same spatial volume, even with the same separation distance th, compared to the pyramid type of the third embodiment.

[0098] FIG. 17 is an explanatory diagram illustrating an example of the distance cost node 302-i according to the fifth embodiment. FIG. 18 is an explanatory diagram illustrating an example of visualization of the gradient by the distance cost node 302-i according to the fifth embodiment. In FIG. 17, the distance node 810 executes a first subtraction process 811 and a multiplication process 1712. The multiplication process 1712 multiplies the design parameters x[i] and (1 / th0 2 ,1 / th1 2 ,…,1 / thn 2 ) and output the multiplication result 17120.

[0099] In addition, the cost node 820 executes an addition process 1721, a third subtraction process 1722, and a ReLU activation process 822. The addition process 1721 is a process of adding elements of each axis j of the multiplication result 17120 and outputting an addition result 17210. The third subtraction process is a process of subtracting the addition result 17210 from 1 and outputting a third subtraction result 17220.

[0100] The ReLU activation process 822 inputs the third subtraction result 17220 into the ReLU function for activation, and outputs the distance cost C2i. Specifically, for example, if the third subtraction result 17220 is 0 or less, the ReLU activation process 822 outputs a gradient of 0. On the other hand, if the third subtraction result 17220 is greater than 0, the ReLU activation process 822 outputs a gradient of 1. For example, in the state of FIG. 11B, if the point at the destination of the gradient descent is within the ReLU field B1 having a semi-ellipsoid shape as shown in FIG. 18, a gradient of 1 is output.

[0101] According to the fifth embodiment, by using the Euclidean distance as the cost, gradient vectors are generated using all axes j so that the distance can be minimized. Therefore, it is useful when the number of dimensions of the design parameters x[i] of each axis j is the same. In addition, since the ReLU field B1 shown in FIG. 17 has a semi-ellipsoid shape, the volume (area) projected onto the parameter space is smaller than that of the quadrangular pyramid of the third embodiment. When a ReLU field is generated in the parameter space, the area that can be selected is narrowed by the projected volume. Therefore, according to the fifth embodiment, more candidate points can be searched for. EXAMPLES

[0102] Next, a sixth embodiment will be described. In the sixth embodiment, differences from the first to fifth embodiments will be mainly described, and therefore descriptions of common parts with the first to fifth embodiments will be omitted. The sixth embodiment is an example in which a cost adjustment layer is connected between the inference model 301 and the cost addition node 303 in the graph structure.

[0103] FIG. 19 is an explanatory diagram showing an example of a graph structure according to the fifth embodiment. In the fifth embodiment, there are two inference models 301a and 301b, but there may be only one inference model 301. When there is no need to distinguish between the inference models 301a and 301b, they are simply referred to as inference model 301. Similarly, the suffixes a and b are omitted for the output costs C1ia, C1ib, AD(C1ia), and AD(C1ib). Also, FIG. 18 shows, as an example, a state in which distance cost nodes 302-1 and 302-2 are connected.

[0104] Cost adjustment layers 1900a and 1900b are connected between the inference models 301a and 301b and the cost addition node 303. When the cost adjustment layers 1900a and 1900b are not to be distinguished from each other, they are simply referred to as the cost adjustment layer 1900. The cost adjustment layer 1900 adjusts the cost C1i output from the inference model 301 and outputs AD(C1i) to the cost addition node 303.

[0105] Fig. 20 is an explanatory diagram showing an example of the cost adjustment layer 1900. The cost adjustment layer 1900 has a subtraction unit 2001, a ReLU activation unit 2002, a multiplication unit 2003, an addition unit 2004, and a summing unit 2005. Each of the subtraction unit 2001, the ReLU activation unit 2002, the multiplication unit 2003, and the addition unit 2004 has j blocks, and the j-th block is connected in series. Fig. 19 shows an example where j=5. The j-th blocks of the subtraction unit 2001, the ReLU activation unit 2002, the multiplication unit 2003, and the addition unit 2004 are denoted as 2001-j to 2004-j, respectively.

[0106] C1ij, which is an element of axis j of cost C1i from the inference model 301, is input to the j-th subtraction block 2001-j of the subtraction unit 2001. The summation unit 2005 adds the addition results from the addition block 2004-j, and outputs the result as a cost adjustment result AD(C1i) to the cost addition node 303. If the cost C1i is x and the cost adjustment result AD(C1i) is y, the cost adjustment layer 1900 is expressed by the following formula (4).

[0107]

Number

[0108] FIG. 21 is an explanatory diagram showing an example of a graph indicating the cost input-output relationship. Specifically, for example, the subtraction unit 2001 subtracts the x-axis of the control points p[0],..., p[N] from the input x to calculate (x - p[i].x). Next, the ReLU activation unit 2002 applies ReLU to the subtracted value (x - p[i].x) to calculate ReLU(x - p[i].x). Next, the multiplication unit 2003 multiplies by the slope grad of each region corresponding to between adjacent control points in the x-axis direction to calculate grad[i] × (ReLU(x > p[i].x)).

[0109] Next, the addition unit 2004 adds the offset of each region corresponding to between adjacent control points in the x-axis direction to calculate |grad[i] × (ReLU(x > p[i].x)) + offset[i]|. Finally, the summation unit 2005 takes the sum and outputs it.

[0110] FIG. 22 is an explanatory diagram showing another example of the cost adjustment layer 1900. FIG. 23 is an explanatory diagram showing another example of a graph indicating the cost input-output relationship. In FIG. 22, the slope grad[N] is grad[N] = 0. As a result, it is possible to extend the gradient to the region on the right side of the rightmost control point, that is, the region where (max(p.x) < x). Also, a branch 2200 in which the node of x - p[0].x is inverted is added between the subtraction unit 2001 and the ReLU activation unit 2002. As a result, it is possible to extend the gradient to the region on the left side of the leftmost control point, that is, the region where (min(p.x) > x). In the branch 2200, the value of the multiplication node becomes -grad[0] and the offset value becomes 0.

[0111] In this way, by adopting the cost adjustment layer, it is possible to finely adjust the orientation of the candidate point Pi. In addition, the cost adjustment layer may be changed in the middle of the repetition of obtaining the candidate point Pi. By this operation, it is also possible to adjust the cost C3i of the five points added to the candidate point output list in the first half so that it is lower than a certain value, and to adjust the cost C3i of the five points added to the candidate point output list in the second half so that it is equal to or higher than the certain value. In this way, a group of candidate points incorporating multiple strategies can be obtained.

[0112] In addition, the structure of the cost adjustment layer can be freely chosen as long as the input and output dimensions are equal, but if it is implemented in a configuration including the ReLU activation unit 2002 and the summation unit 2005, control points appear, making it easy to intuitively control and useful.

[0113] In this way, according to the above-mentioned first to sixth embodiments, the searched candidate points can be separated from each other by a graph structure in which distance cost nodes are additionally connected, and multiple solutions that are experimentally meaningful can be obtained. Also, since it is implemented as a part of a neural network, it is possible to efficiently search for candidate points Pi using a trained inference model 301. Also, it is easy to customize the separation distance th, the termination criterion, the upper limit number of gradient descent counts, the grid spacing d, and the cost adjustment layer 1900, and it is easy to adjust the tendency of the candidate solutions, making it suitable for experimental planning.

[0114] The present invention is not limited to the above-described embodiments, and includes various modified examples and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those having all of the configurations described. Also, a part of the configuration of one embodiment may be replaced with a configuration of another embodiment. Also, a configuration of another embodiment may be added to a configuration of one embodiment. Also, a part of the configuration of each embodiment may be added, deleted, or replaced with another configuration.

[0115] Furthermore, each of the aforementioned configurations, functions, processing units, processing means, etc. may be realized in hardware, for example by designing some or all of them as an integrated circuit, or may be realized in software by a processor interpreting and executing a program that realizes each function.

[0116] Information such as programs, tables, files, etc. that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0117] In addition, the control lines and information lines shown are those considered necessary for the explanation, and do not necessarily show all the control lines and information lines necessary for implementation. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0118] 100 Experimental Planning System 101 Experimental Planning Equipment 102 Learning Device 103 Experimental Equipment 301 Inference Model 300 Input node (input section) 302-i Distance cost node (distance cost calculation unit (step S403), quantization unit (step S1600)) 303 Cost addition node (cost addition unit (step S403)) 304 Loss function output node (modification unit (step S403), determination unit (step S404), setting unit (step S407)) 500 Design Parameter Space 1900 Cost Adjustment Layer

Claims

1. An experimental planning device having a processor for executing a program and a storage device for storing the program, the device being capable of communicating with an experimental device, The processor, an initial correction process in which initial design parameters in a design parameter space for the experimental device are input as input parameters into an inference model for evaluating the experimental device, an initial evaluation cost is calculated as the evaluation result, and the initial design parameters are repeatedly corrected by gradient descent based on the initial evaluation cost until a first condition regarding the number of corrections is satisfied; an initial candidate point addition process for adding the modified initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether or not the number of the candidate points in the candidate point output list exceeds a predetermined number; a distance cost node addition process for adding a distance cost node that sets the candidate point last added to the candidate point output list as a starting point and calculates a distance cost at a new position in the design parameter space obtained by gradient descent from the starting point when the number of the candidate points in the candidate point output list does not exceed the predetermined number; a correction process in which a design parameter which is the candidate point last added to the candidate point output list is input as the input parameter into the inference model to calculate an evaluation cost which is the evaluation result, a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes added so far, and the correction of the design parameter by gradient descent of the sum of the evaluation cost and each distance cost is repeatedly performed until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold value is satisfied; a candidate point addition process for adding a modified design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; an output process of outputting the candidate points to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number; An experimental design apparatus for executing the above.

2. An experimental planning device having a processor for executing a program and a storage device for storing the program, the experimental planning device being capable of communicating with an experimental device, The processor, an initial correction process in which initial design parameters in a design parameter space for the experimental device are input as input parameters into an inference model for evaluating the experimental device, an initial evaluation cost is calculated as the evaluation result, and the initial design parameters are repeatedly corrected by gradient descent based on the initial evaluation cost until a first condition regarding the number of corrections is satisfied; an initial candidate point addition process for adding the modified initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether or not the number of the candidate points in the candidate point output list exceeds a predetermined number; a distance cost node activation process for activating one inactive distance cost node among a plurality of inactive distance cost nodes as a distance cost node that calculates a distance cost at a new position in the design parameter space obtained by gradient descent from the starting point by setting the candidate point last added to the candidate point output list as a starting point when the number of the candidate points in the candidate point output list does not exceed the predetermined number; a correction process in which a design parameter which is the candidate point last added to the candidate point output list is input as the input parameter into the inference model to calculate an evaluation cost which is the evaluation result, a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes activated so far, and the correction of the design parameter by gradient descent of the sum of the evaluation cost and each distance cost is repeatedly performed until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold value is satisfied; a candidate point addition process for adding a modified design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; an output process of outputting the candidate points to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number; An experimental design apparatus for executing the above.

3. An experimental planning apparatus according to claim 1 or claim 2, The distance cost node addition process adds a ReLU field including the starting point to the design parameter space; In the modification process, the processor modifies the design parameters by performing gradient descent along the boundary of the ReLU field outside the ReLU field. An experimental planning device characterized by:

4. An experimental planning apparatus according to claim 1 or 2, the design parameter space is gridded; the gradient descent is performed on the gridded design parameter space for the quantized design parameters; An experimental planning device characterized by:

5. An experimental planning apparatus according to claim 1 or 2, the distance cost is calculated based on a Euclidean distance between the starting point and a new position of the design parameter; An experimental planning device characterized by:

6. An experimental planning apparatus according to claim 1 or 2, In the initial modification process, the processor adjusts the evaluation results from the inference model to calculate the initial evaluation cost; In the modification process, the processor adjusts the evaluation result from the inference model to calculate the evaluation cost. An experimental planning device characterized by:

7. An experimental planning method executed by an experimental planning device having a processor that executes a program and a storage device that stores the program, the experimental planning device being capable of communicating with an experimental device, comprising: The processor, an initial correction process in which initial design parameters in a design parameter space for the experimental device are input as input parameters into an inference model for evaluating the experimental device, an initial evaluation cost is calculated as the evaluation result, and the initial design parameters are repeatedly corrected by gradient descent based on the initial evaluation cost until a first condition regarding the number of corrections is satisfied; an initial candidate point addition process for adding the modified initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether or not the number of the candidate points in the candidate point output list exceeds a predetermined number; a distance cost node addition process for adding a distance cost node that sets the candidate point last added to the candidate point output list as a starting point and calculates a distance cost at a new position in the design parameter space obtained by gradient descent from the starting point when the number of the candidate points in the candidate point output list does not exceed the predetermined number; a correction process in which a design parameter which is the candidate point last added to the candidate point output list is input as the input parameter into the inference model to calculate an evaluation cost which is the evaluation result, a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes added so far, and the correction of the design parameter by gradient descent of the sum of the evaluation cost and each distance cost is repeatedly performed until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold value is satisfied; a candidate point addition process for adding a modified design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; an output process of outputting the candidate points to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number; A method for designing an experiment, comprising:

8. An experimental planning method executed by an experimental planning device having a processor that executes a program and a storage device that stores the program, the experimental planning device being capable of communicating with an experimental device, comprising: The processor, an initial correction process in which initial design parameters in a design parameter space for the experimental device are input as input parameters into an inference model for evaluating the experimental device, an initial evaluation cost is calculated as the evaluation result, and the initial design parameters are repeatedly corrected by gradient descent based on the initial evaluation cost until a first condition regarding the number of corrections is satisfied; an initial candidate point addition process for adding the modified initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether or not the number of the candidate points in the candidate point output list exceeds a predetermined number; a distance cost node activation process for activating one inactive distance cost node among a plurality of inactive distance cost nodes as a distance cost node that calculates a distance cost at a new position in the design parameter space obtained by gradient descent from the starting point by setting the candidate point last added to the candidate point output list as a starting point when the number of the candidate points in the candidate point output list does not exceed the predetermined number; a correction process in which a design parameter which is the candidate point last added to the candidate point output list is input as the input parameter into the inference model to calculate an evaluation cost which is the evaluation result, a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes activated so far, and the correction of the design parameter by gradient descent of the sum of the evaluation cost and each distance cost is repeatedly performed until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold value is satisfied; a candidate point addition process for adding a modified design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; an output process of outputting the candidate points to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number; A method for designing an experiment, comprising:

9. An experimental planning system having an experimental apparatus and an experimental planning apparatus, The experimental planning apparatus includes a processor that executes a program, a storage device that stores the program, and a communication interface that can communicate with the experimental apparatus; The processor, an initial correction process in which initial design parameters in a design parameter space for the experimental device are input as input parameters into an inference model for evaluating the experimental device, an initial evaluation cost is calculated as the evaluation result, and the initial design parameters are repeatedly corrected by gradient descent based on the initial evaluation cost until a first condition regarding the number of corrections is satisfied; an initial candidate point addition process for adding the modified initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether or not the number of the candidate points in the candidate point output list exceeds a predetermined number; a distance cost node addition process for adding a distance cost node that sets the candidate point last added to the candidate point output list as a starting point and calculates a distance cost at a new position in the design parameter space obtained by gradient descent from the starting point when the number of the candidate points in the candidate point output list does not exceed the predetermined number; a correction process in which a design parameter which is the candidate point last added to the candidate point output list is input as the input parameter into the inference model to calculate an evaluation cost which is the evaluation result, a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes added so far, and the correction of the design parameter by gradient descent of the sum of the evaluation cost and each distance cost is repeatedly performed until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold value is satisfied; a candidate point addition process for adding a modified design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; an output process of outputting the candidate points to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number; An experimental design system comprising:

10. An experimental planning system having an experimental apparatus and an experimental planning apparatus, The experimental planning apparatus includes a processor that executes a program, a storage device that stores the program, and a communication interface that can communicate with the experimental apparatus; The processor, an initial correction process in which initial design parameters in a design parameter space for the experimental device are input as input parameters into an inference model for evaluating the experimental device, an initial evaluation cost is calculated as the evaluation result, and the initial design parameters are repeatedly corrected by gradient descent based on the initial evaluation cost until a first condition regarding the number of corrections is satisfied; an initial candidate point addition process for adding the modified initial design parameters that satisfy the first condition to a candidate point output list as candidate points to be applied to the experimental apparatus; a determination process for determining whether or not the number of the candidate points in the candidate point output list exceeds a predetermined number; a distance cost node activation process for activating one inactive distance cost node among a plurality of inactive distance cost nodes as a distance cost node that calculates a distance cost at a new position in the design parameter space obtained by gradient descent from the starting point by setting the candidate point last added to the candidate point output list as a starting point when the number of the candidate points in the candidate point output list does not exceed the predetermined number; a correction process in which a design parameter which is the candidate point last added to the candidate point output list is input as the input parameter into the inference model to calculate an evaluation cost which is the evaluation result, a distance cost at a new position of the design parameter in the design parameter space obtained by gradient descent of the design parameter from the starting point using each of the distance cost nodes activated so far, and the correction of the design parameter by gradient descent of the sum of the evaluation cost and each distance cost is repeatedly performed until the first condition or the second condition that the differential value of the sum is equal to or less than a threshold value is satisfied; a candidate point addition process for adding a modified design parameter that satisfies the first condition or the second condition to the candidate point output list as a new candidate point; an output process of outputting the candidate points to the experimental apparatus when the number of the candidate points in the candidate point output list exceeds the predetermined number; An experimental design system comprising:

Images