Data processing device, program, and data processing method

The data processing device simplifies and parallelizes the calculation of combinatorial optimization problems with multivalued variables by generating transition candidates and calculating evaluation function changes, enhancing computational efficiency and reducing memory needs.

JP2024167800APending Publication Date: 2024-12-04FUJITSU LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2023084138
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-22
Publication Date
2024-12-04

AI Technical Summary

Technical Problem

Existing Ising machines face challenges in efficiently calculating combinatorial optimization problems expressed with multivalued variables due to complex data movement and difficulty in performing parallel trials.

Method used

A data processing device and method that generates transition destination candidates for multivalued variables, calculates the change in evaluation function values, and selects variables for transition based on these candidates, enabling parallel processing of combinatorial optimization problems.

Benefits of technology

This approach simplifies data movement and enables parallel computation of combinatorial optimization problems with multivalued variables, improving computational efficiency and reducing memory requirements compared to binary state variables.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024167800000001_ABST
    Figure 2024167800000001_ABST
Patent Text Reader

Abstract

To allow a combinatorial optimization problem expressed by a multi-valued variable to be computed by parallel trials.SOLUTION: A data processing device includes: a storage unit 11 configured to store evaluation function information 11a indicating an evaluation function of a combinatorial optimization problem represented by using a plurality of multi-valued variables; and a processing unit 12 configured to generate, based on the evaluation function information 11a, transition destination candidates that are a part of respective transitionable ranges of values of the plurality of multi-valued variables, calculate, for each of the plurality of multi-valued variables, a change amount of the values of the evaluation function associated with a transition to the transition destination candidates, based on the evaluation function information 11a, determine the multi-valued variables for which the transition is accepted, from among the plurality of multi-valued variables, based on the change amount, and cause the values of the determined multi-valued variables to transition to the values of the transition destination candidates.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

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

[0002] Ising machines that use an Ising-type evaluation function are devices that can calculate large-scale discrete optimization problems that von Neumann-type computers are not good at. When using an Ising machine, a combinatorial optimization problem is converted into an Ising model that represents the behavior of the spin of a magnetic material. The Ising machine then searches for the state of the Ising model that minimizes the value of the Ising-type evaluation function by using Markov chain Monte Carlo methods such as simulated annealing and replica exchange (also known as parallel tempering). The state that results in the smallest of the minimum values ​​of the evaluation function is the optimal solution. The state of the Ising model can be expressed by a combination of the values ​​of multiple state variables. The value of each state variable can be 0 or 1.

[0003] The Ising-type evaluation function is defined, for example, by the following equation (1).

[0004]

number

[0005] The term on the right side is the sum of the values ​​of the two state variables (0 or 1) and the weight value (which represents the strength of the interaction between the two state variables) for all combinations of all state variables in the Ising model, without any omissions or overlaps. x i is the state variable with identification number i, x j is the state variable with identification number j, and W ij is the weight value indicating the magnitude of interaction between state variables with identification numbers i and j. The two terms on the right side are the sum of the products of the bias coefficient and the state variable for each identification number. b i indicates the bias coefficient for identification number = i.

[0006] Also, x i The change in the value of the evaluation function due to the change in the value of i ) is expressed by the following equation (2).

[0007]

number

[0008] In formula (2), x i When changes from 1 to 0, Δx i becomes -1, and the state variable x i When changes from 0 to 1, Δx i is 1. Note that h i is called the local field, and Δx i Depending on h i Multiplying by the sign (+1 or -1) is ΔE i It becomes.

[0009] An Ising machine, for example, ΔE i If x is smaller than the noise value obtained based on the random number and temperature parameter values, i The solution is searched for by repeating the process of inverting the value of and updating the local field.

[0010] ΔE i Calculation of x i Since the process of determining whether to invert the value of can be performed multiple times in parallel, parallel trials for multiple state variables are possible. Incidentally, some combinatorial optimization problems can be expressed using multi-valued variables, such as an assignment problem. In an assignment problem, an assignment state indicating which destination an element is assigned to can be expressed by a multi-valued variable. Conventionally, a technique has been proposed for searching for a solution to an assignment problem using an Ising machine (see, for example, Patent Documents 1 and 2). [Prior art documents] [Patent documents]

[0011] [Patent Document 1] Patent Publication No. 2023-1055 [Patent Document 2] Patent Publication No. 2022-165250 Summary of the Invention [Problem to be solved by the invention]

[0012] In the technology that uses an Ising machine to search for a solution to the allocation problem, the allocation state, which is a multi-valued variable, is expressed as x, which is either 0 or 1. i To convert to x i The number of x i When the number of x becomes large, i Regarding ΔE i The movement of data used in the calculation can be complicated, making parallel trials difficult.

[0013] In one aspect, the present invention has an object to make it possible to compute a combinatorial optimization problem expressed with multi-valued variables through parallel trials. [Means for solving the problem]

[0014] In one embodiment, there is provided a data processing device having: a memory unit that stores evaluation function information regarding an evaluation function for a combinatorial optimization problem expressed using a plurality of multi-valued variables; and a processing unit that generates transition destination candidates which are a part of the possible transition ranges of values ​​of each of the plurality of multi-valued variables based on the evaluation function information, calculates, for each of the plurality of multi-valued variables, an amount of change in the value of the evaluation function associated with a transition to the transition destination candidate based on the evaluation function information, determines a multi-valued variable from among the plurality of multi-valued variables that will accept the transition based on the amount of change, and transitions the value of the determined multi-valued variable to the value of the transition destination candidate.

[0015] Also, in one embodiment, a program is provided. Also provided in one embodiment is a method for processing data. Effect of the Invention

[0016] In one aspect, the present invention is capable of calculating a combinatorial optimization problem expressed with multiple variables through parallel trials. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 illustrates an example of a data processing device according to a first embodiment; [Diagram 2] FIG. 13 is a diagram illustrating an example of transition destination candidates of n multi-valued variables. [Diagram 3] FIG. 1 illustrates an example of a data processing device according to a second embodiment. [Figure 4] FIG. 13 is a diagram illustrating an example of a current state and a transition destination candidate state. [Diagram 5] FIG. 13 is a diagram showing an example of matrix elements read out from the F memory. [Figure 6] FIG. 13 is a diagram showing an example of matrix elements read out from a D memory. [Figure 7] 10 is a timing chart illustrating an example of the operation of the data processing device according to the second embodiment. [Figure 8] 13 is a flowchart illustrating an example of a processing procedure of the data processing device according to the second embodiment. [Figure 9] FIG. 13 illustrates an example of a data processing device according to a third embodiment. [Figure 10] 13 is a timing chart illustrating an example of the operation of the data processing device according to the third embodiment. [Figure 11] FIG. 13 is a diagram illustrating an example of a current state of a plurality of transition destination candidates and states of the transition destination candidates. [Figure 12] FIG. 13 is a diagram showing an example of matrix elements read out from a D memory. [Figure 13] FIG. 13 illustrates an example of a data processing device according to a fifth embodiment. [Figure 14] FIG. 13 is a diagram showing an example of matrix elements read out from a ΔD memory. [Figure 15] 13 is a timing chart showing an example of the operation of the data processing device according to the fifth embodiment. [Figure 16] FIG. 1 illustrates an example of hardware of a computer that is an example of a data processing device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] Hereinafter, an embodiment of the invention will be described with reference to the drawings. A data processing device according to each of the following embodiments searches for a solution to a combinatorial optimization problem expressed by a plurality of multi-valued variables. One such combinatorial optimization problem is an assignment problem such as a quadratic assignment problem (QAP (Quadratic Assignment Problem)). The assignment problem is an NP-hard combinatorial optimization problem with various real-world applications, such as vehicle routing and block placement in an FPGA (Field-Programmable Gate Array). The assignment problem is expressed using a flow matrix and a distance matrix.

[0019] In a QAP that allocates n elements to n destinations, the flow matrix and the distance matrix are expressed by the following equations (3) and (4).

[0020]

number

[0021]

number

[0022] The flow matrix (F) is a matrix with n rows and n columns. ij is a matrix element in row i and column j, and represents the flow amount (for example, the cost of transporting materials between facilities) between elements (for example, facilities) with identification numbers = i, j.

[0023] The distance matrix (D) is an n-row, n-column matrix. d kl is the distance in k rows and l columns, and represents the distance between the locations where the identification numbers k and l are assigned (for example, the locations where the facilities are located). The QAP is calculated by searching for an assignment that minimizes the value of the evaluation function expressed by the following equation (5).

[0024]

number

[0025] In equation (5), φ i (i=1 to n) is a multi-valued variable that indicates the allocation of the element with the identification number = i. For this reason, φ i φ j The subscript iφ on b indicates the distance between the allocation destination of the element with the identification number = i and the allocation destination of the element with the identification number = j. i The element with the identification number = i is φ i represents the bias coefficient when assigned to . Note that if the sign of the evaluation function is changed, the QAP calculation is performed by searching for the state where the evaluation function value is maximized.

[0026] (First embodiment) FIG. 1 illustrates an example of a data processing device according to a first embodiment. The data processing device 10 of the first embodiment includes a storage unit 11 and a processing unit 12.

[0027] The storage unit 11 may include a volatile semiconductor memory such as a RAM (Random Access Memory), or may include a non-volatile storage such as a HDD (Hard Disk Drive) or a flash memory. The storage unit 11 may also include both a volatile semiconductor memory and a non-volatile storage.

[0028] The storage unit 11 stores evaluation function information 11a relating to an evaluation function for a combinatorial optimization problem expressed using a plurality of (hereinafter, n) multi-valued variables. The evaluation function information 11a includes information such as possible transition ranges of the values ​​of the n multi-valued variables and coefficients included in the evaluation function. Each of the n multi-valued variables may be an integer variable or a continuous variable. Furthermore, the n multi-valued variables may have different possible transition ranges.

[0029] Figure 1 shows an example of n multi-valued variables, φ1 to φn The subscripts 1 to n represent the identification numbers of the multi-valued variables. In the example of FIG. 1, φ i is an integer variable, and φ i The possible transition range is m values: 1, 2, ..., m.

[0030] The coefficients included in the evaluation function include, for example, the matrix elements (flow amount and distance) of the flow matrix and distance matrix as shown in equations (3) and (4). The storage unit 11 may further store calculation conditions for the combinatorial optimization problem. For example, when a search is performed using the replica exchange method, the calculation conditions include the number of replicas and the value of a temperature parameter to be set for each replica.

[0031] The processing unit 12 can be realized by an electronic circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA. However, the processing unit 12 can also be realized by a processor such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a DSP (Digital Signal Processor). The processor executes a program stored in a memory such as a RAM (which may be the storage unit 11). A collection of processors may be called a multiprocessor or simply a "processor." Furthermore, the processing unit 12 may include a processor and an electronic circuit such as an ASIC or an FPGA.

[0032] FIG. 1 shows an example of the procedure of the processing performed by the processing unit 12. The processing unit 12 generates transition destination candidates, which are part of the possible transition range of values, for each of the n multi-valued variables based on the evaluation function information 11a (step S1).

[0033] For example, the processing unit 12 randomly generates one transition destination candidate from within the transition possible range for each of the n multi-valued variables. The processing unit 12 may randomly generate two or more transition destination candidates from within the transition possible range for each of the n multi-valued variables (see the fourth embodiment described later).

[0034] When a combinatorial optimization problem has constraints, transition destination candidates are generated to satisfy the constraints. For example, for a multi-valued variable for which a transition destination candidate that does not satisfy a constraint has been generated, transition destination candidates may be generated randomly until the constraint is satisfied.

[0035] The processing unit 12 may generate transition destination candidates according to a predetermined rule, rather than randomly. In the example of FIG. i Transition candidate φ i ' is generated as m. 2 is a diagram showing an example of transition destination candidates for n multi-valued variables. The horizontal axis represents the identification number of the multi-valued variable, and the vertical axis represents the value of the multi-valued variable. Plot group 15a represents the current values ​​or initial values ​​of the n multi-valued variables. Plot group 15b represents the values ​​of transition destination candidates for the initial values ​​or current values ​​of the n multi-valued variables.

[0036] In the example of FIG. 2, each multi-valued variable can have two possible values, and the transition destination for the current value of each multi-valued variable is narrowed down to one. After the process of step S1, the processing unit 12 calculates the amount of change in the value of the evaluation function accompanying the transition to the transition destination candidate, for each of the n multi-valued variables, based on the evaluation function information 11a (step S2).

[0037] When the evaluation function is expressed by equation (5), φ i From φ' i The change in the transition from i →φ' i ) can be expressed by the following equation (6).

[0038]

number

[0039] In equation (6), the subscript φ' i φ j The distance d between the current allocation destination of the element with identification number = j and the candidate allocation destination of the element with identification number = i is the distance between the current allocation destination of the element with identification number = j and the candidate allocation destination of the element with identification number = i. j The candidate for the assignment of the element with the identification number = i is φ' i It can be expressed as follows. The subscript φ i φ j The distance d between the allocation destination of the element with identification number = j and the current allocation destination of the element with identification number = i is φ i It can be expressed as follows. The subscript iφ' i b with the identification number = i is φ' i is the bias coefficient when assigned to the subscript iφ i b with the identification number = i is the element φ i is the bias factor when assigned to

[0040] After the process of step S2, the processing unit 12 determines a multi-valued variable that accepts a transition to a transition destination candidate based on the calculated change amount (step S3). For example, the processing unit 12 determines φ i From φ' i The processing unit 12 determines whether to accept the transition to φ. The predetermined value is, for example, a noise value obtained based on a random number and the value of the temperature parameter. i If the amount of change accompanying the transition is smaller than log(rand)×T, which is an example of a noise value obtained based on a uniform random number (rand) between 0 and 1 and a temperature parameter (T), the transition is accepted.

[0041] The processing unit 12 may accept the transition of the multi-valued variable that minimizes the value of the evaluation function based on the calculated amount of change. The processing unit 12 transitions the value of the multi-valued variable that receives the transition to the value of the transition destination candidate (step S4).i The value of φ' is a transition candidate. i An example is shown in which the value of φ' is transitioned to φ'. i If = m, then φ i The value of changes to m. Note that the original value of φ i The value of φ' is a transition candidate. i The value is:

[0042] When there are multiple multi-valued variables for which transitions have been accepted, the processing unit 12 may select one from among those multi-valued variables. In that case, for example, the processing unit 12 selects a multi-valued variable for which the value of the evaluation function is the smallest due to the amount of change accompanying the transition from among the multiple multi-valued variables for which transitions have been accepted. The processing unit 12 may select a multi-valued variable to which a transition is randomly made from among the multiple multi-valued variables for which transitions have been accepted.

[0043] After the process of step S4, the process from step S2 is repeated. At this time, the processing unit 12 may calculate a difference value of the change amount of the evaluation function accompanying the transition by the process of step S4 in the process of step S2, and update the change amount using the difference value. φ j is φ' j When the transition to i →φ' i When j≠i, the difference value of the change amount accompanying the transition is expressed by the following equation (7).

[0044]

number

[0045] The processing unit 12 reads out the matrix elements of the flow matrix and the distance matrix included in the formula (7) from the storage unit 11, and calculates δΔE(φ i →φ' i ) can be calculated. In addition, when j=i, φ i →φ' i The difference in the amount of change accompanying the transition is expressed by the following equation (8).

[0046]

number

[0047] The processing unit 12 can update the amount of change when j≠i using the difference value shown in equation (7) according to the following equation (9).

[0048]

number

[0049] The processing unit 12 can update the amount of change when j=i using the difference value shown in equation (8) according to the following equation (10).

[0050]

number

[0051] Incidentally, equation (9) can also be expressed as equation (11) below.

[0052]

number

[0053] In equation (11), Δd in the brackets of the second term on the right side is k=φ j or k=φ' j Then, it is expressed by the following equation (12).

[0054]

number

[0055] δΔE(φ i →φ' i An example of a configuration for efficiently calculating ) will be described in the second embodiment. The processing of steps S2 to S4 can be performed in parallel for n multi-valued variables. In other words, the processing of steps S2 to S4 can be realized by a machine (represented as multi-color spin machine 12a in FIG. 1) that is an extension of an Ising machine capable of parallel trials so as to be capable of multi-valued processing. However, the processing of steps S2 to S4 may also be performed serially for n multi-valued variables.

[0056] Note that, each time the process of step S3 is repeated a predetermined number of times, the process of step S1 described above may be performed again to regenerate transition destination candidates for each multi-valued variable. Alternatively, if the process of step S3 is repeated a predetermined number of times but the minimum value of the evaluation function is not updated, transition destination candidates for each multi-valued variable may be regenerated.

[0057] When performing the simulated annealing method, the processing unit 12 decreases the value of the above-mentioned temperature parameter (T) according to a predetermined temperature parameter change schedule while the processes of steps S2 to S4 are repeated. Then, the processing unit 12 outputs, for example, the values ​​of the n multi-valued variables when a predetermined search end condition is satisfied as a search result for a solution to the combinatorial optimization problem (for example, displays on a display device not shown). Note that the processing unit 12 may update the value of the evaluation function expressed by, for example, Equation (5) every time a transition is accepted, and may hold the minimum value so far and the values ​​of the n multi-valued variables when the minimum value was obtained. In that case, the processing unit 12 may output, as a search result, the values ​​of the n multi-valued variables corresponding to the minimum value stored when the search end condition is satisfied.

[0058] When the processing unit 12 performs the replica exchange method, the processing unit 12 performs the processes of steps S2 to S4 shown in FIG. 1 for each of a plurality of replicas to which different temperature parameter values ​​are set. The processing unit 12 performs replica exchange every time the processes of steps S2 to S4 are repeated a predetermined number of times. For example, the processing unit 12 randomly selects two of the plurality of replicas. The processing unit 12 then exchanges the temperature parameter value or the n multi-valued variable values ​​between the two selected replicas with a predetermined exchange probability based on the difference in the evaluation function value between the replicas or the difference in the temperature parameter value. For example, the processing unit 12 updates the value of the evaluation function expressed by formula (5) every time a transition is accepted in each replica, and holds the minimum value so far and the values ​​of the n multi-valued variables when the minimum value is obtained. Then, the processing unit 12 outputs, as a search result, the values ​​of the n multi-valued variables corresponding to the minimum value among all replicas among the minimum values ​​stored in each replica, when a predetermined search end condition is satisfied, for example.

[0059] As described above, the processing unit 12 of the data processing device 10 of the first embodiment generates transition destination candidates, which are part of the transition possible range of values, for each of the n multi-valued variables based on the evaluation function information 11a. Then, for each of the n multi-valued variables, the processing unit 12 calculates the amount of change in the value of the evaluation function associated with the transition to the transition destination candidate based on the evaluation function information 11a, and determines the multi-valued variable that will accept the transition based on the amount of change. Furthermore, the processing unit 12 transitions the value of the determined multi-valued variable to the value of the transition destination candidate.

[0060] This makes it easier to perform parallel calculations than when each of n multi-valued variables is converted into a large number of binary state variables of 0 or 1, since it avoids the complexity of data movement used to calculate the amount of change in the evaluation function. As a result, combinatorial optimization problems expressed with multi-valued variables can be calculated through parallel trials. As parallel trials become possible for combinatorial optimization problems expressed with multi-valued variables, the calculation efficiency of such combinatorial optimization problems is improved.

[0061] In addition, when calculating an assignment problem in which n elements are assigned to m destinations using state variables that are either 0 or 1, the number of state variables that represent the assignment state is n × m, and the number of weighting coefficients between the state variables is n 2 ×m 2 In contrast, by using a multi-valued variable that represents the allocation state, the coefficients to be stored are n 2 The flow amount of pieces and m 2 When the amount of change is updated using the difference value as described above, the amount of change is stored, but the amount of data required for storage is on the order of n × m, so even if the amount of flow and the distance are combined, the memory capacity can be made smaller than when a binary state variable is used.

[0062] (Second embodiment) The data processing device of the second embodiment searches for a solution to an assignment problem expressed using a flow matrix and a distance matrix.

[0063] FIG. 3 illustrates an example of a data processing device according to the second embodiment. The data processing device 20 of the second embodiment includes a storage unit 21 and a processing unit 22. The storage unit 21 includes an F memory 21a and D memories 21b and 21c. The F memory 21a and the D memories 21b and 21c may be, for example, a volatile semiconductor memory such as a RAM, or a non-volatile storage such as a flash memory.

[0064] The F memory 21a stores the flow matrix expressed by the equation (3). The D memory 21b stores a distance matrix expressed by equation (4) that includes matrix elements arranged based on the current values ​​(representing the current states) of n multi-valued variables.

[0065] The D memory 21c stores a distance matrix expressed by equation (4) including matrix elements arranged based on the values ​​(representing the states of the transition destination candidates) of each of the n multi-valued variables.

[0066] When multiple replicas are used for solution search, such as when replica exchange is performed, the above two types of distance matrices are stored for each replica. The flow matrix may be shared among multiple replicas.

[0067] The processing unit 22 has a δΔE calculation circuit 22a, a ΔE update circuit 22b, a ΔE holding circuit 22c, a transition acceptance determination circuit 22d, a column swapping circuit 22e, a state holding and updating circuit 22f, and a control circuit 22g.

[0068] The ΔE calculation circuit 22a calculates φ j is φ' j The difference in the amount of change in the evaluation function value when the transition occurs is calculated in n parallel steps (expressed in the above formula (7)). Hereinafter, the amount of change is represented as ΔE, and the difference is represented as δΔE.

[0069] 3 shows an example of the δΔE calculation circuit 22a. The δΔE calculation circuit 22a has registers 22a1, 22a2, 22a3, 22a4, and 22a5, adders 22a6, 22a7, and 22a8, and a multiplier 22a9.

[0070] The register 22a1 stores the flow amount (f ij ) to hold the The register 22a2 stores the allocation destination (φ j ) and the current allocation of the element with identification number = i (φ i ) to maintain the distance.

[0071] The register 22a3 stores the allocation candidate (φ' j ) and the assignment destination of the element with the identification number = i (φ i ) to maintain the distance. The register 22a4 stores the allocation destination (φ j ) and the candidate for assignment of the element with identification number = i (φ' i ) to maintain the distance.

[0072] The register 22a5 stores the allocation candidate (φ' j ) and the candidate for assignment of the element with identification number = i (φ' i ) to maintain the distance.

[0073] Adder 22a6 outputs a value obtained by subtracting the distance held in register 22a2 from the distance held in register 22a3. Adder 22a7 outputs a value obtained by subtracting the distance held in register 22a4 from the distance held in register 22a5.

[0074] Adder 22a8 outputs a value obtained by subtracting the output value of adder 22a6 from the output value of adder 22a7. Multiplier 22a9 outputs the product of the flow amount held in register 22a1 and adder 22a8. The output of multiplier 22a9 becomes δΔE.

[0075] The calculation of δΔE is performed by using the circuit configuration described above, and the n multi-valued variables φ1 to φ n can be performed in parallel. When i≠j, the ΔE update circuit 22b updates ΔE by adding ΔΔE calculated by the ΔΔE calculation circuit 22a to the original ΔE. When i=j, the ΔE update circuit 22b updates ΔE by inverting the sign of ΔE. The ΔE update circuit 22b performs the update process of ΔE on n multi-valued variables φ1 to φ for i=1 to n. n can be performed in parallel.

[0076] The ΔE holding circuit 22c holds the above ΔE relating to n multi-valued variables. When a solution search is performed using a plurality of replicas, such as when replica exchange is performed, a ΔE holding circuit 22c is provided for each replica.

[0077] The transition acceptance determination circuit 22d determines a multi-valued variable that accepts a transition to a transition destination candidate based on the calculated ΔE. The transition acceptance determination circuit 22d determines whether or not to accept a transition of the value of the multi-valued variable based on, for example, a comparison result between ΔE and a predetermined value supplied from the control circuit 22g. The predetermined value is, for example, a noise value obtained based on a random number and a temperature parameter value. The transition acceptance determination circuit 22d accepts the transition when, for example, ΔE associated with the transition is smaller than log(rand)×T, which is an example of a noise value obtained based on a uniform random number (rand) between 0 and 1 and a temperature parameter (T). The transition acceptance determination circuit 22d outputs an identification number of the multi-valued variable that has accepted the transition.

[0078] The column permutation circuit 22e permutes the columns of the distance matrix in the D memory 21b and the distance matrix in the D memory 21c, which are specified by the identification numbers of the multi-valued variables that accept the transition. When the reading from the distance matrix for calculating δΔE is performed column by column, the column permutation circuit 22e may permutate rows instead of columns.

[0079] The reason for swapping the columns or rows of the distance matrix in this way is as follows. Although not shown, the δΔE calculation circuit 22a, the ΔE update circuit 22b, the ΔE hold circuit 22c, and the transition acceptance determination circuit 22d each have a φ i The n circuit blocks are arranged in the order of the identification number=i. The n f ij In order to facilitate parallel propagation, the f ij The part that stores i is placed near the circuit block corresponding to i. In order to use the same parallel propagation to read the matrix elements of the distance matrix, i and φ j The distance between the first identification number (e.g., column number) is i and the second identification number (e.g., row number) is φ j This is why columns or rows are swapped in the distance matrix.

[0080] An example of the replacement will be described later (see FIG. 6). The state holding and updating circuit 22f holds the values ​​of n multi-valued variables, and transitions the value of the multi-valued variable corresponding to the identification number output by the transition acceptance determination circuit 22d to a transition destination candidate value. The state holding and updating circuit 22f also specifies the matrix element to be read from the F memory 21a to the δΔE calculation circuit 22a according to the identification number of the multi-valued variable that accepts the transition.

[0081] Furthermore, the state holding and updating circuit 22f specifies matrix elements to be read from the D memories 21b and 21c to the δΔE calculation circuit 22a according to values ​​before and after the transition of the multi-valued variable that accepts the transition. When a solution search is performed using multiple replicas, such as when replica exchange is performed, the state holding and updating circuit 22f holds the values ​​of n multi-valued variables for each replica.

[0082] The control circuit 22g controls each unit of the data processing device 20. The control circuit 22g also generates and regenerates transition destination candidates for each multi-valued variable, and rearranges the elements of the distance matrix stored in the D memory 21c so as to correspond to the generated transition destination candidates. The control circuit 22g also generates a predetermined value used for comparison with ΔE, for example, and supplies it to the transition acceptance determination circuit 22d. The control circuit 22g also controls pipeline processing, which will be described later.

[0083] Fig. 4 is a diagram showing an example of a current state and a transition destination candidate state in an assignment problem of assigning eight facilities to eight locations.

[0084] The current state is the position where the eight facilities are currently located (current facility positions), and is represented by the multi-valued variables φ1 to φ8. 1 to 8 are facility numbers. For example, as shown in Figure 4, the facility with facility number = 1 is located at position φ1 = 2, and the facility with facility number = 6 is located at position φ6 = 4.

[0085] The candidate destination states are the positions of the eight facility candidate destinations, and are represented by multi-valued variables φ'1 to φ'8. For example, as shown in Figure 4, the candidate destination position of the facility with facility number = 1 is φ'1 = 4, and the candidate destination position of the facility with facility number = 6 is φ'6 = 2.

[0086] In the current state as shown in FIG. 4, when the transition of φ6 to φ'6 is accepted, an example of the matrix elements read out for the calculation of δΔE is shown below. FIG. 5 is a diagram showing an example of matrix elements read out from the F memory.

[0087] When the transition of φ6 to φ'6 is accepted, the matrix element (f i6 ) is read out. FIG. 6 is a diagram showing an example of matrix elements read out from the D memory.

[0088] The matrix elements (distances) of the distance matrix stored in D memories 21b and 21c are arranged in the order of location identification numbers (1 to 8) along the columns of the matrix, and in the order of facility numbers along the rows. In the distance matrix stored in D memory 21b, the matrix elements are arranged based on the current values ​​of φ1 to φ8. In the distance matrix stored in D memory 21c, the matrix elements are arranged based on the values ​​of transition destination candidates φ'1 to φ'8 of φ1 to φ8.

[0089] In the distance matrices stored in the D memories 21b and 21c, a row is specified by the value of a multi-valued variable that accepts a transition or the value of a transition of a multi-valued variable, and the matrix elements included in that row are read out.

[0090] When the transition of φ6 to φ'6 is accepted, the current position (φ6) of the facility with facility number = 6 and the current position (φ i In the example shown in FIG. 4, since φ6=4, the distance matrix φ jFurther, the matrix elements in the fourth row are read out from the D memory 21b. The destination candidate position (φ'6) of the facility with facility number = 6 and the current position (φ i In the example shown in FIG. 4, since φ′6=2, the distance matrix φ j =The matrix elements in the second row are read out.

[0091] On the other hand, the current position (φ6) of the facility with facility number = 6 and the candidate destination position (φ' i In the example shown in FIG. 4, since φ6=4, the distance matrix φ j Further, the matrix elements in the fourth row are read out from the D memory 21c. The position (φ'6) of the candidate destination of the facility with the facility number = 6 and the position (φ' i In the example shown in FIG. 4, since φ′6=2, the distance matrix φ j =The matrix elements in the second row are read out.

[0092] Distance matrix φ from D memory 21b, 21c j The matrix elements in the fourth row can be read out in parallel at the first read timing. j = Reading out the matrix elements in the second row can be performed in parallel at the second read timing.

[0093] After the calculation of the difference value shown in formula (7), the matrix elements in the sixth column of each distance matrix stored in the D memories 21b and 21c are swapped by the column swapping circuit 22e. As a result, the distance matrices in the D memories 21b and 21c reflect the updated state.

[0094] Next, an example of the operation of the data processing device 20 according to the second embodiment will be described. 7 is a timing chart showing an example of the operation of the data processing device according to the second embodiment. In the example of the operation of FIG. 7, an example of pipeline processing using four replicas is shown. The processing includes a trial phase and an update phase.

[0095] The trial phase includes a process in which the transition acceptance determination circuit 22d acquires ΔE from the ΔE holding circuit 22c, and a process in which the transition acceptance determination circuit 22d performs an acceptance determination of a transition. The update phase includes a process of reading out matrix elements of the flow matrix from the F memory 21a, a process of reading matrix elements of the distance matrix from the D memories 21b and 21c, and a process of replacing columns of the distance matrix in the D memories 21b and 21c. The D memories 21b and 21c are provided for each replica. For this reason, in FIG. 7, the D memories 21b and 21c of replicas 0 to 3 are represented as "D0 memory", "D0 memory", "D1 memory", "D2 memory", and "D3 memory", respectively. The update phase also includes a process in which the δΔE calculation circuit 22a holds matrix elements of the flow matrix or matrix elements of the distance matrix (represented as F / D data in FIG. 7), calculates δΔE, and updates ΔE.

[0096] Between timing t1 and timing t4, a process of acquiring ΔE is performed in the order of replica 0, replica 1, 2, and 3. Between timing t2 and timing t5, acceptance determination is performed in the above order, starting with replica 0. For each replica, an identification number (denoted as "Idk") of the multi-valued variable whose transition has been accepted is output from the transition acceptance determination circuit 22d.

[0097] Between timing t3 and timing t9, a process of reading out matrix elements of the flow matrix from the F memory 21a and a process of reading out matrix elements of the distance matrix from the D memories 21b and 21c are performed. Reading out matrix elements of the distance matrix is ​​performed in two cycles for each replica. For example, as shown in FIG. 6, this is because matrix elements of two rows are read out twice in parallel from the D memories 21b and 21c.

[0098] From timing t4 to t10, the registers 22a1 to 22a5 of the ΔΔE calculation circuit 22a hold the matrix elements read from the F memory 21a and the D memories 21b and 21c. From timing t4 to t7, the matrix elements used to calculate ΔΔE for replica 0 are held, from timing t5 to t7, the calculation of ΔΔE is performed, and from timing t7 to t8, the calculation result is finalized. After timing t7, the same process as that for replica 0 is performed for replicas 1, 2, and 3, in that order. Also, from timing t8 onwards, ΔE for each replica is updated each time the calculation result of ΔΔE for each replica is finalized.

[0099] After timing t11, in the distance matrix stored in the D memories 21b and 21c as shown in Fig. 6, columns including n matrix elements are swapped in parallel for each replica. Since the n matrix elements included in each column are swapped between two columns one by one, the swapping is completed in n cycles.

[0100] FIG. 8 is a flowchart of an example of a processing procedure of the data processing device according to the second embodiment. First, under the control of the control circuit 22g, initialization of each unit of the data processing device 20 (step S10), data setting, and operation setting (step S11) are performed.

[0101] In the processing of steps S10 and S11, a flow matrix is ​​written to the F memory 21a, and a distance matrix including matrix elements arranged based on the initial values ​​of n multi-valued variables is written to the D memory 21b. Furthermore, for each multi-valued variable, transition destination candidates are generated, for example, as shown in Fig. 4. Then, the distance matrix including matrix elements arranged based on the values ​​of the transition destination candidates is written to the D memory 21c.

[0102] Furthermore, an initial value of the evaluation function expressed by equation (5) is calculated and held based on the initial values ​​of the n multi-valued variables and the evaluation function information (flow matrix, distance matrix, bias coefficient, etc.). Also, for each multi-valued variable, an initial value of ΔE expressed by equation (6) is calculated and held in the ΔE holding circuit 22c. Also, when the replica exchange method is performed, an initial value of the temperature parameter (T) is set for each replica. A different value is set for the temperature parameter for each replica. In addition, a termination condition for the solution search, etc. are set.

[0103] The control circuit 22g determines whether or not to change the transition destination candidate (step S12). For example, every time the acceptance determination process of step S16 described below is repeated a predetermined number of times, the control circuit 22g determines to change the transition destination candidate. The control circuit 22g may determine to change the transition destination candidate when the minimum value of the evaluation function is not updated even after the process of step S16 is repeated a predetermined number of times. When it is determined to change the transition destination candidate, the process of step S13 is performed, and when it is not determined to change the transition destination candidate, the process of step S15 is performed.

[0104] In the process of step S13, the control circuit 22g changes the transition destination candidate, and rewrites the distance matrix stored in the D memory 21c to one based on the changed transition destination candidate value. Furthermore, the control circuit 22g calculates ΔE based on the current multi-valued variable value and the changed transition destination candidate value, and holds it in the ΔE holding circuit 22c (step S14).

[0105] In the process of step S15, the transition acceptance determination circuit 22d acquires n ΔE from the ΔE holding circuit 22c. Then, the transition acceptance determination circuit 22d determines whether or not to accept a transition to a transition destination candidate for each of the n multi-valued variables based on the n ΔE (step S16). If there is a multi-valued variable that accepts a transition to a transition destination candidate, the transition acceptance determination circuit 22d selects a multi-valued variable to be transitioned so as to satisfy a constraint condition (step S17).

[0106] The control circuit 22g determines whether or not there is a multi-valued variable that has accepted a transition (and that has been selected in the process of step S17) (step S18). If it is determined that there is a multi-valued variable that has accepted a transition, the process of step S19 is performed. If it is determined that there is no multi-valued variable that has accepted a transition, the process of step S23 is performed.

[0107] In the process of step S19, a process of reading out the matrix elements of the flow matrix from the F memory 21a, and a process of reading out the matrix elements of the distance matrix from the D memories 21b and 21c are performed. The ΔΔE calculation circuit 22a calculates ΔΔE using each of the read matrix elements (step S20). Then, the ΔE update circuit 22b updates ΔE of the n multi-valued variables using the calculated ΔΔE (step S21).

[0108] The column permutation circuit 22e performs a column permutation process for permuting the columns of the distance matrix stored in the D memories 21b and 21c in response to the transition of the values ​​of the accepted multi-valued variables, for example, as shown in FIG. 6 (step S22).

[0109] The control circuit 22g judges whether or not to end the search (step S23). For example, if the search time reaches the timeout limit, it is judged that the search is ended, and the search ends. If it is judged that the search is not to be ended, the process from step S12 is repeated.

[0110] If it is determined that the search is to be ended, the control circuit 22g outputs the search result (step S24) and ends the process. When performing the simulated annealing method, the control circuit 22g decreases the value of the above-mentioned temperature parameter (T) according to a predetermined temperature parameter change schedule while the processes of steps S12 to S23 are repeated. Then, the control circuit 22g outputs, for example, the values ​​of the n multi-valued variables when a predetermined search end condition is satisfied as a search result for a solution to the assignment problem (for example, displays on a display device not shown). Note that the control circuit 22g may update the value of the evaluation function expressed by, for example, equation (5) every time a transition is accepted, and may hold the minimum value so far and the values ​​of the n multi-valued variables when the minimum value was obtained. In this case, the control circuit 22g may output, as a search result, the values ​​of the n multi-valued variables corresponding to the minimum value stored when the search end condition is satisfied.

[0111] When the control circuit 22g performs the replica exchange method, the control circuit 22g performs the processes of steps S12 to S23 for each of a plurality of replicas to which different temperature parameter values ​​are set. The control circuit 22g performs replica exchange every time the processes of steps S12 to S23 are repeated a predetermined number of times. For example, the control circuit 22g randomly selects two of the plurality of replicas, and exchanges the temperature parameter value or the n multi-valued variable values ​​between the two selected replicas with a predetermined exchange probability based on the difference in the evaluation function value between the replicas or the difference in the temperature parameter value. For example, the control circuit 22g updates the value of the evaluation function expressed by equation (5) every time a transition is accepted in each replica, and holds the minimum value so far and the values ​​of the n multi-valued variables when the minimum value is obtained. Then, the control circuit 22g outputs, as a search result, the values ​​of the n multi-valued variables corresponding to the minimum value among all replicas among the minimum values ​​stored in each replica, for example, when a predetermined search end condition is satisfied.

[0112] According to the data processing device 20 described above, since the complicated data movement used to calculate the amount of change in the evaluation function can be avoided compared to the case where each of n multi-valued variables is converted into a large number of binary state variables of 0 or 1, parallel calculation becomes easier. Therefore, a combinatorial optimization problem expressed by multi-valued variables can be calculated by parallel trials. By enabling parallel trials for combinatorial optimization problems expressed by multi-valued variables, the calculation efficiency of such combinatorial optimization problems is improved.

[0113] Furthermore, the data processing device 20 holds ΔE associated with the transition of each of the n multi-valued variables, and updates ΔE using the calculated δΔE, thereby eliminating the need to recalculate ΔE itself.

[0114] The data processing device 20 also has an F memory 21a that stores a flow matrix. The data processing device 20 also has a D memory 21b that stores a distance matrix including matrix elements arranged based on the current values ​​of n multi-valued variables, and a D memory 21c that stores a distance matrix including matrix elements arranged based on the values ​​of the transition destination candidates of the n multi-valued variables. When calculating δΔE, the matrix elements of the flow matrix specified by the identification number of the multi-valued variable that accepts the transition are read from the F memory 21a. The matrix elements of the distance matrix specified by the values ​​before and after the transition of the multi-valued variable are read from the D memories 21b and 21c. This allows δΔE to be calculated efficiently in parallel for n multi-valued variables.

[0115] (Third embodiment) The data processing device of the third embodiment is different from the data processing device 20 of the second embodiment in that the circuits are made common.

[0116] Fig. 9 is a diagram illustrating an example of a data processing device according to a third embodiment. In Fig. 9, the same reference numerals are used for the elements illustrated in Fig. 3. Note that in Fig. 9, the column switching circuit 22e and the state holding and updating circuit 22f illustrated in Fig. 3 are omitted.

[0117] A storage unit 31 of a data processing device 30 according to the third embodiment includes an F memory 21a, D memories 21b and 21c, as well as a D memory 31a and a memory control circuit 31b. The D memory 31a stores a distance matrix common to each replica. The elements of the distance matrix stored in the D memory 31a are arranged in the order of position identification numbers (1 to 8) in both the row and column directions. When swapping columns of the distance matrix stored in the D memories 21b and 21c, two columns corresponding to the two columns to be swapped are read out from the distance matrix stored in the D memory 31a. Then, one is written into the position of the column to be swapped in the D memory 21b, and the other is written into the column to be swapped in the D memory 21c, thereby performing the swapping.

[0118] In the data processing device 20 of FIG. 3 described above, such a D memory 31a may be used for permuting columns. The memory control circuit 31b controls the write and read operations for the F memory 21a and the D memories 21b, 21c, and 31a under the control of the control circuit 33. In order to calculate δΔE, the memory control circuit 31b reads the above-mentioned five matrix elements included in the formula (7) from the F memory 21a and the D memories 21b and 21c at different read timings using a common read path.

[0119] The δΔE calculation circuit 32 includes registers 32a, 32b, and 32d, and an arithmetic circuit 32c. The register 32a holds the matrix elements (flow amount and distance) read from the F memory 21a and the D memories 21b and 21c. The register 32b holds the matrix elements read from the D memories 21b and 21c.

[0120] The five matrix elements held in the registers 32a and 32b are input to the arithmetic circuit 32c in sequence, and the arithmetic circuit 32c outputs the intermediate calculation result of equation (7) or δΔE based on the matrix elements and the intermediate calculation result held in the register 32d.

[0121] For example, the arithmetic circuit 32c calculates φ j and φi The distance between j and φ i The distance between j and φ' i The distance between j and φ i The calculation in the parentheses on the right side of the formula (7) is performed by sequentially adding or subtracting the distance between the flow amount (f ij ) and the intermediate calculation result held in register 32d, and outputting the product thereof to obtain δΔE.

[0122] The register 32d holds the intermediate calculation result output by the arithmetic circuit 32c or the calculation result δΔE. The control circuit 33 has the same function as the control circuit 22g of the processing unit 22 of the data processing device 20 shown in Fig. 3. Furthermore, the control circuit 33 instructs the memory control circuit 31b on the data read timing and the like to cause the δΔE calculation circuit 32 to calculate δΔE as described above.

[0123] The data processing device 30 may store whether the current value of each multi-valued variable is the original value or the value of the generated transition destination candidate. In this case, the matrix elements of the distance matrix used to calculate δΔE can be appropriately read out without swapping columns between the distance matrices stored in the D memories 21b and 21c. However, when regenerating the transition destination candidates, the control circuit 33 rearranges the matrix elements of the distance matrix stored in the D memory 21b according to the current values ​​of the multi-valued variables.

[0124] Fig. 10 is a timing chart showing an example of the operation of the data processing device according to the third embodiment. In the operation example of Fig. 10, an example of pipeline processing using four replicas is shown, similar to the timing chart shown in Fig. 7.

[0125] Between timing t20 and timing t23, a process of acquiring ΔE is performed in the order of replica 0, replica 1, 2, and 3. Between timing t21 and timing t24, acceptance determination is performed in the above order, starting with replica 0. For each replica, the identification number of the multi-valued variable whose transition has been accepted is output from the transition acceptance determination circuit 22d.

[0126] Between timing t22 and timing t25, a process of reading out matrix elements of the flow matrix from the F memory 21a and a process of reading out matrix elements of the distance matrix from the D memories 21b and 21c are performed. In the data processing device 30 of the third embodiment, the reading out of matrix elements of the distance matrix is ​​performed in four cycles for each replica.

[0127] From timing t23 to t29, the registers 32a and 32b of the ΔΔE calculation circuit 32 hold the matrix elements read from the F memory 21a and the D memories 21b and 21c. From timing t23 to t26, the matrix elements used to calculate ΔΔE for replica 0 are held, from timing t24 to t26, the calculation of ΔΔE is performed, and from timing t26 to t27, the calculation result is finalized. After timing t26, the same process as that for replica 0 is performed for replicas 1, 2, and 3, in that order. Also, from timing t27 onwards, ΔE for each replica is updated each time the calculation result of ΔΔE for each replica is finalized.

[0128] After timing t30, a write process is performed to swap columns including n matrix elements in the distance matrix stored in the D memories 21b and 21c, using the matrix elements of the distance matrix read from the D memory 31a. The write process is performed in parallel for each replica. The n matrix elements included in each column are swapped between two columns one by one, so the swapping is completed in n cycles. As described above, when the data processing device 30 stores whether the current value of each multi-valued variable is the original value or the value of the generated transition destination candidate, it is not necessary to swap columns between the distance matrices stored in the D memories 21b and 21c.

[0129] According to the data processing device 30 of the third embodiment as described above, the δΔE calculation circuit 32 processes five matrix elements in a common arithmetic circuit 32c. This makes it possible to reduce the circuit scale more than the δΔE calculation circuit 22a shown in FIG.

[0130] (Fourth embodiment) In the second and third embodiments, the description has been given assuming that there is one transition destination candidate for the value of each multi-valued variable, but there may be multiple transition destination candidates for the value of each multi-valued variable.

[0131] Fig. 11 is a diagram showing an example of the current state and the states of the transition destination candidates for multiple transition destination candidates. Fig. 11 shows an example of the current state and the states of two transition destination candidates (positions of the destination candidates) in an assignment problem for assigning eight facilities to eight positions. In the example of Fig. 11, the placement of multiple facilities at one position is permitted.

[0132] The state of the first destination candidate is represented by multi-valued variables φ'1 to φ'8. For example, as shown in FIG. 11, the position of the first destination candidate for the facility with facility number = 1 is φ'1 = 4, and the position of the second destination candidate is φ"1 = 6. The position of the first destination candidate for the facility with facility number = 6 is φ'6 = 6, and the position of the second destination candidate is φ"6 = 1.

[0133] In the current state as shown in FIG. 11, when the transition of φ6 to φ″6 is accepted, an example of the matrix elements read out for calculating δΔE is shown below. Regarding the matrix elements to be read from the F memory, similarly to the case where the transition of φ6 to φ'6 shown in FIG. 5 is accepted, the matrix element (f i6 ) is read out.

[0134] FIG. 12 is a diagram showing an example of matrix elements read out from the D memory. When there are a first transition destination candidate state and a second transition destination candidate state, in addition to the D memory 21b corresponding to the current state and the D memory 21c corresponding to the first transition destination candidate state, the D memory 21d corresponding to the second transition destination candidate state is used.

[0135] The matrix elements (distances) of the distance matrices stored in D memories 21b, 21c, and 21d are arranged in the order of location identification numbers (1 to 8) along the columns of the matrix, and in the order of facility numbers along the rows. In the distance matrix stored in D memory 21b, the matrix elements are arranged based on the current values ​​of φ1 to φ8. In the distance matrix stored in D memory 21c, the matrix elements are arranged based on the values ​​of the first transition destination candidates φ'1 to φ'8 of φ1 to φ8. In the distance matrix stored in D memory 21d, the matrix elements are arranged based on the values ​​of the second transition destination candidates φ"1 to φ"8 of φ1 to φ8.

[0136] Multi-valued variable φ j φ” j When j≠i, the difference value of the amount of change accompanying the transition of the value of the multi-valued variable with identification number=i to the first transition destination candidate is expressed by the following formula (13).

[0137]

number

[0138] In addition, when j=i, the difference value of the amount of change accompanying the transition of the value of the multi-valued variable with identification number=i to the first transition destination candidate is expressed by the following formula (14).

[0139]

number

[0140] Also, the multi-valued variable φ j φ” j When j≠i, the difference value of the amount of change accompanying the transition of the value of the multi-valued variable with identification number=i to the second transition destination candidate when j≠i is expressed by the following formula (15).

[0141]

number

[0142] In addition, when j=i, the difference value of the amount of change accompanying the transition of the value of the multi-valued variable with identification number=i to the second transition destination candidate is expressed by the following formula (16).

[0143]

number

[0144] When the transition of φ6 to φ″6 is accepted (when j=6), in order to calculate δΔE in equation (13), the current position (φ6) of the facility with facility number = 6 and the current position (φ i In the example shown in FIG. 11, since φ6=5, the distance matrix φ j Further, the position (φ″6) of the second destination candidate facility with facility number=6 and the current position (φ i In the example shown in FIG. 11, since φ″6=1, the distance between the distance matrix φ j =The matrix elements in the first row are read out.

[0145] On the other hand, the current position (φ6) of the facility with facility number = 6 and the position (φ' i In the example shown in FIG. 11, since φ6=5, the distance matrix φ j Furthermore, the position (φ'6) of the first candidate destination of the facility with facility number = 6 and the position (φ' i In the example shown in FIG. 11, since φ′6=1, the distance matrix φ j =The matrix elements in the first row are read out.

[0146] The distance read from D memory 21b to calculate δΔE in equation (15) is the same as that read from D memory 21b to calculate δΔE in equation (13). On the other hand, to calculate δΔE in equation (15), the current position (φ6) of the facility with facility number = 6 and the position (φ" i In the example shown in FIG. 11, since φ6=5, the distance matrix φ j Furthermore, the position (φ"6) of the second destination candidate facility with facility number = i and the position (φ" i In the example shown in FIG. 11, since φ″6=1, the distance between the distance matrix φ j =The matrix elements in the first row are read out.

[0147] Although the configuration of the data processing device of the fourth embodiment is not shown, a configuration using a plurality of transition destination candidates can be realized by appropriately modifying the configurations shown in Figs. 3 and 9. For example, the D memory 21d shown in Fig. 12 is added to the configurations shown in Figs. 3 and 9. Moreover, ΔΔE in equations (13) to (16) can be calculated by modifying the ΔΔE calculation circuits 22a and 32 shown in Figs. 3 and 9. Furthermore, the ΔE holding circuit 22c holds φ i is φ i ΔE(φ i →φ i ') and φ i is φ i ” ΔE(φ i →φ i The ΔE update circuit 22b uses ΔΔE calculated based on the formulas (13) and (14) to update ΔE(φ i →φ i ') and update ΔE(φ i →φ i The transition acceptance determination circuit 22d updates ΔE(φ i →φ i ') and ΔE(φi →φ i ”), a multi-valued variable that accepts a transition to the first transition destination candidate or the second transition destination candidate is determined.

[0148] It is not necessary to simultaneously determine whether or not to accept a transition to the first transition destination candidate and the second transition destination candidate. When determining whether or not to accept a transition to the first transition destination candidate using the D memory 21c, it is not necessary to determine whether or not to accept a transition to the second transition destination candidate using the D memory 21d. The same is true vice versa. The transition destination candidate may be switched between the first transition destination candidate and the second transition destination candidate as appropriate.

[0149] As described above, the data processing device of the fourth embodiment has D memories 21c and 21d that store distance matrices including matrix elements arranged based on the values ​​of each transition destination candidate. This makes it possible to reduce the processing overhead caused by regenerating transition destination candidates. In addition, for example, while performing a solution search using the D memory 21c, it is also possible to rewrite the distance matrix in the D memory 21d to a distance matrix corresponding to a new transition destination candidate.

[0150] (Fifth embodiment) Fig. 13 is a diagram illustrating an example of a data processing device according to the fifth embodiment. In Fig. 13, the same elements as those illustrated in Fig. 9 are denoted by the same reference numerals.

[0151] δΔE in equation (7) can be transformed into the following equation (17).

[0152]

number

[0153] The right-hand side of equation (17) is φ' j and φ' i and the distance between j and φ i and the difference between the distances j and φ' i and the distance between j and φi This includes the difference between the distance and In the data processing device 40 of the fifth embodiment, the storage unit 41 has a ΔD memory 41a that stores a matrix including matrix elements expressed by the above two differences. The differences are arranged in a matrix so that they can be read by specifying the row or column number. Hereinafter, the matrix based on the above differences is called a differential distance matrix.

[0154] FIG. 14 is a diagram showing an example of matrix elements read out from the ΔD memory. φ of the difference distance matrix stored in the ΔD memory 41a j The row matrix elements are φ j and φ' i and the distance between j and φ i The difference between the distance and the difference matrix φ' j The row matrix elements are φ' j and φ' i and the distance between j and φ i is the difference between the distance

[0155] When the transition of φ6 to φ'6 is accepted, the matrix element in the φ6 row of the differential distance matrix is ​​read from the ΔD memory 41a in order to calculate δΔE. In the example shown in FIG. 14, since φ6=5, the matrix element in the φ6=5th row of the differential distance matrix is ​​read. Furthermore, the matrix element in the φ'6 row of the differential distance matrix is ​​read from the ΔD memory 41a. In the example shown in FIG. 14, since φ'6=6, the matrix element in the φ'6=6th row of the differential distance matrix is ​​read.

[0156] Such a differential distance matrix is ​​generated by differential calculation (Δd calculation) using the distance matrix stored in the D memory 31a. The differential calculation is performed, for example, by the control circuit 33. When a solution search is performed using multiple replicas, a differential distance matrix is ​​generated for each replica.

[0157] The data processing device 40 may store whether the current value of each multi-valued variable is the original value or the value of the generated transition destination candidate. In this case, while the transition destination candidates are constant, the matrix elements used in the Δd calculation can be appropriately read out without rearranging the columns in the distance matrix stored in the D memory 31a. However, when regenerating the transition destination candidates, the control circuit 33 rearranges the matrix elements of the distance matrix according to the current values ​​of the multi-valued variables and the values ​​of the transition destination candidates. Then, the control circuit 33 performs the Δd calculation and initializes the difference distance matrix stored in the ΔD memory 41a.

[0158] The ΔΔE calculation circuit 42 of the data processing device 40 has registers 42a and 42b. The register 42a holds the matrix elements (flow amount and distance difference value (Δd)) read from the F memory 21a and the ΔD memory 41a. The register 42b holds the matrix elements read from the ΔD memory 41a.

[0159] The arithmetic circuit 32c outputs an intermediate calculation result or δΔE of the equation (17) based on the matrix elements or intermediate calculation results held in the registers 42a, 42b, and 32d. For example, the calculation circuit 32c performs the calculation in the parentheses on the right side of the formula (17) by subtracting the difference between the two Δd values ​​held in the registers 42a and 42b. The result of this calculation is held in the register 32d as an intermediate calculation result. Then, the calculation circuit 32c subtracts the flow amount (f ij ) and the intermediate calculation result held in register 32d, and outputting the product thereof to obtain δΔE.

[0160] The other elements of the data processing device 40 are the same as those shown in FIG. Fig. 15 is a timing chart showing an example of the operation of the data processing device according to the fifth embodiment. In the example of the operation of Fig. 15, an example of pipeline processing using four replicas is shown, similar to the timing chart shown in Fig. 7.

[0161] Between timing t40 and timing t43, a process of acquiring ΔE is performed in the order of replica 0, replica 1, 2, and 3. Between timing t41 and timing t44, acceptance determination is performed in the above order, starting with replica 0. For each replica, the identification number of the multi-valued variable whose transition has been accepted is output from the transition acceptance determination circuit 22d.

[0162] Between timing t42 and timing t48, a process of reading out the matrix elements of the flow matrix from the F memory 21a and a process of reading out the matrix elements of the differential distance matrix from the ΔD memory 41a are performed. In the data processing device 40 of the fifth embodiment, the reading out of the matrix elements of the differential distance matrix is ​​performed in two cycles for each replica.

[0163] Between timings t43 and t49, the registers 42a and 42b of the ΔΔE calculation circuit 42 hold the matrix elements read out from the F memory 21a and ΔD memory 41a. Between timings t43 and t46, the matrix elements used to calculate ΔΔE for replica 0 are held, between timings t44 and t46, the calculation of ΔΔE is performed, and between timings t46 and t47, the calculation result is finalized. After timing t46, the same process as that for replica 0 is performed for replicas 1, 2, and 3, in that order. Also, after timing t47, ΔE for each replica is updated each time the calculation result of ΔΔE for each replica is finalized.

[0164] When regenerating transition destination candidates, the control circuit 33 rearranges the matrix elements of the distance matrix in the D memory 31a according to the current values ​​of the multi-valued variables and the values ​​of the transition destination candidates after timing t50. Then, the control circuit 33 performs a write process to perform a Δd calculation and initialize the differential distance matrix stored in the ΔD memory 41a. The write process is performed in parallel for each replica.

[0165] According to the data processing device 40 of the fifth embodiment as described above, the difference in distance used to calculate δΔE is stored in advance in the ΔD memory 41a, thereby making it possible to reduce the circuit scale of the δΔE calculation circuit 42. Also, the time required to calculate δΔE can be shortened.

[0166] The processing contents of each of the above-described embodiments (eg, FIG. 1, FIG. 8, etc.) can be realized in software by causing a data processing device to execute a program. The program may be recorded on a computer-readable recording medium. For example, a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory may be used as the recording medium. Magnetic disks include flexible disks (FDs) and HDDs. Optical disks include compact discs (CDs), recordable CD-Rs (recordable) / rewritable RWs (rewritable), digital versatile discs (DVDs) and DVD-Rs / RWs. The program may be recorded on a portable recording medium and distributed. In that case, the program may be copied from the portable recording medium to another recording medium and executed.

[0167] FIG. 16 is a diagram illustrating an example of hardware of a computer that is an example of a data processing device. The computer 50 includes a CPU 51, a RAM 52, a HDD 53, a GPU 54, an input interface 55, a media reader 56, and a communication interface 57. The above units are connected to a bus.

[0168] The CPU 51 is a processor including an arithmetic circuit that executes program instructions. The CPU 51 loads at least a part of the program and data stored in the HDD 53 into the RAM 52 and executes the program. The CPU 51 may include multiple processor cores, for example, to execute the processing of multiple replicas in parallel. The computer 50 may also include multiple processors. A set of multiple processors (multiprocessor) may be called a "processor."

[0169] The RAM 52 is a volatile semiconductor memory that temporarily stores programs executed by the CPU 51 and data used in calculations by the CPU 51. Note that the computer 50 may include a type of memory other than the RAM 52, or may include multiple memories.

[0170] The HDD 53 is a non-volatile storage device that stores software programs such as an OS (Operating System), middleware, and application software, as well as data. The programs include, for example, a program that causes the computer 50 to execute a process for searching for a solution to the combinatorial optimization problem as described above. The computer 50 may also include other types of storage devices, such as a flash memory or an SSD (Solid State Drive), or may include multiple non-volatile storage devices.

[0171] The GPU 54 outputs an image (for example, an image showing a calculation result of a combinatorial optimization problem) to a display 54a connected to the computer 50 in accordance with an instruction from the CPU 51. As the display 54a, a CRT (Cathode Ray Tube) display, a liquid crystal display (LCD: Liquid Crystal Display), a plasma display (PDP: Plasma Display Panel), an organic EL (OEL: Organic Electro-Luminescence) display, or the like can be used.

[0172] The input interface 55 receives an input signal from an input device 55a connected to the computer 50 and outputs the signal to the CPU 51. The input device 55a may be a pointing device such as a mouse, a touch panel, a touch pad, or a trackball, a keyboard, a remote controller, a button switch, etc. Furthermore, a plurality of types of input devices may be connected to the computer 50.

[0173] The medium reader 56 is a reading device that reads programs and data recorded on a recording medium 56a. For example, a magnetic disk, an optical disk, a magneto-optical disk (MO: Magneto-Optical disk), a semiconductor memory, etc. can be used as the recording medium 56a. Magnetic disks include FDs and HDDs. Optical disks include CDs and DVDs.

[0174] The medium reader 56 copies the program and data read from the recording medium 56a to another recording medium such as the RAM 52 or the HDD 53. The read program is executed by the CPU 51, for example. The recording medium 56a may be a portable recording medium, and may be used for distributing the program and data. The recording medium 56a and the HDD 53 may also be referred to as computer-readable recording media.

[0175] The communication interface 57 is an interface that is connected to a network 57a and communicates with other information processing devices via the network 57a. The communication interface 57 may be a wired communication interface that is connected to a communication device such as a switch by a cable, or may be a wireless communication interface that is connected to a base station by a wireless link.

[0176] (Modification) In the data processing device 20 of the second embodiment described above, the columns or rows of the distance matrix are swapped depending on the current values ​​of the multi-valued variables and the values ​​of the transition destination candidates. In this swapping, the elements of the first identification number=i are swapped with the elements of the second identification number=φ i This can be applied when searching for a solution to the problem of assigning a node to a destination.

[0177] However, if the first identification number = i is assigned to the second identification number = ψ i When a combinatorial optimization problem is formulated as a problem of allocating to the elements of, the roles of the flow matrix and the distance matrix are reversed from the above case. In that case, the columns or rows of the flow matrix are swapped depending on the values ​​of the current multi-valued variables and the values ​​of the transition destination candidates.

[0178] When the columns or rows of the flow matrix are interchanged, for example, processing based on the following equations (18) and (19) is performed instead of equations (11) and (12).

[0179]

number

[0180]

number

[0181] In equation (19), k=ψ j or k=ψ' j It is. The processing based on equations (18) and (19) is i In order to perform parallel processing for (i = 1 to n), the difference value Δf in equation (19) is ik (or identification number = ψ i ,ψ' i Then, the amount of flow between the element with the identification number = k and the element with the identification number = i is stored in a memory with a first identification number i and a second identification number k.

[0182] For the embodiments other than the second embodiment, the roles of the flow matrix and the distance matrix are exchanged, and φ i Instead of ψ i By using the above formula, it is possible to convert the flow matrix into one that corresponds to the interchange of columns or rows.

[0183] While one aspect of the data processing device, the program, and the data processing method of the present invention has been described above based on the embodiment, these are merely examples and the present invention is not limited to the above description. [Explanation of symbols]

[0184] 10 Data Processing Device 11 Storage section 11a Evaluation function information 12 Processing section 12a Multicolor Spin Machine

Claims

1. a storage unit that stores evaluation function information regarding an evaluation function for a combinatorial optimization problem expressed using a plurality of multi-valued variables; a processing unit that generates transition destination candidates that are a part of a transition possible range of values ​​of each of the multiple variables based on the evaluation function information, calculates an amount of change in the value of the evaluation function associated with a transition to the transition destination candidate for each of the multiple multi-valued variables based on the evaluation function information, determines a multi-valued variable that will accept the transition from among the multiple multi-valued variables based on the amount of change, and transitions the value of the determined multi-valued variable to the value of the transition destination candidate; A data processing device having:

2. The processing unit includes: a holding circuit for holding the amount of change accompanying the transition of each of the plurality of multi-valued variables; a calculation circuit for calculating, for each of the plurality of multi-valued variables, a difference value of the amount of change associated with a change in the value of the multi-valued variable that receives the transition; an update circuit that updates the amount of change by using the difference value; 2. The data processing apparatus of claim 1, comprising:

3. the combinatorial optimization problem is an assignment problem expressed using a flow matrix and a distance matrix, The storage unit is a first memory for storing a first matrix, the first matrix being one of the flow matrix and the distance matrix; a second memory for storing a second matrix, the second matrix being the other of the flow matrix and the distance matrix, the second matrix including matrix elements arranged based on values ​​of the plurality of multi-valued variables; a third memory configured to store a third matrix, the third matrix being the other of the flow matrix and the distance matrix, the third matrix including matrix elements arranged based on values ​​of the transition destination candidates of the plurality of multi-valued variables; having the calculation circuit calculates the difference value based on a first matrix element of the first matrix specified by an identification number of the multi-valued variable receiving the transition, a second matrix element of the second matrix and a third matrix element of the third matrix specified by a value before the transition of the multi-valued variable receiving the transition, and a fourth matrix element of the second matrix and a fifth matrix element of the third matrix specified by a value after the transition of the multi-valued variable.

3. A data processing apparatus according to claim 2.

4. one of rows or columns of the second matrix and the third matrix is ​​designated by a value of the multi-valued variable undergoing the transition or a value of the multi-valued variable after the transition; the second matrix element, the third matrix element, the fourth matrix element, or the fifth matrix element included in the designated row or column is read from the second memory or the third memory; 4. A data processing apparatus according to claim 3.

5. the second matrix elements and the third matrix elements are read out in parallel from the second memory and the third memory at a first read timing; the fourth matrix element and the fifth matrix element are read out in parallel from the second memory and the third memory at a second read timing; 4. A data processing apparatus according to claim 3.

6. the first to fifth matrix elements are read from the first memory, the second memory, or the third memory at different read timings using a common read path; 4. The data processing apparatus according to claim 3, wherein said calculation circuit inputs said first to fifth matrix elements to a common arithmetic circuit to calculate said difference value.

7. 4. The data processing apparatus according to claim 3, wherein the processing unit swaps columns or rows of the second matrix and the third matrix, which are specified by the identification number of the multi-valued variable that receives the transition, between the second matrix and the third matrix.

8. for each of the plurality of multi-valued variables, the transition destination candidates include a first transition destination candidate and a second transition destination candidate; the third matrix is ​​arranged based on values ​​of the first transition destination candidates, the storage unit further includes a fourth memory configured to store a fourth matrix, the fourth matrix being the other of the flow matrix and the distance matrix, the fourth matrix including matrix elements arranged based on values ​​of the second transition destination candidates.

4. A data processing apparatus according to claim 3.

9. the combinatorial optimization problem is an assignment problem expressed using a flow matrix and a distance matrix, The storage unit is a first memory for storing a first matrix, the first matrix being one of the flow matrix and the distance matrix; a second memory that stores a fourth matrix including matrix elements expressed by a difference between values ​​of matrix elements of a second matrix, which is the other of the flow matrix and the distance matrix and is arranged based on values ​​of the plurality of multi-valued variables, and values ​​of matrix elements of a third matrix, which is the other of the flow matrix and the distance matrix and is arranged based on values ​​of the transition destination candidates of the plurality of multi-valued variables; having the calculation circuit calculates the difference value based on a first matrix element of the first matrix specified by an identification number of the multi-valued variable receiving the transition, a second matrix element of the fourth matrix specified by a value of the multi-valued variable receiving the transition before the transition, and a third matrix element of the fourth matrix specified by a value of the multi-valued variable after the transition.

3. A data processing apparatus according to claim 2.

10. generating transition destination candidates that are a part of possible transition ranges of the values ​​of each of the multiple multi-valued variables based on evaluation function information related to an evaluation function for a combinatorial optimization problem represented using the multiple multi-valued variables, the evaluation function information being stored in a memory; calculating, for each of the plurality of multi-valued variables, an amount of change in a value of the evaluation function associated with a transition to the transition destination candidate based on the evaluation function information; determining a multi-valued variable that will receive the transition from among the plurality of multi-valued variables based on the amount of change; transitioning the determined value of the multi-valued variable to the transition destination candidate value; A program that causes a computer to carry out processing.

11. a storage unit stores evaluation function information regarding an evaluation function for a combinatorial optimization problem expressed using a plurality of multi-valued variables; a processing unit generating transition destination candidates which are a part of possible transition ranges of the values ​​of each of the plurality of multi-valued variables based on the evaluation function information; the processing unit calculates, for each of the plurality of multi-valued variables, an amount of change in a value of the evaluation function associated with a transition to the transition destination candidate based on the evaluation function information; the processing unit determines, from among the plurality of multi-valued variables, a multi-valued variable that will receive the transition based on the amount of change; the processing unit transitions the determined value of the multi-valued variable to the transition destination candidate value; Data processing methods.

Citation Information

Patent Citations

  • Program, data processing method, and data processing apparatus

    JP2022165250A

  • Program, data processing apparatus, and data processing method

    JP2023001055A