JP3421552B2 - Look-up table design method and recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses various look-up tables such as a parameter look-up table that decides a control parameter to be used based on a sensor input and a decision table that decides an appropriate strategy based on an observed situation. A general-purpose lookup table design method for automatic design or design support using a forward calculation model or calculation simulator, especially because it is not possible to obtain the output value in real time by calculation from the input value online When it is necessary to decide the output value in the lookup table and store it in the lookup table, a general-purpose tool for automating or assisting the design of the contents to be stored in the lookup table (cases of input value and output value in each case) Various lookup table design methods That.

[0002]

2. Description of the Related Art A look-up table can be easily implemented as hardware and can operate at high speed, for example, as an associative memory using means for quantizing an input and converting it into an address, and an operation test is also easy. , Traffic signal control parameter adjusting device, elevator county management mode determining device, home appliance control parameter adjusting device, etc. are used in an extremely wide range of fields, centering on systems that require real-time performance.

Here, the lookup table divides an input into several cases, stores an output corresponding to each case, and when an arbitrary input is given, searches an output corresponding to the output and outputs the output. A software module or a hardware module that can be used.

As a method of using the look-up table, online / real time calculation is performed from input values.
Since it is not possible to automatically determine the output value, it is necessary to determine the output value in advance, and it is possible to automatically determine the output value by calculation, but to save the calculation hardware, look There are two types of purpose of use: storing in an up table and using. In the following, the lookup table used for the former purpose will be treated.

In order to create a look-up table (for example, a parameter look-up table that determines the control parameter to be used based on the sensor input, a decision table that determines an appropriate strategy based on the observed situation, etc.) It is necessary to design cases and output values in each case. In a conventional general look-up table design, a designer classifies input information in advance based on know-how and determines output values corresponding to each case by another means. Here, a conventional look-up table designing method will be described by taking a look-up table used for traffic signal control as an example.

For the control of traffic signals in Japan, regarding the traffic signal group belonging to the intersection group (system) to be controlled, information on vehicle detector groups provided at major points (for example, representative intersections) in the system is input. A method called pattern selection control is designed in which a look-up table that outputs control parameters for the traffic signal group is designed, and control parameters are selected from the look-up table at regular time intervals according to traffic conditions and used for control. It is common. In this case, as the control parameters, "cycle" indicating the interval of the green signal start time, "split" indicating the ratio of the green time given to each traffic route intersecting each other, and the deviation of the green start time between the traffic lights in the system The three types of “offset” shown are control parameters that are mainly used. In traffic signal control, one of the goals is to make the flow of the vehicle smoother by adjusting these control parameters according to the traffic conditions. When designing a look-up table, these control parameters can be used for various traffic conditions. It is necessary to make decisions according to each.

A brief description will be given of a design method of the look-up table in the case of this example. First, as a first procedure, the input vector space for the look-up table is divided. That is, the traffic volume at the representative intersection, which is an input given by the vehicle detector group, is classified into several types of predetermined patterns such as quiet, upward direction priority, and downward direction priority. FIG. 16 shows an example of division that is often used. In FIG. 16, the two-dimensional input vector space of the up traffic index (up traffic volume) and the down traffic index (down traffic volume) is divided into quiet, two types of up priority, two types of equality, and two types of down priority. It is divided into a total of 7 sections.

This division of the input vector space is performed based on know-how, and there is no general division method.

Next, as a second procedure, an output vector corresponding to each pattern defined by the above division is defined.
That is, for each pattern, a typical traffic situation in that pattern (in the case of FIG. 16, a representative value of the upstream traffic index and a typical value of the downstream traffic index in each pattern) is set, and, for example, a design method called through band maximization The "cycle", "split" and "offset" suitable for the typical traffic situation are determined by an optimization method using a road traffic simulator or the like, and are fitted into the corresponding patterns.

In the optimization using the above-mentioned street traffic simulator, for example, a design support tool called TRANSYT is mainly used. TRANSYT was developed by the British Transport and Road Research Institute (TRRL) to support street traffic phenomenon analysis and signal parameter design.
It is the most widely used tool in the world. This T
RANSYT uses a macro model that describes traffic flow on a vehicle group basis, sets average traffic volume for each direction on the road network,
It has the function of simulating the movement of the vehicle group within the cycle period, and the function of calculating the evaluation value based on the delay and the number of stops generated at each signalized intersection by simulation and optimizing the offset and split by the hill climbing method. As the evaluation criteria for optimization, evaluation values P. I. And the evaluation value P.
I. Find the parameter that minimizes. P. I. = Σ (delay time + α × number of stops) Through such a procedure, the look-up table for traffic signal control is designed.

[0011] The above contents are edited and published by the Traffic Engineering Research Society of Japan, "Guide of Traffic Signals" and edited by Koshi.
It is explained in detail in “Traffic Engineering” published by Technical Shoin.

Similar to the method of designing the look-up table for traffic signal control described above, generally in designing a look-up table, the input vector space is divided (input) by a method based on experience or know-how or a trial-and-error method. Vector case), and then the representative value of the input vector in each division is determined, for example, by using the central value of each division, and the optimum output corresponding to each representative value is determined in the forward direction. The most widely used method is to determine the output evaluation value using a model and determine it by trial and error iteration of verification by a calculation simulator and store it as the output of each division.

The forward model and the calculation simulator mean an arithmetic expression, a software module, or a hardware for which the corresponding output is uniquely obtained by calculation when the internal parameters and the input vector to the simulator (the output vector of the lookup table) are determined. A wear module is shown, and a one-to-many mapping whose corresponding output is not uniquely determined even if the same internal parameter and input vector are specified is not included.

However, the conventional lookup table designing method as described above has the following two major problems.

(Problem 1) Points requiring time and effort for design When designing a new type of look-up table without know-how in the past, 1) determination of division of input space 2) determination of output value in each division 3) Since it is necessary to repeat the three steps of simulation or experiment verification as to whether or not the entire lookup table works without problems by trial and error, the number of iterations of the entire work becomes enormous, and the design is troublesome. There is a problem. In particular, when information from a plurality of sensors is input and there are various types of inputs, there is generally no method for analytically obtaining the output, and the design is unavoidably trial-and-error, which often requires time and effort. For example, in the elevator group management control, a lookup table for determining the priority weight that determines which call to give priority to by using all call information, call frequency, time, etc. as an input is an analytical priority weight. It is difficult to design because there is no method for obtaining

(Problem 2) It is difficult to determine a robust representative value in each division of the input vector space. As an input vector of the point lookup table, an observation value obtained by a sensor or information obtained by processing the observation value is used. However, the information observable by these sensors is often insufficient to completely estimate the state of the system S that is the output target (that is, the control target) of the lookup table. In other words, if there is an unobservable internal parameter (hereinafter referred to as internal parameter p) whose state cannot be estimated based on the observed value obtained by the set sensor in the system S, the representative value of each division The internal parameter p (p is a vector) to be set in the forward model or the simulator is not uniquely determined from (the median value, etc.), and the representative value and the internal parameter p have a one-to-many relationship.
More specifically, in general, the observable value r | p under a certain internal parameter p is uniquely obtained, but conversely, the observed value r |
Since the internal parameter p | r under the observation of is backward processing, it cannot be uniquely determined.

In this case, in the conventional method, first, a representative observable value r of division of the input vector space is determined, and one internal parameter p1 (p1 is an element of P) from the set of internal parameters P | r for r. This internal parameter p1 is set in the forward model or simulator of the system S, the optimum output under this internal parameter p1 is obtained by the search method, and stored in the lookup table. Therefore, if the internal parameter p of the system S when the look-up table is used is significantly different from the assumed value p1, the output may not be effective and there is a problem in robustness. Furthermore, in a simulation including non-linear processing, the set P |
It is often difficult to even find r.

This point will be specifically described by taking a traffic signal control system as an example. The example shown in FIG. 17 is 1-1.
This is an example of a system consisting of 15 intersections and 31 links numbered 5. Note that the 31 links flow into the intersection 1 from the down direction, the up link and the down link sandwiched between adjacent intersections 1 to 15, and the link 15 flows from the up direction into the intersection 15. Incoming links are those that flow in from the side at intersection 8.

To simplify the description, this system has a link that flows in a large amount from the side at the central intersection 8, and the signal control device of the intersection county has two vehicle detectors provided at the main intersection 1. The amount of uphill traffic and the amount of downhill traffic obtained respectively are measured, and the "offset" of all intersections in the system is determined by a lookup table. Here, it is assumed that the offset is the amount of deviation of the blue start time of each intersection based on the time when blue starts at the main intersection 1 (in this example, 14 offsets are set). ).

In this example, it is impossible to estimate the ratio of the inflow traffic from the side in the up traffic obtained by the vehicle detector with only two vehicle detectors.
Therefore, in the conventional method, an internal parameter (three traffic volumes) is determined by making an assumption that half of the upstream traffic volume is inflow traffic, and is set to the above-mentioned TRANSSYT to optimize the offset by searching. By doing so, the offset of each division is determined. Therefore, if the actual inflow traffic is near the assumed value (for example, half the upflow traffic), the lookup table works effectively, but the actual inflow traffic deviates significantly from the assumed value. When it comes,
The look-up table may not work effectively, and there is a problem in robustness.

When the look-up table is used, the output vector determined for the division to which the input vector belongs is output. That is, the output vector determined for each division is widely applied to the system S even for an input vector whose distance in the input vector space is far from the representative value of the input vector used for determining the output vector. Therefore, when each division of the input vector space is large (rough division), the output vector optimized for the representative value does not always function effectively even for the input vector far from the representative value. There is.

[0022]

As described above, in the conventional look-up table designing method, for example, when designing a look-up table of a new type that has no past know-how, determination of division of the input space, The number of repetitions of the whole work is enormous because it is necessary to repeat the three steps of determining the output value in each division and verifying by simulation or experiment as to whether the entire lookup table works without problems. Therefore, there is a problem that it takes time to design. Moreover, there is a problem that it is difficult to determine a robust representative value in each division of the input vector space.

The present invention has been made in consideration of the above circumstances, and has the most suitable and robust look depending on a given combination of input-output variables regardless of the amount of available input information. It is an object of the present invention to provide a look-up table design method capable of automatically designing or supporting design of the contents of a look-up table.

[0024]

SUMMARY OF THE INVENTION The present invention generally comprises:
The content of the lookup table is automatically designed by off-line learning on the computational simulation model or the forward model. The present invention simplifies and mimics the “mechanism of the acquired system that produces a wide variety of antibodies against a vast variety of antigens” of the immune system. In other words, by imitating a part of the principle of the immune system of ecology, the heuristic search means and the memory means are integrated, and the off-line learning is performed so as not to destroy the diversity of knowledge. However, since the present invention does not directly correspond to the detailed principle of the immune system, the description of the living body immune system itself will be omitted here. D. Farmer, N.M. H. Packard,
"The immune system, adapta
ion, and machine learnin
g ", Physica D22, pp. 187-20
4, 1986 and the like.

In the present invention, a genetic algorithm (GA) is used.
As in the case of applying the coding unit of information to the gene in, the coding unit of the input vector and the output vector for the lookup table is applied to the antibody. That is, an antibody means a coding unit of an input vector and an output vector. Antibodies are expressed as numbers or symbolic vectors.

In the present invention, the index of effectiveness is applied to the antibody concentration. That is, the antibody concentration has the same meaning as the reinforcement value in a method based on reinforcement learning in general, and serves as a standard for selecting antibodies.
Reinforcement learning is a type of machine learning that adapts to the environment by updating the weights to prioritize outputs that give rewards as inputs, and learning is performed even if a clear teacher output is not given. It has the feature that

The evaluation value of the antibody is evaluated from the result of executing the calculation simulation or the forward model calculation under the condition of a certain internal parameter for the portion corresponding to the output vector of the antibody. It is a scale that indicates whether the target system could be successfully controlled by, and is a value determined by the antibody alone.

The look-up table designing method according to the present invention (Claim 1) performs the learning for examining the relationship between the input vector and the output vector of the look-up table used for controlling the desired system, and then the result of the learning. A lookup table designing method for determining the contents of the lookup table based on
A plurality of antibody data including a pair of an input vector and an output vector with respect to a target lookup table, which are generated upon the establishment of a predetermined condition, are stored in association with a concentration indicating the relative superiority thereof. From the antibody information storage means, a plurality of antibody data to be evaluated is determined by obtaining antibody data having an input vector included in a predetermined neighborhood range on the input vector space,
The evaluation value is obtained for each individual antibody data subject to evaluation by a predetermined method, and the relative superiority of the evaluation value of each antibody data in consideration of all the obtained evaluation values is taken as a reference, and each antibody data is evaluated. It is characterized by updating the density of.

Further, the lookup table designing apparatus according to the present invention (Invention 1) is an antibody memory for storing a plurality of antibodies, which are coding units of input and output vectors to the lookup table, and their concentrations in association with each other. Means, data changing means for rewriting the antibody concentration stored in the antibody storage means to an updated value, candidate antibody producing means for producing a new antibody to be added to the antibody storage means, and the designated antibody And an antibody evaluation value determining means for obtaining an evaluation value of, a parameter input means for inputting a learning parameter and a look-up table design parameter designated by the user, and managing an antibody learning process based on the input learning parameter. , A lookup table based on the entered lookup table design parameters The learning / design management means for managing the measurement process and the lookup table creation means for determining the contents of the lookup table based on the contents stored in the antibody storage means after the learning process are completed are provided. When updating the concentration of an antibody, a part corresponding to the input vector of the antibody in the antibody group located in the vicinity in the input vector space is based on the relative merit of the evaluation value of the antibody. The concentration of the antibody is updated.

According to the present invention, in the learning process, the candidate antibody producing means generates a variety of coding units (antibodies) of the input vector and the output vector, which are the basis of the contents of the lookup table, and stores them in the antibody storing means. In addition, each neighborhood antibody is evaluated independently by the antibody evaluation value determination means within the neighborhood based on the input vector of a certain step, and then the concentration of each neighborhood antibody is determined based on the relative superiority between the neighborhood antibodies. Is updated by means of data modification, increasing the concentration of antibody expected to have the ability to perform better control while maintaining antibody diversity, and lookup after the learning process. In the table design process, the contents of the lookup table are determined based on the contents stored in the antibody storage means. Also it is obtained by allowing design a lookup table having excellent appropriate and robust without hassle without.

Further, according to the present invention, even if the designer does not prepare the teacher data for learning like a neural network, the lookup table can be automatically designed by repeating the trial and error autonomously on the calculation simulation. It will be possible.

The present invention (Invention 2) is the same as the above Invention 1, except that the input vector space is divided into meshes of a size specified for each dimension of the input vector of the lookup table, and the input vector input as the search key is obtained. The neighborhood antibody that retrieves / extracts the coding contents or the antibody storage address information of all the neighborhood antibodies contained in the neighborhood mesh within the neighborhood of the size specified for each dimension with respect to the mesh containing Search means is further provided, and the evaluation values of all the vicinity antibodies extracted by the vicinity antibody search means are respectively obtained by the antibody evaluation value determination means, and the extraction is performed by a predetermined calculation using all the obtained evaluation values. It is characterized in that the concentration is updated for each of the generated neighboring antibodies.

Here, an example of detection by the neighborhood antibody search means will be described. When the input vector is two-dimensional, the number of mesh divisions is eight in both dimensions, and the size of the neighborhood is one mesh in both dimensions, FIG. In, it seems that the coordinates corresponding to the input vector of the antibody stored in the antibody storage means are represented by the black circles in the input vector space, and the coordinates occupied by the input vector input as the search key are x.
As indicated by the marks, the hatched 9
Antibodies (in this case, four antibodies) included in one mesh (neighboring mesh) are extracted as the neighboring antibodies by the neighboring antibody search means.

In the present invention, by limiting the range in which the antibody competes to within the neighborhood mesh of the input vector space, the diversity in the input vector space is maintained and each mesh (each address of the lookup table) is supported. Antibodies to be obtained will be evenly obtained. That is, when the antibody evaluation value determination means uses the calculation simulator or the forward model for the evaluation of the antibody, any of the observed values r | p that can be observed by the fluctuation of the internal parameter p in the calculation simulator or the forward model. Corresponding antibodies are uniformly obtained.

The present invention (Invention 3) is the above Invention 2, wherein the evaluation values of all the neighboring antibodies extracted by the neighboring antibody searching means are obtained by the antibody evaluation value determining means, and all the obtained evaluation values are obtained. The average value is calculated from
The concentration of the antibody whose concentration is to be updated is updated according to the magnitude relationship between the evaluation value of the antibody whose concentration is to be updated and the average value.

According to such a concentration updating method, even when there is a large variation in the market value of the evaluation value of the antibody depending on the mesh, an antibody having an excellent evaluation value relative to the market value in the vicinity can be obtained. Since the density increases, it is not necessary to set an absolute criterion for the evaluation value for each mesh, and the learning parameters can be set significantly easily. For example, the evaluation index P. I. When is used as an evaluation value, the reference value for good or bad changes significantly according to the average of the total traffic volume of the mesh, so the density increase / decrease standard differs depending on the mesh. Need not be set individually.

As an example of the density updating method, there is a method of updating the density based on the following formula. Where Fj is antibody j
Is the evaluation value of. In the following equation, the larger the evaluation value, the better the evaluation. n is the number of antibodies in the vicinity, Cj is the concentration of antibody j, and Δ is the concentration increase / decrease value. ΣFk / n ≦ Fj → Cj = Cj + Δ ΣFk / n> Fj → Cj = Cj−Δ Here, the range of the total sum of Fk is K = 1 to n.

Alternatively, based on the following formula, there is a method in which the concentration is regarded as the winning rate of each antibody and the concentration is changed. Here, Sj is the number of steps that have passed since the antibody was added to the database, and is stored in the antibody storage means. ΣFk / n ≦ Fj → Cj = (SjCj + 1) / (S
j + 1) ΣFk / n> Fj → Cj = SjCj / (Sj + 1) Here, the range of the sum of Fk is K = 1 to n.

As other density updating methods, various methods such as a method of using a weighted average according to the Euclidean distance between the antibody and the input in the input vector space, a method of increasing or decreasing the value according to the rank in the neighboring antibody from the density, etc. The method of can be considered.

According to the present invention (claim 2), in the invention according to claim 1, an internal parameter of the first calculation simulator or the forward model is randomly or preliminarily specified every time each learning is repeated. The first calculation simulation or the forward model calculation is changed, and observable information designated in advance is extracted from the result of the first calculation simulation or the forward model calculation, and this is converted to obtain the input vector. A mesh including the input vector among a plurality of meshes formed by dividing the input vector space into the number of divisions specified for each dimension of the input vector of the lookup table is used as a starting point, and To obtain all antibody data contained in the neighboring meshes by adding the specified number of adjacent meshes for each dimension Through this, a plurality of antibody data to be evaluated are determined, and for each antibody data to be evaluated, the value of the portion corresponding to the output vector of the coding content is converted to convert the second calculation simulator or forward model. And perform the second calculation simulation or the forward model calculation in which the same internal parameter as that of the first is set, and the evaluation value of each antibody data is calculated based on the result of the second calculation simulation or the forward model calculation. Characterized by seeking.

In the present invention (invention 4), in the above-mentioned invention 1, the input vector space is divided into meshes of a size specified for each dimension of the input vector of the lookup table, and the mesh is input as a search key. With respect to the mesh containing the input vector, the coding contents or the antibody storage address information of all the neighboring antibodies contained in the neighboring mesh within the neighborhood of the size designated for each dimension are searched and extracted from the antibody storage means. The antibody evaluation value manual decision stage further comprises a neighborhood antibody search means, and has a calculation simulator or a forward direction model, a parameter automatic changing means and a simulation result evaluation means, and for each learning step, the calculation simulator or the forward direction model The internal parameters are changed randomly or by a pre-specified procedure, and the first calculation Or forward direction model operation is performed, pre-specified observable information is extracted from the result of the first calculation simulation or the forward direction model operation, this is converted into an input vector, and the input vector is used as a search key. The neighborhood antibody search means extracts the neighborhood antibody from the antibody storage means, and for each of the extracted neighborhood antibodies, the value of the portion corresponding to the output vector of the coding content is converted and input to the calculation simulator or forward model. Then, the second calculation simulation or the forward model calculation is performed, and the evaluation value of each of the extracted neighboring antibodies is obtained based on the result of the second calculation simulation or the forward model calculation. .

In the present invention, in the first simulation for obtaining the input vector, the observable value vector r of the target system S of the look-up table corresponding to the internal parameter p is obtained, and the second vector for evaluating the antibody is obtained. By the simulation, under the setting of the internal parameter p, the evaluation value F | when the portion x corresponding to the output vector of the neighboring antibody of the input vector r ′ | p (corresponding to the observable value vector r) is input to the simulator It is possible to estimate p, x. When F | p, x is used as an antibody evaluation criterion and learning is performed while changing p, when p | r ′ is always in the vicinity of a certain input vector r ′, F is always F.
It is possible to increase the concentration of antibody with a relatively good and robust output x. This allows
It is possible to solve the problem 2 of the conventional method described above without the trouble of creating and analyzing the backward model of the target system S.

According to the present invention (Invention 5), in the above Invention 4, the candidate antibody producing means has observable information designated in advance from the result of the first simulation or the forward model calculation for each learning step. Is extracted and the input vector created by converting this and the output vector created by a predetermined procedure are both coded to generate a new candidate antibody. .

In the present invention, after the first simulation, the input vector r '| p is based on the specified observable value.
In addition to this, a predetermined procedure, for example, the optimum output vector x | p under the setting of p is obtained by a search method such as a hill climbing method, and finally both are combined and coded into an antibody. In addition to the hill climbing method, various methods are conceivable, such as a method for generating an output vector by means of mutation, crossover, etc. using random generation such as a genetic algorithm or another antibody in the vicinity as a template.

Further, according to the present invention, the internal parameter p
No antibodies other than those in the vicinity of the input vector, which may appear when f is changed, are not generated from the beginning. Therefore, it is possible to avoid creating / adding unused candidate antibodies, improve learning efficiency, and save memory.

The present invention (Invention 6) is the above Invention 4 or 5.
In the above, the candidate antibody producing means extracts the observable information designated in advance from the result of the first simulation or the forward direction model calculation and converts the input observable information for each learning step. It is characterized in that a new candidate antibody is generated only when the number of the neighboring antibodies extracted by inputting a vector as a search key into the searching means is lower than a predetermined specified number.

According to the present invention, it is possible to prevent uneven distribution of the number of antibodies existing near each mesh. If the number of neighboring antibodies is too large, it takes calculation time to evaluate the antibody at the time of updating the concentration, and if it is too small, relative evaluation in the neighborhood according to Invention 3 cannot be performed, but according to the present invention, it is always 2 or more. The appropriate number of antibodies, which is less than or equal to the number, is maintained.

The present invention (Invention 12) is the above Invention 4 or 5, wherein the candidate antibody producing means is designated in advance for each learning step based on the result of the first calculation simulation or forward model calculation. When the number of the neighboring antibodies extracted by inputting the input vector created by extracting the observable information and converting the observable information as the search key into the search means is less than a predetermined number specified in advance In addition, it is characterized in that new antibodies are repeatedly produced until the specified number is reached.

According to the present invention, the number of antibodies existing in the vicinity of each mesh is further averaged as compared with the case of Invention 6 above.

The present invention (Invention 7) is based on the above Inventions 1 to 6 and 1.
In item 2, in each learning step, an antibody having a concentration lower than the minimum threshold value among the antibodies having the updated concentration (neighboring antibody) is deleted from the antibody storing means.

The present invention makes it possible to prevent the maintenance of incompatible antibodies.

The present invention (Invention 13) is based on the above Inventions 1-6.
In 12, the antibody having the lowest concentration among the antibodies (neighboring antibodies) whose concentration has been updated is deleted from the storing means at each learning step. According to the present invention, maintenance of incompatible antibodies can be prevented. The present invention is preferably used in combination with Invention 12 described above.

The present invention (Invention 11) includes the above Inventions 1 to 7,
12 and 13, the size of the neighborhood used in the neighborhood antibody search means is gradually reduced according to the progress of learning.

According to the present invention, when the user cannot determine an appropriate size of the neighborhood, the neighborhood is made large while the number of antibodies is small, thereby increasing the chances of changing the concentration of the antibody of each mesh and making the robust antibody robust. Is the priority to get
By decreasing the neighborhood as learning progresses and the number of antibodies increases, priority will be given to obtaining antibodies specialized in a narrow area of the input vector space, and effective neighborhood size according to the progress of learning can be set. Can be set.

The present invention (Claim 3) relates to Claim 1 or 4.
In determining the contents of the lookup table, each of a plurality of meshes formed by dividing the input vector space into a division number specified for each dimension of the input vector of the lookup table. Regarding, from the antibody information storage means after the learning, all the antibody data contained in the neighborhood mesh formed by adding the target mesh and the number of adjacent meshes starting from this mesh and specified by each dimension. It is characterized in that a portion corresponding to the output vector of the coding content of the antibody data having the maximum density among them is written to the address corresponding to the target mesh of the lookup table.

The present invention (Invention 8) is the above Invention 1 or 4, wherein the lookup table creating means divides the input vector space into a size designated for each dimension of the input vector of the lookup table. By inputting the median value of the mesh as a search key to the neighboring antibody searching means, a portion corresponding to the output vector of the coding content of the antibody having the maximum concentration among the extracted neighboring antibodies is stored in the lookup table. The writing in the address corresponding to the mesh is performed sequentially for all the meshes, so that the antibodies stored in the antibody storage unit are converted into a lookup table.

According to the present invention, the output vector of the best antibody for each mesh can be stored in the look-up table and the learning result can be converted into the look-up table. Also, the lookup table obtained by the present invention can be implemented as the associative memory described above.

The present invention (claim 4) relates to claim 1 or 4.
In determining the contents of the lookup table, each of a plurality of meshes formed by dividing the input vector space into a division number specified for each dimension of the input vector of the lookup table. Regarding, from the antibody information storage means after the learning, all the antibody data contained in the neighborhood mesh formed by adding the target mesh and the number of adjacent meshes starting from this mesh and specified by each dimension. It is characterized in that the processing is performed to write the coding content of the antibody data having the maximum concentration among them into the lookup table.

The present invention (Invention 9) is the same as Invention 1 or 4, wherein the lookup table creating means divides the input vector space into a size designated for each dimension of the input vector of the lookup table. Enter the median value of the mesh as the search key into the neighborhood antibody search means, and write the coding content of the antibody with the highest concentration among the extracted neighborhood antibodies in the lookup table. It is characterized in that the antibodies stored in the antibody storage means are converted into a look-up table by sequentially performing the steps.

According to the present invention, the output vector of the best antibody for each mesh can be stored in a look-up table and the learning result can be converted into the look-up table. Further, the lookup table of the form obtained by the present invention cannot be implemented as an associative memory, and a means corresponding to the proximity antibody search means is required as the lookup table reading means, but it is obtained by the above-mentioned invention 8. In general, the amount of storage is smaller than that of the lookup table of the above format, and the lookup table can be created at the same time when the processing of the above-mentioned invention 8 is performed.

The present invention (Invention 10) is the above Invention 8 or 9, wherein the lookup table creating means uses the median value of the mesh of the input vector space as a search key for the neighboring antibody searching means at the time of creating the lookup table. It is characterized in that when there is no nearby antibody extracted at the time of input, the mesh is notified to the user.

As a means for notifying the user, a method of visually displaying the coordinates of the mesh median value in the input vector space on a graphical user interface can be considered. When a new candidate antibody is created by the candidate antibody production means, a mesh in which no antibody is present in the vicinity remains after learning. These are internal parameters p
It is either a mesh in which an input vector cannot appear in the neighborhood no matter how it is changed or learning is insufficient, but it is possible to provide the user with information for distinguishing the former from the latter. .

The present invention (Claim 5) includes Claims 1 and 2.
In the invention described in any one of 1, the plurality of the first
Processing for generating a plurality of the input vectors in parallel by the calculation simulator or the forward model, and obtaining an evaluation value of the individual antibody data to be evaluated for each of the generated plurality of input vectors Are executed in parallel by a plurality of the second calculation simulators or forward models.

The present invention (Invention 14) is the same as Invention 1 above.
In 13 to 13, a plurality of the antibody evaluation value determining means is provided,
It is characterized in that the processes in the respective antibody evaluation value determining means are executed in parallel.

According to the present invention, it is possible to shorten the learning time by parallel processing the simulation calculation or the forward model calculation which requires the longest calculation time.

The present invention (Claim 6) includes Claims 1 and 2.
In the invention described in any one of 1,
The user interface is used to display the configuration diagram of the system to be controlled / adjusted, and the information about the sensor configuration input by the graphical user interface and the observation values and lookup table obtained by the specified sensor configuration are displayed. It is characterized by receiving information defining a relationship with an input vector and designing a look-up table based on the inputted information.

Preferably, the integrated performance of the designed look-up table is evaluated and presented. This makes it possible to comprehensively support the observation system design and the look-up table design.

The information on the sensor configuration is the type of the sensor, the installation location, etc., and is input, for example, by dragging and dropping a sensor item indicating the desired sensor on the configuration diagram being displayed. The information that defines the relationship between the observed value and the input vector to the lookup table is, for example, the designation that the observed value of a certain sensor is used as the input vector, the designation that the observed value of a certain sensor is converted to the data of another unit, There are various options such as designation of averaging the observed values of a plurality of sensors, and for example, the desired sensor arranged on the drawing being displayed and the processing content thereof are selected by inputting with a mouse.

According to the present invention, it is not necessary to analyze the backward model of the target system S according to the type of information used as an input to the lookup table, and learning is performed by almost the same method regardless of the type of input vector. As a result, the user can select the input to the look-up table on the spot and immediately learn.

For example, when applied to a look-up table design for determining traffic signal control parameters, whether or not the road / lane / position where the vehicle detector is placed and the average value of plural vehicle detector information are used as an input, Furthermore, the user is allowed to freely select on the GUI how to combine relatively expensive heterogeneous sensors such as a vehicle speed sensor and an image processing device, and a lookup table automatically designed based on the specified input. By comparing the control adjustment performance, it is possible to support the design of the monitoring system according to the purpose.

The present invention (Invention 15) is the same as Invention 1 above.
14 to 14, the user uses information to be used as an input to the look-up table from among the observable sensor information and the information obtained by integrating some of the sensor information in the system to be controlled / adjusted. Can be selected using the graphical user interface of the parameter input means, and a look-up table corresponding to the selected input is automatically designed by learning, and at the same time the user can check the overall performance of the look-up table designed at the same time. It is characterized by providing comprehensive support for observation system design and lookup table design.

The present invention (Claim 7) includes Claims 1 to 6.
In the invention described in any one of 1,
The user interface is used to display the configuration diagram of the system to be controlled / adjusted, and the graphical user
It accepts information on the internal parameters of the system expressed in the form of an ambiguous probability distribution input by the interface, and when the information is input, the internal parameters given in the ambiguous form follow the given probability distribution. The feature is that learning is performed by randomly adding fluctuations.

This makes it possible to design a look-up table having robustness according to the degree of ambiguity.

That is, according to the present invention, for example, when it is applied to a look-up table design for determining traffic signal control parameters, the traffic volume and the average vehicle speed are determined to be almost constant by the traffic survey. The simulator internal parameters are fixed to concrete values, and there are ambiguous widths like the membership function in fuzzy theory for places with large fluctuations and places where specific values cannot be specified due to insufficient research. You can specify an internal parameter as the value. This facilitates the table design according to the ambiguity and the degree of fluctuation of the control adjustment target system S.

The present invention (Invention 16) is the same as Invention 1 above.
15 to 15, the internal parameters of the system to be controlled / adjusted can be input in the form of an ambiguous probability distribution using the graphical user interface of the parameter input means, and are given in an ambiguous form. It is characterized by automatically designing a robust lookup table according to the degree of ambiguity by randomly changing the internal parameters according to the probability distribution during learning.

The present invention (Claim 8) includes Claim 1 or Claim 2.
In the invention described in (1), a configuration diagram of a system to be controlled / adjusted is displayed using a graphical user interface, and the first information regarding the sensor configuration input by the graphical user interface, the control / adjustment target Information about each device that becomes the third information, the third information that defines the relationship between the observed value obtained by the specified sensor configuration and the input vector of the lookup table, and the output vector of the lookup table and the specified control / adjustment target It is characterized in that the fourth information defining the relationship with the output value to the device is accepted, and the lookup table is designed by learning based on the inputted information.

The information relating to the sensor configuration includes the type of sensor, installation location, information to be measured, etc. For example, by dragging and dropping an item indicating a desired sensor on the configuration diagram being displayed, the type of sensor And the detailed input information is specified by using the property input screen that appears according to the installation location. Further, the information regarding each device to be controlled / adjusted includes the device to be controlled and the amount to be operated. For example, click the desired device on the displayed configuration diagram, and enter the property according to the device. It is input by the operation of selecting using the screen. The default control method may be adopted for devices not specified.

According to the present invention, the input / output to / from the user lookup table is selected on the spot, the learning is immediately performed, and the performance of the created lookup table is displayed. It is possible to design an interactive lookup table by changing the input / output of the uptable, the sensor configuration, and the device to be controlled / adjusted.

For example, when applied to the design of a look-up table for determining a traffic signal control parameter, it is controlled by a traffic signal and a fixed parameter that adaptively change the parameter based on the road on which the vehicle sensor is placed, the position, and the sensor information. Control of a lookup table automatically designed under the specified conditions by allowing the user to freely specify on the GUI, such as the specification of traffic signals and the road to adaptively change the vehicle speed limit display according to the traffic volume. By comparing the performance, it is possible to support the design of the control system according to the purpose.

According to the present invention (claim 9), in the invention according to claim 8, the selection of a part corresponding to a processing function prepared in advance is accepted, and the selected part is visually identified by its type. It is displayed on the screen as a node corresponding to a possible processing function, and to the node corresponding to the input from the sensor specified on the system configuration diagram and each control / adjustment target device specified on the system configuration diagram. , A node corresponding to the lookup table to be learned is displayed on the screen, and the node corresponding to the desired input and the desired node input by the graphical user interface are displayed. A designation to connect with a node corresponding to the processing function by a link is accepted, and the third and fourth information is obtained based on the accepted designation. The features.

According to the present invention, information defining the relationship between the observed value obtained by the specified sensor configuration and the input vector of the lookup table, and the output vector of the lookup table and the specified control / adjustment target device It becomes easy to input the information that defines the relationship with the output value.

According to the present invention (claim 10), in the invention according to claim 2, the first calculation simulation of the control / adjustment target system or the graphical user interface capable of inputting / outputting in a tabular form is performed. A cell corresponding to each observation value that is the result of the forward model operation,
A cell corresponding to each output to the second calculation simulation or the forward model operation of the control / adjustment target system, a cell corresponding to each element of the input vector to the lookup table that is the learning target, and a learning target Corresponding to each element of the input vector to the lookup table, which displays at least a part of cells corresponding to each element of the output vector of a certain lookup table on the screen and is input by the graphical user interface. The control / adjustment target system, which accepts a specification of embedding an arithmetic expression for obtaining a value with a cell value corresponding to each observation value of the control / adjustment target system as an argument, and is input by the graphical user interface The output of the look-up table in the cell corresponding to each output to Accept the specification to embed the arithmetic expression or the default output value that takes the value of the cell corresponding to each element of the cuttle as a part of the argument, and input the observation value and the lookup table input vector obtained by these accepted specifications. The look-up table is designed by learning based on an arithmetic expression indicating the relationship between the output vector of the lookup table and the output value to the control / adjustment target device.

According to the present invention, the information defining the relationship between the observed value obtained by the sensor configuration and the input vector of the lookup table, and the output vector of the lookup table and the output value to the designated control / adjustment target device It becomes easy to input and change the information that defines the relationship with.

According to the present invention (claim 11), in the invention according to claim 10, the first and second systems of the control / adjustment target system are displayed on a graphical user interface capable of inputting and outputting in a tabular format. A cell corresponding to the internal parameter of the calculation simulation or the forward model calculation, some cells corresponding to the calculation result of the second calculation simulation or the forward model calculation of the control / adjustment target system, and each of the antibody data At least a part of the cells corresponding to the evaluation value is displayed on the screen, and the value of the cell corresponding to the internal parameter is argument to the cell corresponding to each of the observation values input by the graphical user interface. The calculation formula corresponding to the first calculation simulation or the forward model calculation for obtaining the observation value as And specify the writing because, the graphical
Designation of embedding a range for changing the internal parameter and a constraint expression in a cell corresponding to the internal parameter input by the user interface, and input by the graphical user interface,
The cell corresponding to the calculation result of the second calculation simulation or the forward direction model calculation includes the value of the cell corresponding to the internal parameter and the cell corresponding to each output to the second calculation simulation or the forward direction calculation. Corresponding to the evaluation value of each of the antibody data input by the graphical user interface when the specification of embedding an arithmetic expression corresponding to the second calculation simulation or the forward model calculation for obtaining the calculation result by using the value as an argument Accepting at least one of the specifications for embedding an evaluation formula for obtaining an evaluation value in the cell by using the values of some cells corresponding to the calculation result of the second calculation simulation or the forward model operation as an argument, and the specification accepted Learning lookup table based on what is typed in And characterized in that the design Ri.

According to the present invention, it becomes easy to create / change a calculation simulation or a forward model and input / change an evaluation formula.

A computer-readable recording medium according to the present invention (Claim 12) displays a configuration diagram of a system to be controlled / adjusted by using a graphical user interface, and inputs it by the graphical user interface. Information relating to the sensor configuration and information defining the relationship between the observation value obtained by the specified sensor configuration and the input vector to the lookup table, and designing the lookup table based on the input information. It is a computer-readable recording medium that stores a program for controlling a computer.

The computer-readable recording medium according to the present invention displays a configuration diagram of a system to be controlled / adjusted by using a graphical user interface, and a sensor configuration input by the graphical user interface. Related information and information defining the relationship between the observation value obtained by the specified sensor configuration and the input vector to the look-up table is accepted, and the look-up table is designed based on the inputted information. It is a computer-readable recording medium that stores a program for controlling a computer to evaluate and present the overall performance of a look-up table.

A computer-readable recording medium according to the present invention (Claim 13) is characterized in that the look-up table is designed by performing a learning on robustness with respect to a predetermined internal parameter of a system to which the look-up table to be designed is applied. Is a program for designing
Using the graphical user interface, display the configuration diagram of the system to be controlled / adjusted, and display the internal parameter information of the system expressed in the form of the ambiguous probability distribution input by the graphical user interface. When the information is input and accepted, learning is performed by adding a random variation according to a given probability distribution with respect to an internal parameter given in an ambiguous form, and based on this learning result, a lookup table It is a computer-readable recording medium that stores a program for controlling a computer so as to determine contents.

In each invention relating to the computer-readable recording medium described above, as a procedure or a function of designing a look-up table to be executed by a computer,
It is possible to use the procedure or function corresponding to the invention related to each method or the invention related to each device described above.

Each of the inventions related to the above apparatus can be realized as an invention related to a method, and each invention related to the above method can be realized as an invention related to an apparatus.

The invention relating to the above apparatus and method can also be realized as a machine-readable medium in which a program for causing a computer to execute a corresponding procedure or means is recorded.

[0092]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

The look-up table designing apparatus according to the present embodiment is roughly controlled automatically by performing off-line learning on a calculation simulation model or a forward direction model simulating the operation of the controlled system. Designs (or supports) the contents of the look-up table, and follows the principle that is the basis of the "mechanism that produces a wide variety of antibodies against a huge variety of antigens" of the biological immune system. Function to generate a wide variety of input and output vector coding units (antibodies) that are the basis of the contents of the lookup table (diversity generation function), and to provide better control while maintaining diversity. When the coding unit that is expected to have the ability to perform the selection function (diversity maintenance function) is executed in parallel, In addition, by devising to maintain diversity even in the selection of antibodies performed by learning and to maintain excellent coding units due to robustness, it is possible to save time without prior know-how. This is to enable the design of a lookup table that is appropriate and has excellent robustness.

In this embodiment, the genetic algorithm (G
Following the fitting of the information coding unit to the “gene” in A), the coding unit of the input vector and the output vector to the lookup table is fitted to the “antibody”. That is, "antibody" means a coding unit of an input vector and an output vector. "Antibody" is expressed as a numerical value or symbol vector.

In the present embodiment, the "concentration of antibody"
The index of efficacy is also applied to, considering the superiority or inferiority of the antibody with other antibodies. That is, the "concentration of antibody" means
It has the same meaning as the reinforcement value in general methods based on reinforcement learning, and serves as a criterion for selecting antibodies. Reinforcement learning is a type of machine learning that adapts to the environment by updating the weights to prioritize outputs that give rewards as inputs, and learning is performed even if a clear teacher output is not given. It has the feature that

The “evaluation value” of an antibody described below is obtained as a result of performing a calculation simulation or a forward model calculation for the antibody under certain conditions. Whether the target system could be successfully controlled by the antibody under those conditions It is a scale indicating whether or not it is a value determined for the antibody alone.

Various systems such as a traffic system, a production system, a physical distribution system, a chemical plant, a biological plant, and an elevator group management system can be considered as the target system to be controlled by the lookup table.

(First Embodiment) A first embodiment of the present invention will be described.

FIG. 1 shows the configuration of the lookup table designing apparatus according to this embodiment. The present look-up table designing device can be realized as hardware, and can also be realized as a program and data including the look-up table itself.

The look-up table designing device comprises an antibody storage unit 5, a candidate antibody production unit 4, an antibody evaluation value determination unit 3, a parameter input unit 2, a learning / design management unit 1, and a look-up table creation unit 6. There is.

The antibody storage unit 5 is a coding unit of an input vector (output value of the sensor or a processed value thereof) of the lookup table and an output vector (control parameter for the target system or data from which it is obtained). A function that associates and stores the antibody and its concentration, rewrites the concentration of the specified antibody with the updated value, deletes the data of the specified antibody and its concentration, and displays the information of the generated antibody. It has a function of storing and a function of searching and extracting an antibody (neighboring antibody) included in a predetermined range of the input vector space. In the search / extraction of the neighboring antibody, for example, when the input vector space is divided into meshes, the mesh to which the designated input vector belongs and some meshes around the mesh are searched and extracted as a predetermined range.

The candidate antibody production section 4 produces a new candidate antibody to be additionally stored in the antibody storage section 5 as needed. For example, when the number of candidate antibodies is equal to or less than the specified number, a predetermined number is generated.

The antibody evaluation value determination unit 3 uses a simulator that models the target system S to which the look-up table is applied based on the coding content of the designated antibody, or a simulation calculation by a forward model or a forward model calculation. By executing the above, the evaluation value of the designated antibody is obtained.

The parameter input section 2 inputs a parameter used for learning or a look-up table design from a user (or another process or device).

The learning / design management unit 1 manages the learning process based on the input learning parameters, and also manages the lookup table designing process based on the input lookup table design parameters.

In the learning process, the neighborhood antibody is extracted by the antibody storage unit 5, the neighborhood antibody is added by the candidate antibody production unit 4 as needed, the neighborhood antibody is evaluated by the antibody evaluation value determination unit 3, and all the neighborhood antibodies are evaluated. A series of processes consisting of updating the concentration of each neighboring antibody in consideration of the value and selecting the antibody to be deleted based on the concentration is repeatedly performed.

In the learning process, the basic parameters for obtaining the neighboring antibody and the internal parameters of the simulator or forward model are changed for each cycle of the series of processes. In the embodiment described later, this change is performed by the antibody evaluation value determination unit 3, and the input vector that is the basis for determining the neighboring antibody is generated based on the above internal parameters.

The lookup table creation unit 6 determines the contents of the lookup table based on the contents stored in the antibody storage unit 5 at that point in the lookup table designing process after the learning process is completed, and the lookup table is created. Store it in the table. For example, as the content of each mesh in the input vector space of the lookup table, an antibody selected based on the concentration (for example, the maximum concentration of all antibodies included in a range centered on the mesh) The content of the lookup table is determined by adopting the coding content of (the antibody to have).

According to the present embodiment, the lookup table can be automatically designed by repeating the trial and error autonomously on the calculation simulation without the designer having to prepare the training teacher data unlike the neural network. Becomes

Some more detailed embodiments will be described below.

(Second Embodiment) A second embodiment of the present invention will be described. The lookup table designing apparatus of the present embodiment has two processing steps as a processing step, a learning step performed first and a lookup table designing step performed next, and the lookup table is processed through these two processing steps. Created. In the learning process, one learning cycle is used as a unit, and the learning cycle is repeated until a predetermined ending condition is satisfied. In one learning cycle, one input vector for the look-up table is selected, and learning is performed on the antibodies in the vicinity of the input vector under the condition of the internal parameters that are changed in each learning cycle (searching / extracting neighboring antibodies, Generation of a new antibody, evaluation of neighboring antibodies, updating of the concentration of neighboring antibodies, selection of antibodies based on the concentration) is performed as needed.

FIG. 2 shows the configuration of the look-up table designing apparatus of this embodiment. FIG. 3 shows a flow of processing at the time of antibody learning according to this embodiment. FIG. 6 shows the flow of processing when creating the lookup table after learning according to this embodiment. The present look-up table designing device can be realized as hardware, and can also be realized as a program and data including the look-up table itself.

As shown in FIG. 2, the look-up table design device includes a learning / design management unit 1, a parameter input unit 2, an antibody evaluation value determination unit 3, a candidate antibody production unit 4, an antibody storage unit 5, a look-up unit. An up table creating section 6 and a look up table storage section 7 are provided.

The antibody storage unit 5 stores an antibody database 52 in which the antibody, which is a coding unit of the input vector and the output vector for the lookup table, and its concentration are stored in association with each other, and the neighboring antibody (a certain input vector is used as a search key). To the neighboring antibody search unit 51 for extracting from the antibody database 52 antibodies having an input vector within a predetermined range), the concentration of the designated antibody is updated, and the designated antibody is deleted and newly generated. Store the antibody and its initial concentration in the antibody database 52
It has a data changing unit 53 for performing.

Candidate antibody producing unit 4 determines, in a certain learning cycle, when the number of antibodies in the neighboring antibodies is not more than the specified number,
A new candidate antibody to be additionally stored in the antibody storage unit 5, that is, data including an input vector coding unit and an output vector coding unit is generated. In the present embodiment, since the input vector that is the basis of the search / extraction of the neighboring antibody is used for the input vector part of the new antibody,
The candidate antibody production unit 4 actually generates an output vector and performs processing for combining this with the input vector.

The parameter input unit 2 inputs learning parameters and lookup table design parameters from the user. Instead of being input by the user, the learning parameter and the look-up table design parameter may be received from another execution process or device via the parameter input unit 2.

The antibody evaluation value determination unit 3 has a parameter automatic change unit 31, a simulation execution unit 32, and a simulation result evaluation unit 33. Parameter automatic change unit 3
1 is the simulation execution unit 3 for each learning cycle.
The internal parameters used in 2 will be changed. The simulation execution unit 32 performs a first simulation process for each observation cycle, using the internal parameter given by the parameter automatic change unit 31, to obtain the observed value r that is the source of the input vector. Further, the simulation executing unit 32 simulates the operation of the controlled object by using the internal parameter given by the automatic parameter changing unit 31 and the control parameter obtained from the output vector of the designated antibody for each learning cycle. The simulation process of is performed. The second simulation process is performed for each designated antibody. The simulation result evaluation unit 33 obtains the evaluation value of the antibody from the result of the second simulation processing (for example, a predetermined parameter indicating the operation result of the controlled object).

In the first simulation process performed by the simulation executing unit 32, a calculation simulator or a forward model in which the target system S to which the lookup table is applied is modeled,
For the purpose of simulating the observed values from the internal parameters, we do not rely on the output of the lookup table for.

In the second simulation processing performed by the simulation executing unit 32, a simulation simulator or a forward model in which the target system S to which the look-up table is applied is modeled, and the target system S is calculated from internal parameters and control parameters. Is used for the purpose of obtaining a predetermined behavior or a predetermined characteristic value of. For example, when designing a look-up table for controlling traffic signals, the above-mentioned TRANSSYT may be used as a calculation simulator.

The learning / design management unit 1 manages the learning process and the look-up table design process based on the input learning parameter and the look-up table design parameter. The input conversion unit 11, the output conversion unit 12,
It has a learning management unit 13 and a design management unit 14.

The input conversion unit 11 converts the observation value data obtained by the first simulation processing of the simulation execution unit 32 into an input vector.

The output conversion unit 12 converts the output vector of the designated antibody into a format suitable for the second simulation process of the simulation execution unit 32.

The learning management unit 13 manages the learning process, and the antibody buffer 131 that stores the coding content and concentration of the antibody to be evaluated or the antibody storage address information indicating the storage position in the antibody database 52, An evaluation value buffer 132 that stores the evaluation values of the evaluated antibodies, a concentration update calculation unit 133 that updates the respective concentrations of all the evaluated antibodies, based on the concentrations It has an antibody deleting unit 134 which deletes the antibody according to a predetermined standard.

The design management unit 14 manages the look-up table design process, and controls the look-up table creation unit 6 such as giving an opportunity to start operation.

The look-up table creating unit 6 determines the contents of the look-up table on the basis of the contents stored in the antibody storage unit 5 and stores the contents in the look-up table storage unit 7. The input scanning unit 61 that controls the up-table to be scanned in mesh units, the antibody buffer 62 that stores the neighboring antibodies to the target mesh, and the maximum that detects the antibody with the highest concentration from the neighboring antibodies stored in the antibody buffer 62 A lookup table for writing the portion corresponding to the output vector of the antibody coding content detected by the concentration antibody detecting unit 63 and the maximum concentration antibody detecting unit 63 or both the input vector and the output vector into the corresponding portion of the lookup table 7. The writing unit 64 is included.

Next, the learning process in this embodiment, that is, the processing procedure during antibody learning will be described with reference to the flowchart in FIG.

(Step S11) In step S11,
For example, GUI (graphical user int
parameter input unit 2 using
Maximum and minimum values of each dimensional variable of the input vector to the lookup table, the number of mesh divisions for each dimension, the size (number) of neighboring meshes for each dimension, the specified number of neighboring antibodies, and the newly generated antibody The learning parameters such as the initial antibody concentration and the look-up table design parameters to be given to

The size (number) of the neighboring mesh is a prescribed value of the number of meshes which shows up to which adjacent range as seen from a certain mesh. For example, when the size is 1 for a certain dimension, that mesh and the meshes on both sides become neighboring meshes in that dimension. Further, for example, when the input vector is two-dimensional, size = 1 for each dimension.
In the case of, there are nine neighboring meshes.

The neighborhood antibody is an antibody included in a certain mesh and the range of the neighborhood mesh. For more details,
It is an antibody having an input vector included within the range of a neighboring mesh centered on the mesh to which the input vector defined in step S13 belongs.

The maximum value, the minimum value, the number of mesh divisions, and the size of the neighborhood mesh are stored and used in the neighborhood antibody search unit 51. The maximum value, the minimum value, the number of mesh divisions, and the size of the neighboring mesh are also used in the look-up table creation unit 6 as a look-up table design parameter.

The specified number of antibodies in the vicinity is determined in step S15.
In step S16, it is used to determine whether or not to start the process of generating a new candidate antibody.

The initial antibody concentration is an initial concentration stored when a new candidate antibody is generated and added to the antibody database 52 and the antibody buffer 131 in step S16, and a fixed value is set.

(Step S12) In step S12,
In the automatic parameter changing unit 31, the antibody evaluation determining unit 3
The processing is performed such that the internal parameters of the calculation simulator or the forward model that models the target system S used in 1. are changed for each learning cycle so as to cover the ranges in which they can occur (or are assumed to occur).

As internal parameters, in addition to the internal parameters of the target system S that can be estimated from the observed values (for example, the downlink traffic volume in the example of FIG. 17), the target system that cannot be known from the previously set observed values. The one corresponding to the internal parameter p of S (p is a vector) (see, for example, FIG.
Inflow traffic volume or its ratio to the total traffic volume in 7) is included. In the present embodiment, in order to satisfy the following two purposes, learning is performed while changing the internal parameter p many times so as to cover a range that can occur (or is assumed to occur). (Purpose 1) Almost all meshes where r'can appear in the input vector space by indirectly varying the input vector r '| p obtained by converting the observable value uniquely found for the internal parameter p. The antibody is comprehensively generated and is improved by learning. (Purpose 2) Input vector r '| p for each mesh
Since the antibody evaluation is performed on some samples of the internal parameter p such that can be included, it is possible to solve the problems of the above-described conventional method and obtain robustness.

As a method of changing the internal parameter, a completely random method, a method of stochastically varying according to the probability that the internal parameter actually changes in the target system S, an internal parameter vector space is equally divided, and a median value is set. Various methods are conceivable, such as a method of comprehensively scanning.

It should be noted that the selection input of the internal parameter changing method is accepted from the parameter input unit 2 and the internal parameter is changed by the method according to this input so that the user can specify the changing method of the internal parameter. Good.

(Step S13) In step S13,
First, under the condition of the internal parameter set in step S12, the simulation execution unit 32 performs a simulation calculation or a forward model calculation as the first simulation process, and the observable value r | Find p. Then, the obtained observable value r | p is converted into the input vector r ′ | p by the input conversion unit 11. The observable value r and the input vector r ′ have a one-to-one relationship.

If the observable value r and the input vector r'are the same, this conversion procedure is unnecessary.

(Step S14) In step S14,
The input vector r'determined in step S13 is input to the neighboring antibody search unit 51 as a search key, and the neighboring antibody search unit 51 inputs the input vector r input as the search key.
For the mesh containing ', the coding contents and concentrations of all the neighboring antibodies contained in the neighboring mesh within the designated neighborhood for each dimension, or the antibody database 5
The antibody storage address information indicating the storage position in 2 is searched and extracted.

For example, as shown in FIG. 4, the input vector is two-dimensional, the number of mesh divisions is eight in both dimensions, and the size of the neighborhood is one mesh in both dimensions. In FIG.
If the “black circle” is the coordinate occupied by the input vector equivalent portion of the antibody stored in the antibody database 52 in the input vector space, and the “x” is the coordinate occupied by the input vector input as the search key, the nine hatched The four antibodies (four black circles) included in the neighborhood mesh are extracted as neighborhood antibodies by the neighborhood antibody search unit 51.

The extracted coding content and concentration of the nearby antibody, or the antibody storage address information is temporarily stored in the antibody buffer 131.

(Steps S15 and S16) Step S1
In S5 and S16, a process of generating a new candidate antibody is performed as needed.

That is, as a result of the search / extraction in step S14, when the number of the extracted neighboring antibodies is less than the predetermined number specified in advance (step S15), the candidate antibody producing unit 4 causes a new candidate. An antibody is generated (step S15).

As a method for generating a new candidate antibody, for example, an output vector is created by a predetermined procedure, and both the input vector defined in step S13 and the created output vector are combined to form an antibody. A method of generating by coding is conceivable.

As a method of creating the output vector, for example, a method of finding an optimum output vector x | p under the setting of the internal parameter p by a search method such as a hill climbing method can be considered. In addition to the hill climbing method, there may be a means for creating an output vector by means of random generation, mutation, crossover or the like using a random generation or another antibody in the vicinity as a template, as in a genetic algorithm. Besides, various methods are conceivable.

FIG. 5 shows an example of antibody production. here,
Sensor input x1 = 0.5, x2 = 3.0, x3 = 1.0
Is the input vector, and the template of the output vector (1.0,
0.6) is finely adjusted to y1 = 1.2, y2 = 0.5
As the output vector and combine them (0.5, 3.
0, 1.0, 1.2, 0.5) is coded as a new antibody.

The newly generated candidate antibody is additionally stored in the antibody database 52 together with the initial antibody concentration through the data changing unit 53. In addition, like the previously extracted neighboring antibody, the coding content and concentration of the generated candidate antibody, or the antibody storage address information is the antibody buffer 1
31 (that is, the generated candidate antibody is added to the neighboring antibody to be evaluated).

[0148] When the number of neighboring antibodies extracted in this way falls below a prespecified specified number, a new candidate antibody is generated, so that the number of antibodies existing in the vicinity of each mesh is biased. Can be prevented. If the number of neighboring antibodies is too large, it takes calculation time to evaluate the antibody at the time of updating the concentration, and if it is too small, the relative evaluation of antibodies in the neighborhood described later cannot be performed. Can always be maintained at an appropriate antibody number of 2 or more and not more than the specified number.

Further, according to the present embodiment, since only the antibodies near the input vector that can appear when the internal parameter p is changed are not generated from the beginning, it is possible to avoid learning by creating / adding unused candidate antibodies. It is possible to improve efficiency and save memory.

(Steps S17 to S19) Step S1
In 7 to S19, relative evaluation of neighboring antibodies is performed. here,
In the antibody buffer 131, information about the neighboring antibody associated with the input vector determined in step S13, that is, information regarding each neighboring antibody extracted in step S14,
When a new antibody is generated in step S16, information about the antibody is temporarily stored.

First, in step S 17, one of the neighboring antibodies temporarily stored in the antibody buffer 131 is transferred from the antibody buffer 131 (the antibody buffer 13
1 stores the coding content and concentration) or from the antibody database 52 (antibody buffer 1
31 when the antibody storage address information is stored),
A part corresponding to the output vector of the coding content of the antibody is extracted, and the output conversion unit 12 converts the extracted part corresponding to the output vector into a format suitable for the second simulation processing by the simulation executing unit 32. To do.

Next, in step S18, the simulation executing section 32 inputs the output data of the output converting section 12 as a control parameter, and uses this and the previously determined internal parameter to perform the second simulation process ( A second calculation simulation or forward model calculation) is performed. The simulation result evaluation unit 33 obtains the evaluation value of the antibody to be evaluated based on the data obtained as a result of this second simulation process. The obtained evaluation value of the antibody is temporarily stored in the evaluation value buffer 132.

Steps S17 and S1 described above
The process consisting of 8 is repeated for each neighboring antibody temporarily stored in the antibody buffer 131 (step S1).
9) Obtain the evaluation value of all the neighboring antibodies targeted in the learning cycle.

In the above first simulation processing, the observable value vector r | p of the target system S of the lookup table corresponding to the internal parameter p is obtained, and the input vector r '| p is obtained from this. In the simulation process of 2, the input vector r ′ | p (one-to-one correspondence with the observable value vector r) under the setting of the internal parameter p
The estimated value F | p, x when the portion x corresponding to the output vector of the neighboring antibody of is input to a simulator or the like is estimated. In the present embodiment, this F | p, x is used as the antibody evaluation standard, and p
By performing learning while changing the value of p, the internal parameter p | r ′
Whatever value of ?, there is always an increase in the concentration of the coded antibody at the output x such that F has relatively good robustness. This makes it possible to solve the problem 2 of the conventional method described above without the trouble of creating and analyzing the backward model of the target system S.

If the output vector is in a format that is suitable for the second simulation processing, the above conversion is unnecessary.

(Step S20) In step S20,
The concentration updating calculation unit 133 performs a predetermined concentration updating calculation using the evaluation values of all the neighboring antibodies for each neighboring antibody, and updates the concentration of the corresponding neighboring antibody stored in the antibody database 52 through the data changing unit 53. I do.

In the concentration update calculation, for example, the average value of the evaluation values of all the neighboring antibodies obtained by the antibody evaluation value determination unit 3 is obtained, and the evaluation value of the antibody for which the concentration is updated and the magnitude relationship between the average values are obtained. Accordingly, the antibody concentration is updated by increasing or decreasing a fixed amount from the antibody concentration.

For example, there is a method of updating the density based on the following formula. Here, Fj is the evaluation value of antibody j. In the following formula, the larger the value, the better the evaluation.
n is the number of antibodies in the vicinity, Cj is the concentration of antibody j, and Δ is the concentration increase / decrease value. ΣFk / n ≦ Fj → Cj = Cj + Δ ΣFk / n> Fj → Cj = Cj−Δ Here, the range of the total sum of Fk is K = 1 to n.

Alternatively, based on the following formula, there is a method of changing the concentration by regarding the concentration as the winning rate of each antibody. Here, Sj is the number of steps that have passed since the antibody was added to the antibody database 52, and is stored in the antibody storage unit 5. ΣFk / n ≦ Fj → Cj = (SjCj + 1) / (S
j + 1) ΣFk / n> Fj → Cj = SjCj / (Sj + 1) Here, the range of the sum of Fk is K = 1 to n.

As other density updating methods, various methods such as a method of using a weighted average according to the Euclidean distance between the antibody in the input vector space and the input, a method of increasing or decreasing the value according to the rank in the neighboring antibody from the density, etc. The method of can be considered.

According to such a concentration updating method, the concentration of an antibody having an excellent evaluation value relative to the market in the vicinity increases even if the market of the antibody evaluation value varies greatly due to the mesh. Since it is not necessary to set the absolute standard of the evaluation value for each mesh, the setting of the learning parameter becomes extremely easy. For example, the evaluation index P. I. When is used as an evaluation value, the reference value of good or bad changes remarkably according to the total amount of traffic on the mesh, and the standard of density increase / decrease also differs depending on the mesh. There is no need to set it separately.

Further, in this embodiment, since the range in which the antibodies compete with each other is limited to the neighborhood mesh of the input vector space as described above, the diversity in the input vector space is maintained and each mesh (look Antibodies corresponding to each address in the up table) will be obtained evenly. That is, r | that can be observed by the fluctuation of the internal parameter p in the calculation simulator or the forward model.
Antibodies corresponding to any of p are uniformly obtained.

(Step S21) In step S21,
In the antibody deleting unit 134, the data changing unit 5 extracts the antibodies whose updated concentrations are below the minimum threshold value from all the neighboring antibodies and deletes them from the antibody database 52.
Give instructions to 3. The data changing unit 53 that receives this instruction
Deletes the information on the designated antibody and its concentration from the antibody database 52. As a result, the antibody database 5
It is possible to prevent the maintenance of incompatible antibodies within.

There are various possible ways to set the initial antibody concentration and the minimum threshold value. As an example, when the initial antibody concentration is set to zero, the minimum threshold value is set to zero. Can be considered. The minimum threshold value may also be set from the parameter input unit 2.

(Step S22) In step S22,
Judge whether the preset learning end conditions have been achieved,
When it is determined that the preset learning end condition is achieved, the learning is ended at that point. If it is determined that the preset learning end condition has not been achieved, the process returns to step S11 to perform the next learning cycle.

In determining the end, various conditions such as the number of iterations, the coverage of internal parameters, and the value of the maximum antibody concentration in each mesh of the input vector space are comprehensively determined to determine whether to continue learning. It is preferable to judge.

By the way, although the size of the neighborhood used in the neighborhood antibody search unit 51 (that is, the size of the neighborhood mesh) is constant in the above, instead, the size of the neighborhood is gradually reduced according to the progress of learning. You may set so that. In this case, the neighborhood mesh size is changed in step S22.

Thus, when the user cannot determine an appropriate size of the neighborhood, the neighborhood is made large while the number of the antibodies is small, thereby increasing the chances of changing the concentration of the antibody of each mesh and obtaining a robust antibody. Is prioritized, and the smaller the neighborhood as learning progresses and the number of antibodies increases,
Priority is given to obtaining antibodies specialized in a narrow region of the input vector space, and the neighborhood size can be set according to the progress of learning.

As described above, the antibody learning is completed,
The data that is the basis for creating the lookup table is generated. At this point, the input vector space of the look-up table is not yet divided.

Next, the lookup table designing process in this embodiment, that is, the processing procedure for creating the lookup table after learning will be described with reference to the flowchart of FIG.

In the look-up table designing process, based on the antibody obtained in the learning process, for each mesh, the output vector for the mesh is selected from the neighboring meshes. This results in a division of the look-up table input vector space.

In the look-up table design process, the design management unit 14 gives a start instruction to the input scan unit 61 in response to the end of the learning process (or in response to an instruction given from the outside after the learning process is completed). Started at.

First, until it is determined in step S38 that the scanning of all meshes is completed, step S3
A series of processes from 1 to S37 is repeated. 1 of this processing
The content of one mesh of the look-up table is determined by repeating the operation once.

That is, first, in step S31, the input scanning unit 61 sets the target to the first mesh on the input vector space at the start of scanning, and thereafter moves the target to the next mesh. In this step S31, the target mesh is sequentially moved for each repetition of the above series of processing so that all the meshes in the input vector space are scanned without exception.

Next, in step S32, the median value of the target mesh is used as a search key in the neighborhood antibody search unit 51 to extract the neighborhood antibody. The extracted coding content and concentration of the nearby antibody or antibody storage address information is temporarily stored in the antibody buffer 62.

It should be noted that the size of the neighborhood mesh used for each dimension used here may be the same value as in the learning process,
You may use the value of the size of the neighborhood mesh input for the look-up table design process in step S11.

For example, the input vector is two-dimensional, the size of the neighborhood is one mesh in both dimensions, and the antibody database 5
It is assumed that the portion (12 adjacent meshes) obtained by cutting out a part of the input vector space in the contents of 2 is in the state as shown in FIG. 7. Here, the number in parentheses in FIG. 7 indicates the concentration of the antibody. In this case, m1
If the target mesh is 9
Antibodies 1 to 10 contained in one mesh are searched,
When the mesh of m2 is the target, the antibodies 5 to 14 included in the nine meshes centering on m2 are searched.

Next, in step S33, it is determined whether or not the neighboring antibody has been extracted.

If no neighboring antibody is extracted, in step S34, the corresponding mesh position is saved for later notification processing to the user, and in step S3.
Return to 1.

On the other hand, if the neighboring antibody has been extracted, the process proceeds to the next step S35.

In step S35, the antibody detection unit 63 detects the antibody having the highest concentration among the neighboring antibodies extracted in step S32 and temporarily stored in the antibody buffer 62.

For example, in the example of FIG. 7 described above, when the mesh of m1 is the target, antibody 3 having the highest concentration (30) is detected from antibodies 1 to 10, and m2 is detected.
When the mesh No. is targeted, the antibody 9 having the highest concentration (29) among the antibodies 5 to 14 is detected.

Next, step S36 and step S3.
7, the lookup table writing unit 6 writes the lookup table in the second format described below and the lookup table in the first format described below.
4. Of course, step S36 and step S
It is also possible to execute only one of the formats 37 to create only one type of lookup table. Further, the user may be allowed to select the format of the lookup table to be created, and the lookup table of one or more formats specified by the user may be created.

When writing the lookup table of the first format, the median value of the mesh in the input vector space is input to the neighboring antibody searching unit 51 as a search key, and the maximum concentration of the neighboring antibodies extracted is set. The portion corresponding to the output vector of the coding content of the antibody possessed is written into the address corresponding to the mesh in the lookup table. By sequentially performing this writing for all meshes, the antibody stored in the antibody storage unit 5 can be converted into a lookup table.

As a result, the best output vector of the antibody for each mesh is stored in the lookup table,
The learning result can be converted into a lookup table. Also,
The resulting look-up table of the first form can be implemented as an associative memory.

When writing the lookup table of the second format, the median value of the mesh in the input vector space is input to the neighboring antibody searching unit 51 as a search key, and the maximum concentration among the neighboring antibodies extracted is set. Write the coding content of the antibody you have in the lookup table. By performing this writing sequentially for all meshes, the antibodies stored in the antibody storage unit 5 can be converted into a lookup table.

The obtained look-up table of the second form cannot be implemented as an associative memory, and requires a look-up table reading means having the same function as the neighborhood antibody search unit 51. In general, the amount of storage is smaller than that of the lookup table of the first type, and it can be created at the same time when the lookup table of the first type is created.

FIG. 8 is a look-up table for determining traffic signal parameters (a look-up table of the second form).
An example is shown.

As described above, the look-up table designing apparatus according to the present embodiment has a process of determining the contents of the look-up table based on various excellent coding units acquired as a result of learning, and thus the look-up table It is not necessary to have the know-how to divide the up table area.

When the look-up table design is completed as described above, in step S39, the neighborhood antibody extracted when the median value of the target mesh in the input vector space is input as the retrieval key to the neighborhood antibody retrieval unit 51. The user is notified of the information indicating the position of the mesh in which there was no.

As a means for this notification, a method of visually displaying the coordinates of the mesh median value in the input vector space on the GUI can be considered.

There may be a method of notifying the user of the mesh side where the antibody exists.

When a new candidate antibody is created by the candidate antibody producing unit 4 according to the above-mentioned procedure, the mesh where no antibody exists in the vicinity remains after learning. These are either a mesh in which the input vector cannot appear in the neighborhood no matter how the internal parameter p is changed, or the learning is insufficient. However, information for allowing the user to distinguish the former from the latter is provided. Can be provided.

Next, another example of the processing procedure at the time of antibody learning in this embodiment will be described.

FIG. 9 shows the flow of another process during antibody learning according to this embodiment.

This procedure is basically the same as the procedure of FIG. 3, but the processing of steps S115 and S116 and S1.
The process of 21 is a part different from the procedure of FIG. Only the different parts will be described below.

In the processing of steps S15 and S16 of FIG. 3, a new candidate antibody is generated when the number of the extracted near-field antibodies is less than the predetermined specified number. However, in the processing of steps S115 and S116. , If the number of extracted antibodies is less than the specified number specified in advance,
New antibodies are repeatedly generated until the specified number is reached.

By doing so, step S1 in FIG.
5, the number of antibodies existing in the vicinity of each mesh is averaged more than in the processing of S16.

Further, in the processing of step S21 of FIG.
Although the antibody whose updated concentration is below the minimum threshold is deleted from the antibody database, in the processing of S121, the antibody with the minimum concentration among the extracted neighboring antibodies is deleted from the antibody database.

..

The configuration (FIG. 2) of this embodiment described above can be modified and implemented as appropriate. For example, the parameter automatic change unit 31 in the antibody evaluation search unit 3
The simulation result evaluation unit 33 may be an internal configuration of the learning / design management unit 1. Further, for example, the input conversion unit 11 and the output conversion unit 1 in the learning / design management unit 11
2 and the antibody buffer 14 have an internal configuration of the antibody evaluation search unit 3, and the evaluation value buffer 132 and the concentration update calculation unit 133 are included.
The antibody deleting unit 134 may be an internal configuration of the antibody evaluation value determining unit 3 or the antibody storing unit 5. In particular, when the present embodiment is implemented by a program and data, it is not always necessary to faithfully follow the form as shown in FIG. 2 to form a hierarchical structure of modules and programming, and various structures that realize functions equivalent to those in FIG. There is a degree of freedom to build a program with.

As described above, according to the present embodiment, in the learning process, a variety of input vector and output vector coding units (as the basis of the contents of the look-up table by the candidate antibody production unit 4 (as necessary) (Antibody) is generated and stored in the antibody database 52, and the antibody evaluation value determination unit 3 is first locally in the input vector space, that is, in the vicinity of a certain input vector.
Each neighboring antibody is evaluated independently by using the
34, the concentration of each neighboring antibody is updated on the basis of the relative excellence between the neighboring antibodies, and the data changing unit 53 updates the unnecessary antibodies, and the data changing unit 53 deletes unnecessary antibodies under the instruction of the antibody deleting unit 134. Therefore, the concentration of the antibody expected to have the ability to perform better control while maintaining the diversity is increased, and stored in the antibody database 52 in the lookup table designing process after the learning process. The content of the lookup table is determined by extracting the antibody with the maximum concentration in the neighboring meshes for each mesh based on the antibody concentration that is stored. A suitable and robust look-up table can be automatically designed.

(Third Embodiment) Next, a third embodiment of the present invention will be described.

FIG. 10 shows the configuration of the look-up table designing apparatus according to this embodiment.

This embodiment is basically the same as the second embodiment, except that n (n is 2 or more) antibody evaluation value determination units 3 _{1 to} 3 _n are provided. .

Only the points of difference between this embodiment and the second embodiment will be described below.

The antibody evaluation value determination unit 3 _{1 to} 3 _{n of} this embodiment
Are respectively the parameter automatic changing unit 31, the first simulation executing unit 321, and the m second units (m is 2 or more).
A simulation execution unit 322 ₁ ~322 _m, m pieces of the simulation result evaluation unit 33 ₁ ~ 33 _m.
The internal configurations of the antibody evaluation value determination units 3 _{1 to} 3 _n are all the same.

The antibody evaluation value determination unit 3 _{1 to} 3 _{n of} this embodiment
The basic function of is similar to that of the antibody evaluation value determination unit of the second embodiment. In addition, the parameter automatic changing unit 3 of the present embodiment
1, the first simulation executing unit 321, a second simulation executing unit 322 ₁ ~322 _m, the simulation result evaluation unit 33 ₁ ~ 33 _m is basically automatic parameter changing section 31 of the first embodiment, The function of the first simulation process of the simulation executing unit 32, the function of the second simulation process of the simulation executing unit 32, and the simulation result evaluating unit 3
Same as 3.

When the present embodiment is implemented by a program, it is not always necessary to prepare in advance n parts corresponding to the antibody evaluation value determining units 3 _{1 to} 3 _n in advance. It may be generated as a process. This point is that the second simulation execution units 322 _{1 to} 32 ₁ of m in each of the antibody evaluation value determination units 3 _{1 to} 3 _n .
2 _m and m-number of simulation result evaluating section 33 _to 333
_The same applies to _m .

The antibody evaluation value determination unit 3 _{1 to} 3 _{n of} this embodiment
In, and to allow parallel processing of m pieces second simulation executing unit 322 ₁ ~322 _m and simulations evaluating the m set of neighboring antibodies of result evaluation unit 33 ₁ ~ 33 _m. As a result, the processing loop (steps S17 to S19) inside the procedure of FIG. 3 can be sped up.

In addition, the automatic parameter changing section 31 of each of the antibody evaluation value determining sections 3 _{1 to} 3 _n cooperates with each other to share different internal parameters.
Process loop outside the procedure (steps S12 to S22)
Are executed simultaneously for n loops.

That is, the present embodiment is configured to speed up each of the double loops in the procedure of FIG.

Further, in the present embodiment, since the n antibody evaluation value determination units 3 _{1 to} 3 _n operate in parallel, the outputs from the antibody evaluation value determination units 3 _{1 to} 3 _n are also performed in parallel. Therefore, it is desirable to perform control such as exclusively giving the right to operate the data change unit 53 to each of the antibody evaluation value determination units 3 _{1 to} 3 _n , and control for maintaining the consistency of the content of the antibody database 52. Also, for example, if the antibody whose concentration was to be updated according to the result of one antibody evaluation value determination unit has already been deleted according to the result of another antibody evaluation value determination unit, cancel the concentration update instruction for that antibody. You may perform various controls.

As a result, the simulation operation or the forward model operation which requires the most calculation time can be processed in parallel, and the learning time can be shortened.

The first calculation simulation and the second calculation simulation
It is possible that the calculation simulation of 1 is executed by the same type of software or hardware module. In this case, one of the second simulation executing unit 322 ₁ ~322 _m in FIG. 10, used when performing the first computation simulation operations.

(Fourth Embodiment) Next, a fourth embodiment will be described.

Since this embodiment has a basic structure similar to that of the second and third embodiments, only different parts will be described below.

In the second and third embodiments, the sensor specifications (the number of sensors, the sensor type, the arrangement of the sensors, etc.) on which the first simulation process is based and the input converter 11 are used.
In, the conversion formula or conversion rule for converting the observed value obtained as a result of the first simulation processing into the input vector was fixed and set in advance.

In this embodiment, the user can input the sensor specifications and the conversion formula or the conversion rule through the GUI of the parameter input unit 2, or the observable sensor information and some of the sensor information are integrated. By making it possible to selectively input information to be used as an input to the lookup table from among the information obtained by doing so, the lookup table designing apparatus according to the present embodiment creates the lookup table based on the user's input. It is something that is done.

In this embodiment, the user has designated the calculation simulator or the forward model of the first simulation processing in the second and third embodiments under the internal parameters set in step S11. In addition to adding a function of outputting an observation value observed in each sensor, a function of using a conversion formula or conversion rule designated by the user is added to the input conversion unit 11.

Also, since the number of dimensions of the input vector varies depending on this conversion formula or conversion rule, in step S11, the maximum and minimum values of the respective dimensional variables of the input vector to the lookup table, The number of mesh divisions for each dimension and the size (number) of neighboring meshes for each dimension are input so that necessary information is available.
Further, each functional block portion in FIG. 2 and FIG. 10 is configured to be able to cope with a change in the number of dimensions of the input vector.

It should be noted that the second in the second and third embodiments
Since the simulation process of No. does not depend on the sensor specifications (because the internal parameters that are the source of the first simulation process are used instead of the observed values and input vectors obtained by the first simulation process), no correction is required. Becomes

Also, the user is allowed to input internal parameters in the form of an ambiguous probability distribution using the GUI of the parameter input unit 2, and the internal parameters given in the ambiguous form are randomly distributed during learning according to the probability distribution. May be added to automatically design a robust lookup table according to the degree of ambiguity of internal parameters.

FIG. 11 shows an image diagram of a GUI screen when this embodiment is applied to a look-up table design for determining a traffic signal control parameter.

Here, on the first screen (200 in the figure), in addition to the traffic signal (204 in the figure) to be controlled, the road / lane / position (2 in the figure) where the vehicle detector (201 in the figure) is placed.
02), or where necessary, by dragging and dropping a sensor item (203 in the figure) prepared as a menu on the GUI, such as a place where a relatively expensive heterogeneous sensor such as an image processing device is placed. Can be freely selected and placed. In addition, the user can determine, on the GUI, for example, whether or not to use the average value of the plurality of vehicle detector information as an input, how to combine the vehicle speed detector and the designated different type sensor, for example, the system diagram being displayed. The desired sensor arranged above and its processing content (not shown) can be input by selecting with a mouse. Furthermore, if necessary, in the second screen (300 in the figure), in the fuzzy theory, links with large fluctuations in traffic volume and average vehicle speed and links for which specific values cannot be specified due to insufficient investigation You can enter simulation parameters as values with ambiguous widths such as membership functions.

Thereafter, the look-up table designing apparatus of this embodiment automatically generates a look-up table according to the designated condition.

Preferably, it is effective to add a function of evaluating and displaying the performance of this automatically generated look-up table to the look-up table designing apparatus of this embodiment.

As a performance evaluation method, a method of referring to the density obtained in the learning process, a lookup table created, and a calculation simulator or a forward direction model operated under various conditions are tested. Various methods are conceivable, such as a method of comprehensively evaluating the look-up table and a method of evaluating when the look-up table cannot be completed.

By doing so, the user takes improvement measures such as changing the sensor position and designating that a more detailed traffic volume survey is to be performed, with reference to the performance of the presented table. be able to.

Further, by comparing the control adjustment performances of the lookup tables automatically designed under the specified input, it is possible to support the design of the monitoring system according to the purpose.

[0231] For example, for a place where the traffic volume and the average vehicle speed are known to be almost constant by the traffic survey,
By fixing the internal parameters of the traffic simulator to specific numerical values, and especially in places with large fluctuations or in places where specific values cannot be specified due to insufficient research, ambiguous widths such as the membership function in fuzzy theory are used. By designating the internal parameter as the value to be held, it becomes easy to design the table according to the ambiguity and the degree of fluctuation of the target system S for control adjustment.

As described above, according to the present embodiment, it is not necessary to analyze the backward model of the target system S in accordance with the type of information used as an input to the lookup table, and it is possible to obtain an approximate difference regardless of the type of input vector. Since learning can be performed by the same method, it becomes possible for the user to select the input to the look-up table on the spot and immediately learn.

Next, the results of a simple simulation experiment performed by creating a prototype system of the look-up table designing apparatus to which the present invention is applied as a program will be described with comparison with the results of the prior art.

In this case, there are 15 intersections, 31
A simulation experiment was performed using a road model consisting of individual links. This road model has a traffic flow that flows in from the side with respect to the upstream traffic flow at the central intersection 8.
The signal controller measures each traffic flow going up and down by the vehicle detector provided at the intersection 1, determines the offsets at the other 14 intersections in the system by using the parameter determination table, and instructs all the intersections. . Here, "cycle" and "split" are fixed. In addition, each observation value was randomly given an observation error within 20%.

The prototype system acquires the coded antibody by learning the observed traffic flow rate as the condition part (input vector) and the offset to be set as the conclusion part (output vector). As the antibody evaluation function, the above-mentioned P.
I. Was used as α = 25. In addition, in order to speed up learning with an emphasis on practical use, TRANSYT
Utilized the optimization function of its own. The role of TRANSYT is to determine the evaluation value of the antibody by performing a traffic flow simulation corresponding to the second calculation simulation process, and to set certain internal parameters (three traffic volumes in FIG. 17) by the hill climbing method. To find the optimal "offset" below. From the optimum offset, the candidate antibody producing unit 4 produces a candidate antibody.

FIG. 12 shows the results of two experiments. After each learning, 100 different cases of randomly generated traffic flows were set, and P. When using the parameter determination lookup table set by the prototype system. I. Average value of P.P. when a lookup table designed by the conventional method is used. I. , And the former is P. I. It is the ratio of the cases that have won in all the cases. In the conventional method used here, the input vector space is divided into seven as shown in FIG. 16, and offset design is performed by optimization by setting a representative traffic flow rate in each division.

In the first example, learning is performed assuming that the traffic flow rate for each of the up and down directions randomly changes within a range of 90% from the reference value and the inflow traffic flow rate is considered to be zero. Further, in the second example, learning is performed assuming that the inflow traffic flow rate also randomly increases or decreases within a range of 90% from the reference value.

As shown in FIG. 12, in any of the examples, it is understood that the control by the look-up table designed by using the look-up table designing apparatus according to the present invention is superior in robustness. In particular, in the second example, it is impossible to estimate the ratio of the inflow traffic from the side in the upflow traffic observation value only with the two observation values, so that the performance superior to the conventional method is obtained. Is shown.

(Fifth Embodiment) This embodiment has the same structure as each of the above-described embodiments (especially the second or third embodiment) as a basic structure, and therefore, different parts will be described below. Will be described.

In this embodiment as well, as in the fourth embodiment, G
It is related to the UI, and the user uses the GU
Through I, sensor type and sensor information conversion formula, control
This makes it easy to specify each device to be adjusted.

13 and 14 are image diagrams of GUI screens when the present embodiment is applied to a look-up table design for determining traffic signal control parameters.

The window 301 in FIG. 13 is the road 303.
A traffic signal to be controlled by a lookup table (30
4 icon) is selected and specified by the user by clicking on the screen (the selected one is indicated by shading here), and the properties such as parameters to be controlled are specified from the sub screen (not shown). This is the window for doing. Sub-screen (not shown) for specifying properties such as using default parameters for the remaining traffic lights
Do more.

On the main screen 301, the GUI
A vehicle detector (icon 305 in the figure) and a line length measuring instrument (30 in the figure) prepared as a menu (302 in the figure) above
(6 icon) and other items that indicate the desired sensor can be freely selected and placed by dragging and dropping them to the desired location on the configuration diagram being displayed. It is input by an operation of designating detailed information such as information to be measured using a property input screen (not shown) that appears as.

In the contents illustrated in FIG. 13, the user selects and specifies the traffic signals 2 to 4 and the vehicle detectors A1 to A.
8 and the matrix length measuring device B1 are dragged and dropped. The selected traffic signals 2-4 are
It is treated as a traffic signal that adaptively changes parameters based on the sensor output, and the other traffic signals are treated as traffic signals that are controlled by fixed parameters.

The window 311 in FIG. 14 shows the relationship between the observed values obtained by the sensor configuration set by the user and the input vector of the lookup table, and the output vector of the lookup table and the output value to the traffic signal selected and designated by the user. In the window for defining the relationship with and on the GUI, the parts corresponding to the processing functions such as average value calculation (313 in the figure) and maximum value selection calculation (314 in the figure) prepared in advance are displayed on the sole box. The node corresponding to the input from the sensor designated by the user (312 in the figure), the node corresponding to the output to each traffic signal designated by the user (316 in the figure), and the lookup table to be learned. A corresponding node (315 in the figure) is displayed on the screen, and a link is provided between the node of each input and the function node added by the user. It indicates that you can connect Ri.

In the contents illustrated in FIG. 14, FIG.
The nodes corresponding to the inputs from the sensors A1 to A8 set in 1. and the nodes corresponding to the outputs to the selected traffic signals 2 to 4, the nodes corresponding to the lookup table, and the 2 added by the user from the sole box. One average value calculation node and one maximum value selection calculation node are displayed, and the average value of the observed values of the vehicle detector A1 and the vehicle detector A2 and the vehicle detector are displayed according to the link connection operation of the user. The average value of the observation value of A3 and the observation value of the vehicle detector A5, the maximum value of the observation value of the vehicle detector A4, the observation value of the vehicle detector A6 and the observation value of the vehicle detector A7, the matrix length measuring device signal 1 It indicates that the observed value is input as the input vector of the lookup table.

When the input by the GUI is completed, the look-up table designing apparatus of this embodiment automatically creates a look-up table according to the specified condition.

In this embodiment, the observation value observed by each sensor designated by the user is newly calculated based on the internal parameters set in step S11 for the first simulation process, and the calculation is performed. When it is impossible, a function of notifying the user is added, and a function of using a conversion formula or conversion rule designated by the user is added to the input conversion unit 11 and the output conversion unit 12.

Since the number of dimensions of the look-up table input vector and output vector is changed by the user, the maximum and minimum of the newly generated dimensional variables of the input vector, the number of mesh divisions, the size (number) of neighboring meshes, and As for the maximum and minimum of the newly generated dimension variables of the output vector, the user inputs necessary information, or default values are automatically set. Each functional block portion in FIGS. 2 and 10 has a configuration capable of coping with a change in the number of dimensions.

As described above, it is preferable to add the function of evaluating and displaying the performance of this automatically generated look-up table to the look-up table designing apparatus of this embodiment.

As described above, according to the present embodiment, the designation of the traffic signal that adaptively changes the parameter based on the sensor information and the traffic signal that is controlled by the fixed parameter, and the vehicle speed limit display is adaptively changed according to the traffic volume. By allowing the user to freely specify on the GUI, such as specifying a road, and comparing the control performances of the lookup tables automatically designed under the specified conditions, it is possible to support the design of the relevant control system according to the purpose.

Information that defines the relationship between the observed value obtained by the specified sensor configuration and the input vector of the lookup table, and the output vector of the lookup table and the output value to the specified control / adjustment target device It makes it easier to enter information that defines the relationship.

Furthermore, after input / output to / from the user lookup table is selected on the spot, learning is immediately performed, the performance of the created lookup table is displayed, and the user inputs the lookup table with reference to this. An interactive look-up table can be designed by changing the output, sensor configuration, and control / adjustment target device.

(Sixth Embodiment) This embodiment has the same structure as each of the embodiments described above (especially the second, third, and fifth embodiments) as a basic structure. The different parts will be described.

This embodiment relates to the GUI as in the fourth and fifth embodiments, and the user inputs the parameter input unit 2
Through the GUI of, sensor type and sensor information conversion formula,
This makes it easy to specify each device to be controlled / adjusted.

FIG. 15 is an image diagram of a GUI screen when this embodiment is applied to a look-up table design for determining a traffic signal control parameter.

In this embodiment, the vehicle detector is provided with the GUI as described using the window illustrated in FIG. 13 in the fifth embodiment, and in the system configuration diagram displayed in such a window. ID numbers of sensors and traffic signals that can be controlled and adjusted are displayed, and new sensors and signals can be registered as in the fifth embodiment.

The window illustrated in FIG. 15 is a GUI capable of inputting and outputting in the same tabular format as the spreadsheet software, and the internal parameters (independent) of the first and second simulations or the forward model of the system are independent. Corresponding to the traffic volume that can be specified by the operator), the cell corresponding to the observed value of each sensor such as the calculated vehicle detector, and each element of the input vector to the lookup table to be learned. A cell, a cell corresponding to each element of the output vector of the lookup table to be learned, a cell corresponding to each output parameter to the traffic light, a second simulation or forward model calculation result (vehicle at each intersection) Total number of stops and total delay time)
All or some of the cells corresponding to (1) and the cells corresponding to the antibody evaluation value are prepared.

In such a GUI environment, the user embeds an arithmetic expression or the like in each cell, and thereby, for example, a change type of sensor type and sensor information, designation of a traffic signal to be controlled / adjusted, weight coefficient in an antibody evaluation formula, It is possible to easily set and change various conditions such as changes. Less than,
Some examples of information that can be set / changed with such a GUI are shown below.

One input operation is to embed an arithmetic expression for obtaining a value with the value of the cell corresponding to each observation value as an argument in the cell corresponding to each element of the input vector to the lookup table, and It is possible to embed an arithmetic expression for obtaining a value by using the value of the cell corresponding to each element of the output vector of the lookup table as a part of the argument in the cell corresponding to the output parameter, or an operation to embed the default output value, as shown in FIG. Information defining the relationship between the observed value and the input vector of the lookup table and information defining the relationship between the output vector of the lookup table and the output value to the control / adjustment target device can be input.

The other operation is to embed an arithmetic expression corresponding to the first calculation simulation or forward model in which the value of the cell corresponding to the internal parameter is used as an argument in the cell corresponding to each observation value. Is.

Another operation is to embed the range of changing each internal parameter and the constraint condition expression in the cell corresponding to the internal parameter.

Another operation is to input the cell value corresponding to the internal parameter and the cell value corresponding to each output parameter to the traffic signal to the cell corresponding to the second calculation simulation or the calculation result of the forward model. Is to embed the arithmetic expression corresponding to the second calculation simulation or the forward model for obtaining the calculation result.

Another operation is to carry out an evaluation formula for obtaining an evaluation value by using the cells corresponding to the evaluation value of the antibody data and the values of some cells corresponding to the calculation results of the second calculation simulation or the forward model as arguments. To embed.

Although FIG. 15 exemplifies the case where all of the above seven types of cells are used, a GUI which supports a part of the cells, that is, a part of the above five operations is provided. You can use it.

When the input by the GUI is completed, the look-up table designing apparatus of this embodiment automatically creates a look-up table according to the specified condition.

As described above, it is preferable to add the function of evaluating and displaying the performance of this automatically generated look-up table to the look-up table designing apparatus of this embodiment.

According to the present embodiment, information defining the relationship between the observed value obtained by the sensor configuration and the input vector of the lookup table, and the output vector of the lookup table and the output to the designated control / adjustment target device. It becomes easy to input / change the information that defines the relationship with the value.

Further, according to the present embodiment, it becomes easy to create / change the calculation simulation or the forward model operation and input / change the evaluation formula.

The GUI in FIG. 15 is, for example, M
Microsoft's Microsoft Excel
It can also be implemented as an add-in system on a commercially available spreadsheet software such as.

Each function described in each of the above embodiments is
It can be realized as hardware or software. When implemented as software, it can also be implemented as a machine-readable medium that records a program for causing a computer to execute the above-described procedures or means.

The present invention is not limited to the above-mentioned embodiments, but can be implemented with various modifications within the technical scope thereof.

[0273]

According to the present invention, the computational simulation experiment is automatically repeated on behalf of the designer, and depending on the given combination of input-output variables, there is little input information available. It becomes possible to automatically design or support the design of the most appropriate and robust look-up table contents.

Further, by using the graphical user interface according to the present invention, it becomes easy to input various information when designing the look-up table.

[Brief description of drawings]

FIG. 1 is a schematic functional block diagram showing the configuration of a lookup table designing device according to a first embodiment of the present invention.

FIG. 2 is a functional block diagram showing the configuration of a look-up table design device according to a second embodiment of the present invention.

FIG. 3 is a flowchart showing a processing flow during antibody learning according to the same embodiment.

FIG. 4 is a diagram for explaining a process of extracting a neighborhood antibody from a neighborhood mesh.

FIG. 5 is a diagram for explaining an example of antibody generation processing.

FIG. 6 is a flowchart showing the flow of processing when creating a lookup table after learning in the same embodiment.

FIG. 7 is a diagram for explaining an example of processing for determining the content of each mesh.

FIG. 8 is a diagram showing an example of a look-up table for determining traffic signal parameters.

FIG. 9 is a flowchart showing the flow of another process during antibody learning of the same embodiment.

FIG. 10 is a functional block diagram showing the configuration of a lookup table designing device according to a third embodiment of the present invention.

FIG. 11 is a diagram showing an example of a GUI screen image when applied to a look-up table design for determining traffic signal control parameters.

FIG. 12 is a diagram for comparing simulation experiment results between a look-up table design device to which the present invention is applied and a conventional look-up table design device.

FIG. 13 is a diagram showing an example of a GUI screen image when applied to a look-up table design for determining traffic signal control parameters.

FIG. 14 is a diagram showing an example of a GUI screen image when applied to a look-up table design for determining traffic signal control parameters.

FIG. 15 is a diagram showing an example of a GUI screen image when applied to a look-up table design for determining traffic signal control parameters.

FIG. 16 is a diagram showing an example of division of a conventional pattern selection control lookup table.

FIG. 17 is a diagram showing an example of a traffic signal control system.

[Explanation of symbols]

1 ... Learning / design management unit 2 ... Parameter input unit 3, 3 _{1 to} 3 _n ... Antibody evaluation value determination unit 4 ... Candidate antibody production unit 5 ... Antibody storage unit 6 ... Lookup table creation unit 7 ... Lookup table 11 ... input conversion unit 12 ... output conversion unit 13 ... learning management unit 14 ... design management unit 31 ... parameter automatic change unit 32 ... simulation executing unit 33, 33 ₁ ~ 33 _m ... simulation result evaluating section 321: first simulation executing unit 322 _{1 to} 322 _m ... Second simulation executing unit 51 ... Neighboring antibody searching unit 52 ... Antibody database 53 ... Data changing unit 61 ... Input scanning unit 62 ... Antibody buffer 63 ... Maximum concentration antibody detecting unit 64 ... Look-up table writing unit 131 ... Antibody buffer 132 ... Evaluation value buffer 133 ... Concentration update calculation unit 134 ... Antibody deletion unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁷ identification code FI G08G 1/00 G06F 15/60 630 (58) Fields investigated (Int.Cl. ⁷ , DB name) G05B 11/00-13 / 04 G08G 1/00-9/02 G06F 17/50 G01N 1/00-17/08

Claims (13)

(57) [Claims] 1. A learning method for examining a relationship between an input vector and an output vector of a look-up table used for controlling a desired system is performed by using a calculation simulator or a forward model, and then the learning is performed based on a result of the learning. A look-up table design method for determining the contents of a look-up table, comprising an antibody including a pair of an input vector and an output vector for a target look-up table, which is generated upon establishment of a predetermined condition in the learning. In the data,
The evaluation target is obtained by obtaining antibody data having an input vector included in a predetermined neighborhood range in the input vector space from the antibody information storage means that stores a plurality of concentrations showing the relative superiority in association with each other. The evaluation value of each antibody data to be evaluated is determined by a predetermined method, and all the evaluation values obtained are taken into consideration. A method for designing a look-up table, characterized in that the concentration of each antibody data is updated based on the sex, and the antibody data is selected based on the concentration. 2. The learning is performed by repeatedly updating the concentration of the antibody data to be evaluated while changing the evaluation target, and the first calculation simulator or the forward model for each repetition of learning. The internal parameters of R are changed at random or by a predesignated procedure, the first calculation simulation or the forward model calculation is performed, and the observable specified in advance is obtained from the result of the first calculation simulation or the forward model calculation. Information is extracted, this is converted to generate an input vector, and the input vector space is divided into a specified number of divisions for each dimension of the input vector of the lookup table. The mesh containing the input vector is added with the number of adjacent meshes starting from this point and specified by each dimension. By determining all antibody data included in the neighborhood mesh, multiple antibody data to be evaluated are defined, and for each antibody data to be evaluated, the value of the part corresponding to the output vector of the coding content is converted. Then, the same internal parameters as the first are assumed and input to a second calculation simulator or a forward model, and a second calculation simulation or a forward model operation is performed to
The look-up table designing method according to claim 1, wherein the evaluation value of each antibody data is obtained based on the result of the calculation simulation or the forward model calculation. 3. When determining the contents of the look-up table, for each of a plurality of meshes formed by dividing the input vector space into the number of divisions specified for each dimension of the input vector of the look-up table. From the antibody information storage means after the learning, all the antibody data included in the neighborhood mesh formed by adding the target mesh and the number of adjacent meshes starting from this and specified by each dimension are obtained. 3. A process for writing a portion corresponding to the output vector of the coding content of the antibody data having the maximum density in the address corresponding to the target mesh in the lookup table is performed. Lookup table design method described. 4. When determining the contents of the look-up table, for each of a plurality of meshes formed by dividing the input vector space into the number of divisions specified for each dimension of the input vector of the look-up table. From the antibody information storage means after the learning, all the antibody data included in the neighborhood mesh formed by adding the target mesh and the number of adjacent meshes starting from this and specified by each dimension are obtained. The lookup table designing method according to claim 1 or 2, wherein the coding content of the antibody data having the maximum concentration among them is written into the lookup table. 5. A plurality of the first calculation simulators or forward models are used to generate a plurality of the input vectors in parallel, and each of the plurality of generated input vectors is an evaluation target individual. 3. The look-up table designing method according to claim 1, wherein the process of obtaining the evaluation value of the antibody data is executed in parallel by a plurality of the second calculation simulators or forward models. 6. A configuration diagram of a system to be controlled / adjusted is displayed by using a graphical user interface, and information about a sensor configuration input by the graphical user interface and an observation obtained by the designated sensor configuration are displayed. The look-up table design according to claim 1 or 2, wherein information that defines a relationship between a value and an input vector to the look-up table is accepted, and the look-up table is designed based on the inputted information. Method. 7. A configuration diagram of a system to be controlled / adjusted is displayed using a graphical user interface, and the system is represented in the form of an ambiguous probability distribution input by the graphical user interface. The information of internal parameters is received, and when the information is input, learning is performed by randomly changing the internal parameters given in an ambiguous form according to a given probability distribution.
Alternatively, the look-up table designing method according to item 2. 8. A configuration diagram of a system to be controlled / adjusted is displayed by using a graphical user interface, and first information on a sensor configuration input by the graphical user interface, control / adjustment target, and Information regarding each device, third information that defines the relationship between the observed value obtained by the specified sensor configuration and the input vector of the lookup table, and the output vector of the lookup table and the specified control / adjustment target device 4. The fourth information that defines the relationship with the output value to is received, and the lookup table is designed by learning based on the input information.
Look-up table design method described in. 9. A selection of a component corresponding to a processing function prepared in advance is accepted, and the selected component is displayed on the screen as a node corresponding to a visually recognizable processing function. A node corresponding to the input from the sensor designated on the system configuration diagram, a node corresponding to the output to each control / adjustment target device designated on the system configuration diagram, and a lookup table to be learned. Corresponding node is displayed on the screen, and the designation that should be entered by the graphical user interface should be connected by the link between the node corresponding to the desired input and the node corresponding to the desired processing function. 9. The lookup table according to claim 8, wherein the lookup table is obtained, and the third and fourth information is obtained based on the accepted designation. A total way. 10. A graphical interface capable of tabular input / output
On the user interface, a cell corresponding to each observation value as a result of the first calculation simulation or forward model calculation of the control / adjustment target system, and the second calculation simulation or order of the control / adjustment target system. A cell corresponding to each output to the directional model operation, a cell corresponding to each element of the input vector to the lookup table that is the learning target, and a cell corresponding to each element of the output vector of the lookup table that is the learning target Corresponding to each observed value of the control / adjustment target system in a cell corresponding to each element of the input vector to the look-up table, which is displayed on the screen by at least a part of It accepts the specification of embedding an arithmetic expression that takes the value of the specified cell as an argument. In the cell corresponding to each output to the control / adjustment target system input by the graphical user interface, the value of the cell corresponding to each element of the output vector of the lookup table is used as a part of the argument. Is accepted, or a specification for embedding a default output value is accepted, and an operation expression showing the relationship between the observed value obtained by these accepted specifications and the input vector of the lookup table, and the output vector of the lookup table, and the control The look-up table designing method according to claim 2, wherein the look-up table is designed by learning based on an arithmetic expression indicating a relationship with an output value to the device to be adjusted. 11. A graphical interface capable of input and output in tabular form.
On the user interface, cells corresponding to the internal parameters of the first and second calculation simulations or forward model calculations of the control / adjustment target system, and the second calculation simulations or forward direction of the control / adjustment target system Some cells corresponding to the calculation result of the model operation and at least a part of the cells corresponding to the evaluation value of each of the antibody data are displayed on the screen, and the cells are input by the graphical user interface. Designation of embedding an arithmetic expression corresponding to the first calculation simulation or forward model calculation for obtaining an observation value in the cell corresponding to the observed value with the value of the cell corresponding to the internal parameter as an argument; Corresponding to the internal parameters input by the interface The specification of embedding the range of changing the internal parameter and the constraint condition expression, and the cell corresponding to the calculation result of the second calculation simulation or the forward model calculation, which is input by the graphical user interface, In the second calculation simulation or the forward model calculation, the calculation result is obtained by using the value of the cell corresponding to the internal parameter and the value of the cell corresponding to each output to the second calculation simulation or the forward model calculation as an argument. When a corresponding arithmetic expression is designated to be embedded, the number corresponding to the calculation result of the second calculation simulation or the forward direction model calculation is input to the cell corresponding to the evaluation value of each antibody data input by the graphical user interface. Fill in the evaluation formula to obtain the evaluation value with the value of the cell At least one of the specification writing
11. The look-up table designing method according to claim 10, wherein the look-up table is designed by learning based on the content of one of the received designations that has been input. 12. The lookup according to claim 1 or 2.
Control using a graphical user interface using the table design method.
Information that displays the configuration diagram of the system to be adjusted and that defines the relationship between the sensor configuration input by the graphical user interface and the observation values obtained by the specified sensor configuration and the input vector to the lookup table And a computer-readable recording medium that stores a program for controlling a computer so that a lookup table is designed based on the input information. 13. A program for it to perform the learning about the robustness is designed the look-up table for a given internal parameter of the system to be subject to the look-up table to be designed, according to claim 1 or 2 Lookup table design method described in
Control using a graphical user interface
When a configuration diagram of the system to be adjusted is displayed and information of internal parameters of the system expressed in the form of an ambiguous probability distribution input by the graphical user interface is accepted, and the information is input, To control the computer so that the learning is performed by adding random fluctuations according to the given probability distribution for the given internal parameters in an ambiguous form, and the contents of the lookup table are determined based on this learning result. A computer-readable recording medium storing the program of.