Information processing system, information processing method, and program

The system calculates the expected value of cost term changes to determine constraint coefficients, addressing the inefficiency of conventional methods by reducing the time needed for adjustments and improving the solution finding process in combinatorial optimization problems.

JP2024164601AInactive Publication Date: 2024-11-27FUJITSU LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2023080206
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-15
Publication Date
2024-11-27
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Conventional methods for determining constraint coefficients in combinatorial optimization problems using Ising machines require numerous adjustments and significant time, hindering efficient solution finding.

Method used

An information processing system and method that calculates the expected value of the cost term change during transitions between constraint-satisfying and constraint-violating solutions, allowing for the determination of an appropriate constraint coefficient based on this value, thereby reducing the need for multiple adjustments.

Benefits of technology

This approach significantly reduces the time required to determine constraint coefficients, enhancing the efficiency of solving combinatorial optimization problems by minimizing the occurrence of constraint violations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024164601000001_ABST
    Figure 2024164601000001_ABST
Patent Text Reader

Abstract

To reduce a necessary time for deciding a constraint coefficient.SOLUTION: In an information processing system 10 that calculates a combinatorial optimization problem including a constraint condition based on the energy function of an imaging model represented by the sum of a cost item and a constraint item corresponding to the constraint condition, a processing unit 12 obtains the problem information of the combinatorial optimization problem, calculates, based on the problem information, an expected value of a change quantity of the cost item when a transition from a solution that satisfies the constraint condition to a solution that does not satisfy the constraint condition occurs, decides the coefficient of the constraint item based on the expected value, and outputs, by using the decided coefficient, information on the energy function.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

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

[0002] There is an Ising machine (also called a Boltzmann machine) that calculates multivariate combinatorial optimization problems, which von Neumann computers are not good at, by replacing the problem with an Ising model that represents the behavior of spin in magnetic materials. Methods for solving problems replaced with an Ising model in a practical amount of time include, for example, the Simulated Annealing (SA) method and Markov Chain Monte Carlo (MCMC) methods such as the replica exchange method. Combinatorial optimization problems are formulated by an energy function that includes multiple state variables. For example, an information processing device searches for the ground state of the Ising model that minimizes the value of the energy function by repeatedly trying state transitions by changing the values ​​of the state variables using the MCMC method. The ground state corresponds to the optimal solution of the combinatorial optimization problem.

[0003] The energy function can be expressed as the sum of a cost term and a constraint term (see, for example, Patent Documents 1 and 2). The energy function (E(x)) can be expressed, for example, by the following formula (1).

[0004]

number

[0005] In formula (1), x is x = (x1, x2, …, x n) is a vector of state variables, which are binary variables. C(x) is a cost term, and is the objective function to be minimized. αP(x) is a constraint term. When the constraint conditions that the combinatorial optimization problem must satisfy are satisfied (constraint satisfaction), P(x) = 0, and when the constraint conditions are not satisfied (constraint violation), P(x) > 0. α is the constraint coefficient, and has a value greater than 0. By adding such a constraint term to the cost term, it becomes less likely that a solution that does not satisfy the constraint conditions (constraint violation solution) will be selected.

[0006] In addition, by changing the sign of the energy function, it is possible to search for the state of the Ising model in which the value of the energy function is maximized. The Quadratic Assignment Problem (QAP) is an example of a practical combinatorial optimization problem. The QAP is a problem that seeks an assignment that minimizes the sum of the products of the flow volume between the assignment elements (cost of transporting materials between facilities, etc.) and the distance between the assignment destinations to which each assignment element is assigned, when assigning n assignment elements (facilities, etc.) to n assignment destinations (locations, etc.). The flow volume between the assignment elements and the distance between the assignment destinations can each be expressed using a matrix. In addition, the QAP is subject to the constraint that all assignment elements must be assigned to different assignment destinations.

[0007] Incidentally, if the above constraint coefficient (α in the case of formula (1)) is too small, solutions that violate the constraints are more likely to occur, and conversely, if it is too large, the energy barrier between constraint-satisfying solutions becomes large, hindering the search of the solution space. In addition, the appropriate value of the constraint coefficient varies greatly depending on the individual problem. Conventionally, a method has been proposed in which the constraint coefficient is optimized by repeating the process of adjusting the constraint coefficient based on the search results for solutions to combinatorial optimization problems using an Ising machine. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] International Publication No. 2021-044516 [Patent Document 2] JP 2004-110831 A Summary of the Invention [Problem to be solved by the invention]

[0009] However, in the conventional method, in order to obtain a high-quality solution, it may be necessary to perform multiple adjustments to obtain constraint coefficients that enable a wide range of search, and it takes a long time to determine the constraint coefficients. According to one aspect, the present invention has an object to provide an information processing system, an information processing method, and a program that can reduce the time required to determine a constraint coefficient. [Means for solving the problem]

[0010] In one embodiment, there is provided an information processing system that calculates a combinatorial optimization problem including constraint conditions based on an energy function of an Ising model represented by a sum of a cost term and a constraint term corresponding to the constraint conditions, the information processing system having: a processing unit that acquires problem information of the combinatorial optimization problem, calculates an expected value of an amount of change in the cost term when a transition occurs from a solution that satisfies the constraint conditions to a solution that does not satisfy the constraint conditions based on the problem information, determines a coefficient of the constraint term based on the expected value, and outputs information of the energy function using the determined coefficient; and a search unit that searches for a ground state of the Ising model based on the output information.

[0011] Also, in one embodiment, a method of processing information is provided. Also, in one embodiment, a program is provided. Effect of the Invention

[0012] In one aspect, the present invention can reduce the time required to determine constraint coefficients. [Brief description of the drawings]

[0013] [Figure 1]FIG. 1 illustrates an example of an information processing system according to a first embodiment. [Diagram 2] FIG. 11 illustrates an example of hardware of an information processing system according to a second embodiment. [Diagram 3] FIG. 2 is a diagram illustrating an example of functions of the information processing system. [Figure 4] 13A and 13B are diagrams illustrating an example of the presence or absence of a constraint violation in a QAP and an example of the constraint violation energy. [Diagram 5] FIG. 13 is a diagram illustrating an example of a relationship between a state and a value of an energy function. [Figure 6] 1 is a flowchart illustrating a process flow of an example of an information processing system. [Figure 7] FIG. 13 is a diagram showing an example of matrix expansion. [Figure 8] FIG. 13 is a diagram showing an example of change in element qij selected from Qij in a certain allocation state. [Figure 9] 10 is a flowchart showing a process flow of an information processing system of a comparative example. [Figure 10] FIG. 13 is a diagram illustrating the relationship between the value of a constraint coefficient and a calculation result of an instance in a comparative example. [Figure 11] 11 is a diagram showing the relationship between the value of an adjustment coefficient and the calculation result of an instance in the present embodiment. FIG. [Figure 12] 13 is a flowchart illustrating an example of a process for determining an adjustment coefficient. [Figure 13] FIG. 13 is a diagram showing the relationship between the value of the adjustment coefficient and the calculation results of the same type of instance with different problem sizes. [Figure 14] This is a diagram showing xikxjk to be added to the quadratic term in equation (16) and the linear term in the form of a matrix. [Figure 15] This is a diagram showing xikxil to be added to the quadratic term in equation (17) and the linear term in the form of a matrix. [Figure 16] FIG. 13 is a diagram illustrating an example of a QUBO matrix of constraint terms. [Figure 17] FIG. 13 is a diagram showing an example of a QUBO matrix for the entire energy function. [Figure 18]13 is a flowchart illustrating a process flow of an example of an information processing system according to the third embodiment. [Figure 19] FIG. 13 is a diagram illustrating an example of allocation of elements of a submatrix dklF to a submatrix fijD. [Figure 20] 13 is a flowchart illustrating an example of a process for matrix modulation. [Figure 21] 13 is a diagram showing the relationship between the value of the adjustment coefficient and the calculation result of an instance when matrix modulation is used. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, an embodiment of the invention will be described with reference to the drawings. (Policy for determining constraint coefficients) The energy function of the Ising model can be expressed by the above-mentioned formula (1). Here, the constraint coefficient α should be set to a value such that the constraint violating solution is larger than the constraint satisfying solution in order to prevent the constraint violation. However, the constraint satisfying solution has a range from an optimal solution with the smallest value of C(x) to a solution with a large value of C(x) even if the constraint condition is satisfied, and it is not clear which value is appropriate to use as a reference. In addition, the value of the optimal solution is not clear in the first place. As described above, it has been proposed to optimize the constraint coefficient by repeating the process of adjusting the constraint coefficient based on the search result of the solution of the combinatorial optimization problem by the Ising machine, but the adjustment takes time. Therefore, in this embodiment, α is determined by focusing on the amount of change in the value of C(x) when the constraint satisfying solution transitions to the constraint violating solution.

[0015] In the MCMC method using a temperature parameter, the probability of accepting a transition from a constraint satisfying solution to a constraint violating solution that increases the value of C(x) is low, although this depends on the setting of the temperature parameter value. For this reason, in the following, we consider the case where the value of C(x) decreases due to a transition from a constraint satisfying solution to a constraint violating solution.

[0016] When the value of C(x) decreases due to the transition from a constraint-satisfied solution to a constraint-violated solution, if αP(x) added due to the constraint violation is smaller than the change amount (decrease amount) of C(x), then E(x) in Equation (1) becomes smaller due to that transition. Therefore, the transition to that constraint-violated solution can occur with high probability. Thus, in order to suppress the constraint violation, αP(x) only needs to be a value that compensates at least for the decrease amount of C(x). Next, a method for calculating the expected value of the decrease amount of C(x) will be described.

[0017] (Method for calculating the expected value of the decrease amount of C(x)) In Equation (1), when C(x) is expressed in QUBO (Quadratic Unconstrained Binary Optimization), it becomes as follows in Equation (2).

[0018]

Number

[0019] In Equation (2), Q ij is the coupling coefficient matrix, h i is the bias coefficient vector, and c is the constant term. Q ij If the degree of Q ij is n, then the number of elements q ij (i < j) is n 2 However, since q ij = q ji , q ii = 0 (that is, the diagonal elements are 0), the number of valid q ij is N = n(n - 1) / 2. Hereinafter, the above-mentioned N valid q ij are regarded as the upper triangular elements (excluding the diagonal elements) of the coupling coefficient matrix Q ij , and are denoted as q s (s = 1, …, N). The sum of q s can be expressed by the following Equation (3), and the average value (μ q ) of q s can be expressed by the following Equation (4).

[0020]

Number

[0021]

number

[0022] Also, h i The number of elements of is n, and h i The average value of (μ h ) can be expressed by the following equation (5).

[0023]

number

[0024] Q in Equation (2) ij x i x j x i x j = 1, i.e., x i x j Q if =1 ij The element at the intersection of the i-th column and the j-th row of x in a solution x is selected. i For (i=1,…,n), x i If the number of state variables for which =1 is p, then the selected Q ij The number of elements in is p(p-1) / 2.

[0025] h in Equation (2) i x i x i = 1, i.e., x i If =1, h i The elements of x are selected. i If the number of state variables for which =1 is p, the selected h i The number of elements of is p.

[0026] p x's with value 1 i When one of the switches from 1 to 0 (becomes unselected), Q ij The number of elements that will be unselected is p-1, and the number of elements that will be unselected is h iThe number of elements that will be unselected is 1. ij and i The sum of the non-selected elements is the reduction in C(x).

[0027] The above q s Consider an attempt to extract an element from (s=1,…,N), and the value of the extracted element is X=q1,…,q N is a random variable. However, the above extraction of the p-1 elements that will not be selected is extraction without replacement, and the p-1 elements that will not be selected are Q ij x that exists in the same row or column of and is further inverted i The selection of is not random but a function of x. Therefore, the sequence of random variables {X1,…,X p-1} are not independently and identically distributed.

[0028] However, sampling without replacement can be considered approximately as sampling with replacement when n is large. Also, the inversion of x i In the selection of, the solution by the Ising machine involves a stochastic process. Therefore, if we assume that X is an independent and identically distributed (iid) random variable as an approximation, the random variable sequence {X1,…,X p-1 The expected value of the sum S of} is μ q(1→0) =(p-1)μ q It becomes.

[0029] Similarly, the set of bias coefficients {h s Consider an attempt to extract an element from {s=1,…,N}. The value of the extracted element Y=h1,…,h p Assuming that is a random variable following iid, the expected value of the sum of the random variable sequence {Y1} (which has one element as above) consisting of non-selected elements is μ h(1→0) =μ h It becomes.

[0030] From the above, the expected value of the decrease in C(x) is Δ - C(x)=μ q(1→0) +μ h(1→0) =(p-1)μ q +μ h Such a Δ -If C(x) is used as the constraint coefficient α, when the value of C(x) decreases due to a transition from a constraint satisfying solution to a constraint violating solution, αP(x) added due to the constraint violation can be set to a value sufficient to compensate for the decrease in C(x).

[0031] (Application to QAP) The objective function of QAP is h i Since c is identically 0, C(x) can be expressed by the following equation (6).

[0032]

number

[0033] In QAP, the flow volume between allocation elements and the distance between each allocation destination can be expressed using matrices (see Figure 1 below). C(x) when allocating n allocation elements to n allocation destinations can be expressed by the following formula (7).

[0034]

number

[0035] In equation (7), f ij represents the flow amount of the i-th column and j-th row of the flow matrix F, which represents the flow amount between the assignment elements (the flow amount between the i-th assignment element and the j-th assignment element). d kl represents the distance in the kth column and lth row of the distance matrix D that represents the distance between the assignees (the distance between the kth assignee and the lth assignee). ik represents the allocation of the i-th column and k-th row of the allocation matrix x (see Figure 1 below), which represents the allocation state. If the i-th allocation element is assigned to the k-th allocation destination, then x ik = 1, if the i-th allocation element is not assigned to the k-th allocation destination, x ik =0.

[0036] In addition, P(x) in equation (1) can be expressed as equation (8) in QAP.

[0037]

number

[0038] The first term on the right side is a constraint term for making the number of \(i\) such that \(x ik = 1 for the same \(k\) to be 1, and represents the constraint that one allocation element is assigned to one assignee. The second term on the right side is a constraint term for making the number of \(k\) such that \(x ik = 1 for the same \(i\) to be 1, and represents the constraint that one allocation element is assigned to one assignee. It is possible that \(\alpha\neq\beta\), but in essence, in QAP, since the flow matrix \(F\) and the distance matrix \(D\) are symmetric matrices, \(\beta = \alpha\) can be set, and the constraint term can also be represented as \(\alpha P(x)\) collectively.

[0039] The \(Q\) in Equation (6) ij can be represented by a matrix \(Q\) as shown in the following Equation (9) by the Kronecker product of the flow matrix \(F\) and the distance matrix \(D\). The matrix \(Q\) consists of \(n\times n\) small matrices \(f ij D.

[0040]

Number

[0041] When Equation (9) is expanded, a matrix \(Q\) of degree \(n 2 as shown in Equation (10) is obtained.

[0042]

Number

[0043] Also, the allocation matrix \(x\) can be represented by column vectors with \(n 2 elements as shown in the following Equation (11).

[0044]

Number

[0045] As in Equation (10), the number of elements \(q ij (i < j) of the matrix \(Q\) is \(n4 q ij =q ji , q ii = 0 (i.e., the diagonal elements are 0), so the effective q ij The number of is N=n 2 (n 2 -1) / 2. ij Let be the upper triangular elements (excluding the diagonal elements) of matrix Q, and q s If we write (s=1,…,N), then q s The sum of the above can be expressed as (3), and q s The average value of can be expressed by the above equation (4).

[0046] The expected value of the decrease in C(x) is Δ - C(x)=μ q(1→0) +μ h(1→0) =(p-1)μ q +μ h In the case of QAP, μ h ≡0. Also, in the case of QAP, x is 1 in the constraint satisfaction solution. i The number p of is p=n. Therefore, Δ - C(x) is Δ - C(x)=(p-1)μ q =(n-1)μ q =(n-1)S / N=2S / n 2 The result is (n+1).

[0047] Such a Δ - If C(x) is used as the constraint coefficient α, when the value of C(x) decreases due to a transition from a constraint satisfying solution to a constraint violating solution, αP(x) added due to the constraint violation can be set to a value sufficient to compensate for the decrease in C(x).

[0048] (Δ - Other examples of C(x) As shown in equation (9), the matrix Q is divided into n × n sub-matrices f ij D. The upper triangular and diagonal elements of matrix Q (f ij The number of submatrices that do not include the diagonal element (f ij d klThe number of elements not including (k = l) is n(n - 1). In this case, the effective q ij The number of (i < j) is N’ = n(n - 1)×n(n - 1) / 2 = n 2 (n - 1) 2 / 2. The above-mentioned N’ effective q ij are denoted as q s (s = 1, …, N’), and the sum S’ of q s can be expressed by the following formula (12).

[0049]

Equation

[0050] Here, in the QAP, since the value of the invalid element of the matrix Q is 0, S’ = S. Therefore, the q s average value can be expressed by the following formula (13).

[0051]

Equation

[0052] Therefore, Δ - C(x) is Δ - C(x) = (n - 1)S / N’ = 2(n - 1)S / n 2 (n - 1) 2 = 2S / n 2 (n - 1). Even if such Δ - C(x) is used as the constraint coefficient α, when the value of C(x) decreases due to the transition from the constraint satisfaction solution to the constraint violation solution, αP(x) added due to the constraint violation can be set to a value that compensates for the decrease amount of C(x).

[0053] (First Embodiment) FIG. 1 is a diagram showing an example of the information processing system according to the first embodiment. The information processing system 10 calculates the combinatorial optimization problem based on the energy function of the Ising model as shown in formula (1).

[0054] The information processing system 10 includes a storage unit 11, a processing unit 12, and a search unit 13. The storage unit 11 may be a volatile storage device such as a dynamic random access memory (DRAM) or a non-volatile storage device such as a hard disk drive (HDD) or a flash memory. The storage unit 11 stores various data used in the processing of the processing unit 12. The storage unit 11 also stores problem information of a combinatorial optimization problem. For example, when the optimization problem is a QAP, the problem information includes the flow amount between each allocation element and the distance between each allocation destination.

[0055] The processing unit 12 may include a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc. The processing unit 12 may be a processor that executes a program. The term "processor" may include a collection of multiple processors (multiprocessor).

[0056] The processing unit 12 acquires problem information of the combinatorial optimization problem, formulates the combinatorial optimization problem with an energy function based on the problem information, and outputs information of the energy function to the search unit 13. The processing unit 12 acquires a solution found by the search unit 13, and outputs the acquired solution to the outside. For example, the processing unit 12 may display the contents of the solution on a display device connected to the information processing system 10, or may transmit the solution to another information processing device via a network.

[0057] The search unit 13 receives information on the energy function from the processing unit 12, and searches for a ground state of the Ising model, i.e., a solution that minimizes the value of the energy function, based on the information on the energy function. The search unit 13 outputs a solution obtained by the search to the processing unit 12. The search unit 13 can be realized by dedicated hardware. For example, a search circuit realized by using an integrated circuit such as an FPGA, an ASIC, or a GPU (Graphics Processing Unit) may function as the search unit 13. However, the search unit 13 may also be realized by a processor executing a predetermined program. Search methods used by the search unit 13 include, for example, the SA method, the replica exchange method, and SQA (Simulated Quantum Annealing).

[0058] An example of the operation of the information processing system 10 according to the first embodiment will be described below. When formulating a combinatorial optimization problem using an energy function, the processing unit 12 determines the value of α, which is a coefficient of a constraint term (constraint coefficient).

[0059] First, the processing unit 12 acquires (reads) problem information for a combinatorial optimization problem stored in the storage unit 11. An example of problem information for a QAP is shown in FIG. The problem information of the QAP includes the flow amount between each assignment element and the distance between each assignment destination. The flow amount between each assignment element is represented by a flow matrix F with n rows and n columns, and the distance between each assignment destination is represented by a distance matrix D with n rows and n columns.

[0060] FIG. 1 also shows an example of the allocation matrix x. ik represents the allocation of the i-th column and k-th row of the allocation matrix x, which represents the allocation state. If the i-th allocation element is assigned to the k-th allocation destination, then x ik = 1, if the i-th allocation element is not assigned to the k-th allocation destination, x ik = 0. Note that the assignment matrix x has n elements as shown in the above formula (11). 2 It can also be expressed as a column vector.

[0061] Based on the acquired problem information, the processing unit 12 calculates an expected value of the change in the cost term when a transition occurs from a solution that satisfies the constraint conditions of the combinatorial optimization problem (constraint satisfaction solution) to a solution that does not satisfy the constraint conditions (constraint violation solution).

[0062] When the cost term is expressed as in equation (2), the expected value of the change (decrease) is Δ - C(x)=(p-1)μ q +μ h As mentioned above, p is the sum of n state variables x i For (i=1,…,n), x i = 1. q is the coupling coefficient matrix Q ij N elements q valid in s (s=1,…,N) is the average value of μ h h i (bias coefficient vector), which is the average of n elements.

[0063] The storage unit 11 stores a constraint satisfaction solution as an initial solution, and the processing unit 12 may determine the value of p based on the initial solution, or the value of p itself may be stored in the storage unit 11 in advance. q Or μ h is calculated based on the problem information using the above formulas (3) to (5). When the combinatorial optimization problem is a TSP (Traveling Salesman Problem), etc., the above Δ - C(x)=(p-1)μ q +μ h Therefore, Δ - It is possible to calculate C(x).

[0064] Figure 1 shows the Δ -A calculation example of C(x) is shown. The cost term C(x) of the QAP can be expressed as in the above-mentioned formula (6). The processing unit 12 calculates the matrix Q by the Kronecker product of the flow matrix F and the distance matrix D as shown in the above-mentioned formula (9). Then, the processing unit 12 calculates the average value μ q Based on Δ - Calculate C(x). Note that equations (6), (9), and (10) are also shown in Figure 1.

[0065] In the case of QAP, as mentioned above, we can calculate the average value of the valid elements in the upper triangle (excluding the diagonal elements) of the matrix Q. The number of valid elements is N = n 2 (n 2 -1) / 2, so we divide the sum of the valid element values ​​S by N, S / N=2S / n 2 (n 2 -1) is the average value μ q Furthermore, when the combinatorial optimization problem is QAP, p = n, so Δ - C(x)=2S / n 2 The result is (n+1).

[0066] In addition, the processing unit 12 performs the above-mentioned (Δ - As shown in (other examples of C(x)), Δ - C(x)=2S / n 2 Δ - C(x) may be calculated. In addition, as the problem size (n) of QAP increases, Δ - The n+1 and n-1 in the formula for C(x) are almost equal, and so is n. Therefore, even if you use n instead of n+1 or n-1, the Δ - The value of C(x) is almost the same. Therefore, the above Δ - The formula for calculating C(x) can be changed as appropriate depending on the problem size.

[0067] The processing unit 12 calculates the Δ - The processing unit 12 determines a constraint coefficient α based on C(x). For example, the processing unit 12 determines α=Δ - C(x) or Δ- The value obtained by multiplying C(x) by an adjustment coefficient (described later) may be determined as α.

[0068] Then, the processing unit 12 outputs information on an energy function that formulates the combinatorial optimization problem using the determined α. The output information on the energy function is, for example, matrix data (also called a QUBO matrix) that expresses the energy function of Equation (1) in the QUBO format.

[0069] The search by the search unit 13 is executed, for example, as follows: As an example, consider the case where the SA method or the replica exchange method is used. The search unit 13 calculates the amount of change in the value of the energy function caused by a change in the value of one of the state variables for each of the state variables, and probabilistically accepts the change that reduces the value of the energy function with priority. In this case, in the steepest descent method, if the solution falls into a local solution, it is impossible to escape. Therefore, the search unit 13 uses the Metropolis method or the Gibbs method to determine the transition probability from a certain state of the Ising model to the next state by changing a certain state variable. That is, the search unit 13 probabilistically accepts even a change that increases the value of the energy function according to a comparison between the amount of change in the value of the energy function and the noise value. The noise value is calculated based on the value of the temperature parameter and a random number. The larger the value of the temperature parameter, the larger the amplitude of the noise value. The larger the amplitude of the noise value, the more likely it is that a state transition with a large increase in the value of the energy function is accepted.

[0070] The search unit 13 repeatedly tries state transitions based on such processing from the start of the search until a search termination condition is satisfied, and sets the values ​​of the multiple state variables when the search termination condition is satisfied as a solution. For example, the search termination condition is determined based on whether a predetermined unit time has passed or the number of attempts to change the state variables has reached a predetermined unit number.

[0071] As described above, the processing unit 12 of the information processing system 10 calculates the expected value Δ -The constraint coefficient α is determined based on C(x). As a result, when the value of C(x) decreases due to a transition from a constraint satisfying solution to a constraint violating solution, αP(x) added due to the constraint violation can be set to a value that compensates for the amount of decrease in C(x). This makes it possible to prevent multiple adjustments from being made in order to obtain an appropriate α, and shortens the time required to determine the constraint coefficient.

[0072] (Second embodiment) FIG. 2 illustrates an example of hardware of an information processing system according to the second embodiment. The information processing system 20 includes an information processing device 100 and an optimization device 200 .

[0073] The information processing device 100 formulates the combinatorial optimization problem to generate information of an energy function corresponding to the combinatorial optimization problem, and inputs the generated information of the energy function to the optimization device 200. In the second embodiment, a QAP is handled as an example of a combinatorial optimization problem to be calculated.

[0074] The information processing device 100 includes a CPU 101, a RAM 102, a HDD 103, an IO (Input / Output) interface 104, a GPU 105, an input interface 106, a medium reader 107, and a NIC (Network Interface Card) 108. The CPU 101 is an example of the processing unit 12 of the first embodiment. The RAM 102 is an example of the storage unit 11 of the first embodiment.

[0075] The CPU 101 is a processor that executes program instructions. The CPU 101 loads at least a part of the programs and data stored in the HDD 103 into the RAM 102 and executes the programs. The CPU 101 may include multiple processor cores. The information processing device 100 may have multiple processors. A set of multiple processors may be called a "multiprocessor" or simply a "processor."

[0076] The RAM 102 is a volatile semiconductor memory that temporarily stores programs executed by the CPU 101 and data used in calculations by the CPU 101. The RAM 102 is, for example, a DRAM. Note that the information processing device 100 may include a type of memory other than the RAM 102, or may include multiple memories.

[0077] The HDD 103 is a non-volatile storage device that stores software programs such as an OS (Operating System), middleware, and application software, and data. The information processing device 100 may 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.

[0078] The IO interface 104 is connected to the optimization device 200, and inputs and outputs data to and from the optimization device 200 according to instructions from the CPU 101. For example, the IO interface 104 writes data in the RAM 102 to a register or memory of the optimization device 200, and reads data from the optimization device 200 and writes it to the RAM 102, according to instructions from the CPU 101. As the IO interface 104, for example, a PCI-e (Peripheral Component Interconnect - Express) or the like is used.

[0079] The GPU 105 outputs an image to a display 111 connected to the information processing device 100 in accordance with an instruction from the CPU 101. As the display 111, any type of display can be used, such as a CRT (Cathode Ray Tube) display, a liquid crystal display (LCD: Liquid Crystal Display), a plasma display, or an organic EL (OEL: Organic Electro-Luminescence) display.

[0080] The input interface 106 acquires an input signal from an input device 112 connected to the information processing device 100, and outputs the input signal to the CPU 101. As the input device 112, a pointing device such as a mouse, a touch panel, a touch pad, or a trackball, a keyboard, a remote controller, a button switch, or the like can be used. Furthermore, a plurality of types of input devices may be connected to the information processing device 100.

[0081] The medium reader 107 is a reading device that reads programs and data recorded on the recording medium 113. 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 113. The magnetic disk includes a flexible disk (FD: Flexible Disk) and a HDD. The optical disk includes a compact disc (CD) and a digital versatile disc (DVD).

[0082] For example, the medium reader 107 copies the program or data read from the recording medium 113 to another recording medium such as the RAM 102 or the HDD 103. The read program is executed by the CPU 101, for example. Note that the recording medium 113 may be a portable recording medium, and may be used for distributing the program or data. Also, the recording medium 113 and the HDD 103 may be referred to as computer-readable recording media.

[0083] The NIC 108 is an interface that is connected to the network 30 and communicates with other computers via the network 30. The NIC 108 is connected to a communication device such as a switch or a router via a cable.

[0084] The optimization device 200 is an accelerator that performs a ground state search by hardware using an MCMC method such as the SA method or the replica exchange method based on information of an energy function. The optimization device 200 may be called an Ising machine or a Boltzmann machine. The optimization device 200 has an FPGA 201 and a RAM 202. The FPGA 201 realizes a solution search function for a problem represented by an Ising model. The solution search function of the optimization device 200 may be realized by other types of integrated circuits such as an ASIC or a GPU. The RAM 202 is a memory that stores various data used in the solution search of the FPGA 201. The RAM 202 is, for example, an SRAM (Static RAM).

[0085] However, the optimization device 200 may be hardware that performs a ground state search using the quantum annealing method. Also, instead of the optimization device 200, the CPU 101 or other processor included in the information processing device 100 can execute predetermined software to realize the function of a search unit that uses the SA method, the replica exchange method, the SQA method, or the like.

[0086] FIG. 3 is a diagram illustrating an example of functions of the information processing system. The information processing device 100 includes a storage unit 120, a control unit 130, and a solution output unit 140. The storage unit 120 uses storage areas of a RAM 102 and a HDD 103. The control unit 130 and the solution output unit 140 are realized by a CPU 101 executing a program stored in the RAM 102.

[0087] The optimization device 200 also includes a search unit 210. The search unit 210 is realized by an FPGA 201. The storage unit 120 stores various data used for formulating the QAP by the control unit 130. For example, the storage unit 120 stores information on the flow matrix F and the distance matrix D in the QAP, and information on the matrix Q calculated by the Kronecker product of the flow matrix F and the distance matrix D (see FIG. 1 described above).

[0088] The control unit 130 receives input of the QAP data and formulates the QAP based on the data. The control unit 130 generates information on an energy function that expresses Equation (1) in the QUBO format. The control unit 130 outputs the generated information on the energy function to the optimization device 200 and causes the optimization device 200 to search for a solution based on the energy function.

[0089] The control unit 130 acquires a solution obtained by the search by the optimization device 200, and outputs the acquired solution to the solution output unit 140 if the acquired solution satisfies the constraints and is to be output as a final solution.

[0090] The solution output unit 140 acquires a solution obtained by a search by the optimization device 200, and outputs the acquired solution. For example, the solution output unit 140 may convert the solution acquired from the optimization device 200 into a solution format for the QAP and display it on the display 111, or may transmit information on the solution to another computer via the network 30.

[0091] The search unit 210 acquires information on the energy function output by the control unit 130, and searches for the ground state of the Ising model, i.e., an optimal solution, based on the energy function. The search unit 210 outputs the solution obtained by the search to the information processing device 100.

[0092] (On the relationship between constraint violations in QAP and the energy function) The energy function of the QAP includes a constraint term as shown in formula (8). For example, the constraint condition is "one facility is placed at one position" or "one facility is placed at one position" in a facility placement problem, which is an example of the QAP. Therefore, the search in the search unit 210 is affected by the constraint term.

[0093] FIG. 4 is a diagram showing an example of the presence or absence of a constraint violation in a QAP and an example of the constraint violation energy. Fig. 4 illustrates the constraint terms shown in equation (8). Fig. 4 also illustrates, for example, a portion of the allocation matrix x shown in Fig. 1, and shows the constraint violation energy (constraint violation eg) corresponding to each state in the format of "state:constraint violation eg". The constraint violation eg is a penalty value associated with a constraint violation that is added to the cost term and is calculated based on equation (8). In Fig. 4, a certain state is represented by a number indicating the state.

[0094] For example, state "1" is {x ik}=(x 11 ,x 12 ,x 21 ,x 22 ) = (0,1,1,0). State "1" does not violate the constraint conditions, that is, it is a constraint satisfaction state. Therefore, the constraint violation eg corresponding to state "1" is "0".

[0095] State "2" is {x ik}=(x 11 ,x 12 ,x 21 ,x 22 ) = (0,0,1,0). State "2" violates the constraint. The constraint violation eg corresponding to state "2" is "α + β".

[0096] State "3" is {x ik}=(x 11 ,x 12 ,x 21 ,x 22 ) = (0,0,0,0). State "3" violates the constraint. The constraint violation eg corresponding to state "3" is "2(α + β)".

[0097] State "4" is {x ik}=(x 11 ,x 12 ,x 21 ,x 22 ) = (1,0,0,0). State "4" violates the constraint. The constraint violation eg corresponding to state "4" is "α + β".

[0098] State "5" is {x ik}=(x11 ,x 12 ,x 21 ,x 22 ) = (1,0,0,1). State "5" does not violate the constraint conditions, that is, it is a constraint-satisfied state. Therefore, the constraint violation eg corresponding to state "5" is "0".

[0099] FIG. 5 is a diagram illustrating an example of the relationship between the state and the value of the energy function. In FIG. 5, the state of the Ising model is conveniently shown one-dimensionally. The horizontal axis of FIG. 5 represents the state of the Ising model, and the vertical axis represents energy. Energy represents the value of a cost term or the value of an energy function (including a constraint term). The sequence G1a is a sequence that indicates the relationship between the state and the value of the energy function for a certain problem. The sequence G1b is a sequence that indicates the relationship between the state and the value of the cost term for the problem.

[0100] 4. FIG. 5 also shows the energy difference between series G1a and G1b in states corresponding to states "1" to "5" shown in FIG. The state "1" is a constraint-satisfied state (constraint-satisfied solution), and the constraint violation eg is "0", so the energies of the sequences G1a and G1b match.

[0101] States "2" and "4" are constraint violation states, and since the constraint violation eg is "α+β", the energy of sequence G1a is higher than the energy of sequence G1b by "α+β". State "3" is also a constraint violation state, and since the constraint violation eg is "2(α+β)", the energy of sequence G1a is higher than the energy of sequence G1b by "2(α+β)".

[0102] As shown in Figure 5, in states such as "2" to "4", even though a constraint violation occurs, the value of the cost term represented by the sequence G1b is smaller than that of the constraint satisfying solutions, states "1" and "5". Therefore, in the absence of constraint terms, transitions to states "2" to "4" are more likely to occur.

[0103] For this reason, it is desirable to determine the constraint coefficients α and β of appropriate magnitude. If α and β are too small, solutions that violate the constraints are more likely to occur, and conversely, if they are too large, the energy barrier between solutions that satisfy the constraints becomes large, hindering the search of the solution space.

[0104] For this reason, the control unit 130 of the information processing device 100, like the processing unit 12 of the information processing system 10 of FIG. 1, calculates an expected value Δ - Based on C(x), the constraint coefficient α is determined. Note that β=α.

[0105] In the case of QAP, for example, as mentioned above, α = Δ - C(x)=2S / n 2 By setting it to (n+1), the constraint term added due to a constraint violation can be set to a value that is sufficient to compensate for the amount of decrease in C(x).

[0106] Next, the processing procedure of the information processing system 20 will be described. FIG. 6 is a flowchart illustrating a process flow of an example of the information processing system. The control unit 130 acquires a flow matrix F and a distance matrix D from the QAP data input by the user (step S1).

[0107] The control unit 130 performs matrix expansion (step S2). That is, the control unit 130 calculates the Kronecker product of the flow matrix F and the distance matrix D, and converts the matrix Q ij Generate. FIG. 7 is a diagram showing an example of matrix expansion.

[0108] In Fig. 7, the flow matrix F and the distance matrix D are each 4-by-4 matrices, i.e., Q for n=4. ij An example of the generation of Q is shown. ij is a 16-by-16 matrix. Such a matrix is ​​also called the QUBO matrix of cost terms.

[0109] The allocation elements are represented by A to D, and the allocation destinations are represented by a to d. Also, Fig. 7 shows an example of an allocation matrix x with 4 rows and 4 columns. In the example of Fig. 7, allocation element A is assigned to allocation destination d, allocation element B is assigned to allocation destination a, allocation element C is assigned to allocation destination c, and allocation element D is assigned to allocation destination b.

[0110] The flow matrix F is a matrix in which the flow amounts between the allocation elements A to D are arranged. For example, f AB represents the flow amount between allocation elements A and B. The distance matrix D is a matrix in which the distances between the allocation destinations a to d are arranged. For example, d ab represents the distance between the assignees a and b.

[0111] Matrix Q obtained by the Kronecker product of flow matrix F and distance matrix D ij As shown in Figure 7, Q consists of 4x4 submatrices. The position of each submatrix is ​​represented by (column, row) = (i, j) of matrix Q. The position of an element in the submatrix is ​​represented by (column, row) = (k, l) of matrix Q.

[0112] Matrix Q, which is the product of the flow amount (=0) between identically assigned elements and each distance ij The value of each element in the four diagonal submatrices is 0. In addition, the value of the diagonal elements of each submatrix, which is expressed as the product of the distance between the same allocation destinations (=0) and each flow amount, is also 0.

[0113] Note that the matrix Q ij Since the submatrix included in the lower triangular part of includes the submatrix corresponding to the upper triangular part, the information to be output to search unit 210 needs to be only information about the submatrix corresponding to the upper triangular part (excluding the four diagonal submatrix mentioned above).

[0114] Next, the control unit 130 calculates Δ - Calculate C(x) (step S3). Δ - As described above, C(x) is the expected value of the amount of decrease in C(x) when a transition occurs from a constraint satisfying solution to a constraint violating solution.

[0115] Figure 8 shows the Q ij q elements selected from ij FIG. 13 is a diagram showing an example of a change in When calculating the cost term, the Q corresponding to the state variable with a value of 1 ij Element q at the intersection of row and column of ij In Figure 8, x 13 , x 22 , x 31 , x 44 An example is shown in which the value of is 1. Since there is only one state variable with a value of 1 in each row and column, the constraints of the QAP are satisfied and a constraint satisfying solution is obtained. In this case, the six q ij is selected. Note that here, as mentioned above, Q ij The submatrices corresponding to the upper triangular part of (excluding the four submatrices on the diagonal above) are considered.

[0116] In the example in Figure 8, six q ij The value of C(x) is 140, which is the sum of the values ​​of 12+12+10+8+12+16=70, plus 70 to take into account the lower triangle. where x 31 When the value of q flips (changes from 1 to 0), the number of state variables with a value of 1 in the first row of the assignment matrix x becomes 0. Therefore, the constraints of the QAP are not satisfied, and a solution that violates the constraints is obtained. In this case, the six q ij Of these, three q have values ​​of 12, 8, and 16. ij is not selected. Therefore, the reduction in C(x) is 36.

[0117] The expected value of the decrease in C(x) is Δ - C(x) is, for example, Δ - C(x)=2S / n 2 It can be calculated using the formula (n-1). Next, the control unit 130 determines whether the adjustment coefficient has been determined (step S4). If the control unit 130 determines that the adjustment coefficient has been determined, it performs the process of step S6, and if the control unit 130 determines that the adjustment coefficient has not been determined, it performs the process of step S5.

[0118] In the process of step S5, the control unit 130 determines an adjustment coefficient. The process of step S5 will be described later. After the process of step S5, the control unit 130 performs the process of step S6.

[0119] In the process of step S6, the control unit 130 calculates Δ - The constraint coefficient α is calculated by multiplying C(x) by the adjustment coefficient. After that, the control unit 130 generates information on the energy function in which the formula (1) is expressed in the QUBO format (step S7).

[0120] The control unit 130 sets information about the energy function generated in step S7 in the search unit 210 (step S8). The search unit 210 executes a combinatorial optimization calculation based on the information of the set energy function. The search unit 210 searches for a solution to the QAP through this calculation (step S9).

[0121] The search unit 210 outputs the solution obtained by the search to the control unit 130 (step S10). The control unit 130 acquires the solution from the search unit 210, outputs it to the outside via the solution output unit 140, and ends the process.

[0122] (Comparative Example) FIG. 9 is a flowchart showing a process flow of the information processing system of the comparative example. FIG. 9 shows a processing procedure using a conventional method of optimizing constraint coefficients by repeating the process of adjusting the constraint coefficients based on the search results for a solution to a combinatorial optimization problem using an Ising machine.

[0123] The processes in steps S20 and S21 are the same as those in steps S1 and S2 shown in FIG. In the process of step S22, an initial value (temporary value) of the constraint coefficient α is set. After that, the same processes as steps S7 to S9 shown in Fig. 6 are performed (steps S23, S24, S25), and the assignment matrix x obtained by the combinatorial optimization calculation is output (step S26).

[0124] Then, based on the allocation matrix x, it is determined whether or not the constraint conditions of the QAP are satisfied (step S27). If it is determined that the constraint conditions are satisfied, the process of step S28 is performed, and if it is determined that the constraint conditions are not satisfied, the process of step S30 is performed.

[0125] In the process of step S28, the value of the cost term obtained by the combinatorial optimization calculation is output. Then, it is determined whether the value of the cost term satisfies a predetermined condition (step S29). If it is determined that the condition is satisfied, processing of step S31 is performed. If it is determined that the condition is not satisfied, processing of step S30 is performed.

[0126] In the process of step S30, the constraint coefficient α is updated. For example, if a constraint satisfying solution is not obtained, the constraint coefficient is adjusted to be larger, and if the value of the cost term is larger than a predetermined threshold, the constraint coefficient is adjusted to be smaller. After the process of step S30, the process from step S23 is repeated.

[0127] The process in step S31 is the same as that in step S10 shown in FIG. Next, an example of calculating various instances of the QAP using the method of the comparative example described above will be shown.

[0128] FIG. 10 is a diagram showing the relationship between the value of the constraint coefficient and the calculation result of the instance in the comparative example. FIG. 10 shows an example in which calculations were performed for five instances with a problem size n of n=12 by changing the value of the constraint coefficient using an Ising machine that performs the replica exchange method. The five instances are tai12a, scr12, rou12, nug12, and had12 selected from QAPLIB. The horizontal axis represents the value of the constraint coefficient, and the vertical axis represents the number of replicas for which a constraint satisfying solution was obtained out of 128 replicas.

[0129] When the constraint coefficient is changed from a small value to a large value, in the calculation results for any instance, only constraint violating solutions are obtained in the parts where the constraint coefficient is small, but as the constraint coefficient is increased, constraint satisfying solutions can be obtained. Eventually, constraint satisfying solutions are obtained for almost all replicas. In this way, if the constraint coefficient is made large enough, a constraint satisfying solution can be obtained almost certainly, but generally, the smaller the constraint coefficient is, the closer the cost term value is to the optimal solution (suboptimal solution: a solution whose cost term value is sufficiently close to the optimal solution value) is obtained. Therefore, the appropriate constraint coefficient value is a value where a constraint satisfying solution begins to be obtained, and where most of the solutions are constraint satisfying but some constraint violating solutions remain (the part where the slope of the graph is positive). As shown in Figure 10, even for the same problem size, the appropriate constraint coefficient value varies greatly depending on the instance, from less than one hundred to tens of thousands.

[0130] For example, in the initial state of finding a solution for rou12 (before the relationship shown in FIG. 10 is calculated), the approximate number of the constraint coefficient is unknown, so a sufficiently large value is set as the initial value of the constraint coefficient. If a combinatorial optimization calculation is performed with an appropriate initial value of the constraint coefficient, for example (1) 50,000, a constraint satisfying solution is obtained for all 128 replicas, and it is clear that 50,000 is too large. Next, if the combinatorial optimization calculation is performed with the constraint coefficient value being reduced by half, for example (2) 25,000 (corresponding to the part returning to the processing of step S23 in FIG. 9), all 128 replicas will become constraint violating solutions. Therefore, it is clear that the appropriate range of the constraint coefficient (α) value is 25,000<α<50,000. Next, for example, the constraint coefficient value is set to (3) 40,000 between 25,000 and 50,000, and the combinatorial optimization calculation is performed. However, in this case, all replicas obtain solutions that satisfy the constraints, indicating that this is indeed too large. Next, (4) a combinatorial optimization calculation is performed with 30,000 as the constraint coefficient value. In this case, some replicas obtain solutions that satisfy the constraints, indicating that the constraint coefficients are close to the appropriate value. After this, a combinatorial optimization calculation is performed using several constraint coefficient values, such as values ​​between 30,000 and 40,000 and values ​​slightly smaller than 30,000, and ultimately, for example, 29,500 is determined to be the appropriate constraint coefficient. The above operations can also be incorporated into an algorithm.

[0131] On the other hand, when calculating nug12, etc., if the initial value of the constraint coefficient is 50,000, a solution that violates the constraint will be obtained when the value is changed to, for example, 50,000, 20,000, 10,000, 5,000, 2,000, 1,000, 500, 200, 100, and 50. This results in a large number of wasted trials.

[0132] It is possible to optimize the constraint coefficients using Bayesian estimation instead of manually, but this would still require a large number of trials. In contrast to the comparative example described above, in the information processing system 20 of this embodiment which performs the processing shown in FIG. -By determining the constraint coefficient α based on C(x), the adjustment time for the constraint coefficient α can be shortened. Note that, although Fig. 6 includes step S5 for determining the adjustment coefficient, as described below, the search range for an appropriate adjustment coefficient is limited, so that the increase in calculation time due to the processing of step S5 is smaller than that in the comparative example.

[0133] (Determination of adjustment coefficient) Next, the determination of the adjustment coefficient will be described. The instances (represented using QUBO matrices) solved by the Ising machine have different "statistical properties" (average, variance, matrix density, value distribution, etc.) depending on the problem. - It is desirable to determine the constraint coefficient α by adjusting C(x) to the instance sequence.

[0134] Problems that are formulated as QAP include not only the allocation of factories to factory sites, but also various other problems. For example, the layout of hospital rooms and facilities, the layout of gates at an airport to minimize the total travel distance of transfer passengers, the layout and wiring of printed circuit boards, the key arrangement of keyboards, seating arrangement at a table, and the weight balance of turbine blades. Each of these problems has different "statistical properties."

[0135] QAPLIB is a library that abstracts such problems and has been created as a series of problems of different sizes for each problem, and is used to benchmark solution accuracy and solution speed. 11 is a diagram showing the relationship between the value of the adjustment coefficient and the calculation result of the instance in this embodiment. In FIG. 11, the expected value Δ of the reduction amount of C(x) for five instances with a problem size n of n=12 is calculated by the information processing system 20 of the second embodiment. - After finding C(x), an example is shown in which the value of the adjustment coefficient is changed and calculation is performed using the replica exchange method. The five instances are tai12a, scr12, rou12, nug12, and had12 selected from QAPLIB. The horizontal axis represents the value of the adjustment coefficient, and the vertical axis represents the number of replicas that obtained a constraint satisfying solution out of 128 replicas.

[0136] The variation in the appropriate adjustment coefficient values ​​between instances is significantly smaller than in the comparative example shown in Fig. 10. However, when examined in detail, the statistical properties of each instance are different, and therefore the appropriate adjustment coefficients (the range in which the slope of the graph is positive) differ by several tens of percent. For this reason, in order for the information processing system 20 of the second embodiment to obtain a better solution, it is preferable to determine an appropriate adjustment coefficient for each instance.

[0137] FIG. 12 is a flowchart illustrating an example of the flow of a process for determining an adjustment coefficient. First, the initial value, the increase amount, and the upper limit of the adjustment coefficient are set (step S40). In the process of step S40, the settings may be made based on the initial value, the increase amount, or the upper limit input by the user, or based on the initial value, the increase amount, or the upper limit stored in advance in the storage unit 120.

[0138] Thereafter, the control unit 130 and the search unit 210 perform the processes of steps S41 to S44 similar to the processes of steps S6 to S9 shown in FIG. 6, and the assignment matrix x, which is the search result of the ground state of the Ising model, is output from the search unit 210 (step S45).

[0139] After the process of step S45, control unit 130 determines whether or not the adjustment coefficient has reached an upper limit (step S46). If control unit 130 determines that the adjustment coefficient has not reached the upper limit, it increases the adjustment coefficient by an increment (step S47), and repeats the process from step S41.

[0140] When the control unit 130 determines that the adjustment coefficient has reached the upper limit, it determines an optimal adjustment coefficient based on the allocation matrix x that is the search result obtained up to that point (step S48), and ends the adjustment coefficient determination process. For example, as shown in Fig. 11, the control unit 130 determines the adjustment coefficient based on the allocation matrix x obtained up to that point within a range in which the number of replicas for which a constraint satisfaction solution is obtained increases according to the change in the adjustment coefficient (a range in which the slope is positive), thereby obtaining an appropriate adjustment coefficient.

[0141] The search range for the appropriate adjustment factor is 1×10 0 The range is limited to around 0.05 to 0.1, and the search step (the increase in the adjustment coefficient) is sufficient. The number of adjustments of the adjustment coefficient is determined by the initial value, the increase, and the upper limit set in the process of step S40.

[0142] For this reason, it is expected that the number of times the adjustment coefficient is adjusted can be significantly reduced compared to the number of times the constraint coefficient α is updated in the comparative example. Furthermore, once an adjustment coefficient is determined, that adjustment coefficient can be used when executing a combinatorial optimization calculation of instances of the same type. Therefore, a trial for determining the adjustment coefficient needs to be performed only the first time.

[0143] 13 is a diagram showing the relationship between the value of the adjustment coefficient and the calculation result of the same type of instance with different problem sizes. In FIG. 13, the information processing system 20 of the second embodiment calculates the expected value Δ of the reduction amount of C(x) for 12 taiNa (N is the problem size) with different problem sizes. - This shows an example in which, after finding C(x), calculations were performed using the replica exchange method with different adjustment coefficient values. The horizontal axis represents the adjustment coefficient value, and the vertical axis represents the number of replicas that obtained a constraint satisfying solution out of 128 replicas. In addition, in Fig. 13, the calculation results for each instance are displayed shifted by 150 along the vertical axis.

[0144] In the example of Figure 13, for example, by using an adjustment coefficient (approximately 1.3) that begins to obtain a constraint satisfaction solution for tai10a, the smallest problem size, it is possible to obtain a constraint satisfaction solution for the same type of instance (tai) of 12 different problem sizes.

[0145] In this way, when calculating an adjustment coefficient, the problem size of the problem to be calculated is reduced to search for an appropriate adjustment coefficient, and then that adjustment coefficient is used to solve the actual problem, thereby making it possible to significantly reduce the time required to search for the adjustment coefficient.

[0146] When calculating the adjustment coefficient, it is preferable to use a model that reduces the problem to be optimized (for example, in the case of the allocation of factories to land, the number of factories and the number of land are reduced). However, even in an abstract problem, if there is no statistically significant bias in the flow matrix and the distance matrix, it is possible to calculate the adjustment coefficient using a small-sized flow matrix and distance matrix that are constructed by extracting appropriate rows or columns from each matrix.

[0147] In the above example of determining the adjustment coefficient, a method was used in which an increase amount and an upper limit of the adjustment coefficient were set, and the adjustment coefficient was increased until it reached the upper limit. However, a method in which an increase amount and a lower limit of the adjustment coefficient were set, and the adjustment coefficient was decreased until it reached the lower limit may also be used.

[0148] (QUBO generation example) Next, an example of generating a QUBO based on the determined constraint coefficient α will be described. The constraint term can be expressed, for example, as in the above-mentioned formula (8). The quadratic polynomial multiplied by α, one term on the right side of formula (8), can be expressed as the following formula (14).

[0149]

number

[0150] In equation (14), the quadratic term on the right-hand side can be expressed as equation (15) below.

[0151]

number

[0152] In equation (15), the quadratic term is the sum of the state variables in the range of the off-diagonal elements of the upper triangle (i>j) of the cross product of the state variable vector (see FIG. 14 described later), and doubled for the lower triangle. ik 2 are diagonal elements (see Figure 14 below). x ik is a binary variable that can be 0 or 1, so x ik 2 =x ik From the above, equation (14) can be transformed into the following equation (16).

[0153]

number

[0154] For the same reason as above, the quadratic polynomial multiplied by β in the two terms on the right-hand side of equation (8) can be expressed as the following equation (17).

[0155]

number

[0156] x added to the quadratic term in equation (16) ik x jk When expressed as a matrix, this and the first-order term are as follows: Figure 14 shows the x added to the quadratic term in equation (16). ik x jk 14 is a diagram showing the first and second terms in a matrix. FIG. 14 shows an example in which the problem size n is 4.

[0157] The product of two state variables in region 40 is added to the quadratic term. In Fig. 14, the matrix elements representing the product of two state variables added to the quadratic term are shown with different hatching for each value of k (=1 to 4).

[0158] In FIG. 14, the first-order terms are represented as diagonal elements of a matrix by different hatching for each value of k (=1 to 4). x added to the quadratic term in equation (17) ik x il When expressed as a matrix, this and the first-order term are as follows:

[0159] FIG. 15 shows the x added to the quadratic term in equation (17). ik x il 15 is a diagram showing the first-order term and the second-order term in a matrix. FIG. 15 also shows an example in which the problem size n is 4. The product of two state variables in region 41 is added to the quadratic term. In Fig. 15, the matrix elements representing the product of two state variables added to the quadratic term are represented by different hatching for each value of i (=1 to 4).

[0160] In FIG. 15, the first-order terms are represented as diagonal elements of a matrix by different hatching for each value of i (=1 to 4). FIG. 16 is a diagram illustrating an example of a QUBO matrix of constraint terms.

[0161] Combining the results of Figures 14 and 15, the QUBO matrix for the constraint terms is expressed as shown in Figure 16. The diagonal elements of the QUBO matrix for the constraint terms are -(α+β), and the diagonal elements of each submatrix in the upper triangular part (i>j) are 2α. In addition, each element of the upper triangular part of each of the four diagonal submatrices is 2β. As mentioned above, α=β can be considered.

[0162] FIG. 17 is a diagram showing an example of a QUBO matrix for the entire energy function. The QUBO matrix for the entire energy function in Figure 17 is obtained by adding together the QUBO matrix for the cost terms shown in Figure 7 and the QUBO matrix for the constraint terms shown in Figure 16.

[0163] The control unit 130 outputs such a QUBO matrix as information of the energy function, and sets it in the search unit 210. According to the information processing system 20 of the second embodiment as described above, the same effect as that of the information processing system 10 of the first embodiment can be obtained. That is, it is possible to suppress the occurrence of a large number of adjustments in order to obtain an appropriate constraint coefficient, and it is possible to shorten the time required to determine the constraint coefficient.

[0164] (Third embodiment) FIG. 18 is a flowchart illustrating a process flow of an example of the information processing system according to the third embodiment.

[0165] The processing in steps S50 to S51 is the same as the processing in steps S1 and S2 shown in FIG. After the process of step S51, the control unit 130 performs matrix modulation (step S52). In the process of step S52, Q generated by the Kronecker product of the flow matrix F and the distance matrix D is ij An example of the modulation will be described later.

[0166] After the process of step S52, the control unit 130 ij Based on this, similarly to the process of step S3 shown in FIG. - C(x) is calculated (step S53). In the information processing system according to the third embodiment, the adjustment coefficient can be set to 1.0 for the reasons described below, and Δ - C(x) can be a constraint coefficient α.

[0167] The processing in steps S54 to S57 is similar to the processing in steps S7 to S10 shown in FIG. (Matrix Modulation) C'(x) in the following equation (18) is obtained by adding an arbitrary constant C to the cost term C(x) shown in equation (2).

[0168]

number

[0169] In combinatorial optimization problem calculations, the optimal solution (optimal allocation in the case of QAP) does not change even if the same constant C is always added or subtracted from the cost term, and the value of the cost term in the optimal solution only differs by the amount of the constant C added or subtracted from the cost term. Furthermore, all solutions and their rankings (magnitude relationship) are naturally maintained, not just the optimal solution.

[0170] In QAP, the matrix Q expressed by the Kronecker product of the flow matrix F and the distance matrix D, as shown in the above equation (9), is divided into n × n submatrices f ij D. Given any two allocation elements (e.g., factories) identified by i and j, the possible allocation destinations (e.g., land) are all combinations of the two allocation destinations. ij D represents a cost term corresponding to all combinations of two allocation destinations for any two allocation elements identified by i and j.

[0171] In the constraint satisfaction solution of QAP, each submatrix f ij Any one element f from D other than the diagonal elements ij d kl In other words, the cost terms of the constraint satisfaction solution are chosen based on the respective submatrices f ij Any one element f from D other than the diagonal elements ij d kl Only the value of

[0172] Therefore, for a constraint satisfying solution, all f ij d kl Adding or subtracting the same constant from the cost term will always result in the same optimal solution, as if we simply added or subtracted a constant C from the cost term as above. Also, all the solutions and their rankings are preserved.

[0173] Therefore, in the matrix Q, there are n(n-1) / 2 submatrices f ij For each D, each element f ij d kl Whether or not we add or subtract any uniform constant from the equation, the solution (assignment) and cost term ranking for the constraint satisfaction solution will be the same. ij dkl The operation of adding or subtracting an arbitrary identical constant from corresponds to modulation of the matrix Q.

[0174] On the other hand, if the order of the flow matrix F and the distance matrix D is swapped and the Kronecker product is performed, the resulting small matrix d kl F represents the cost terms corresponding to all combinations of two allocation elements for any two allocation destinations identified by k and l. However, since the Kronecker product is not commutative, the submatrix d kl The elements of F do not form submatrices in the matrix Q shown in equation (9), and each submatrix f ij D is allocated one element at a time.

[0175] Figure 19 shows the submatrix d kl A minor matrix f of elements of F ij FIG. 19 shows an example of allocation to a submatrix d kl d is an example of F 2n A minor matrix f of elements of F ij An example of allocation to D is shown.

[0176] For example, d 2n f is an element of F 12 d 2n is a small matrix f 12 D, and d 2n f element of F 1n d 2n is a small matrix f 1n It is allocated to D. In the constraint satisfaction solution of QAP, the small matrix f ij As in the case of D, the submatrix d kl Any one element f from F other than the diagonal elements ij d kl In other words, the cost term of the constraint satisfaction solution is the small matrix d kl Any one element f other than the diagonal elements of F ij d kl Therefore, for a constraint satisfying solution, the value of all f ij d klAdding or subtracting the same constant from f will result in the same optimal solution as simply adding or subtracting a constant C from the cost term as above. Also, all solutions and their rankings are preserved. Note that the submatrix f ij The elements in the lower triangle of D do not need to be used (they can be invalid elements) because they are the same as the elements in the upper triangle.

[0177] Therefore, the small matrix f ij Similar to D, there are n(n-1) / 2 submatrices d kl For each element f ij d kl Whether or not we add or subtract any uniform constant from , we get the same solution (assignment) and cost term ranking for the constraint satisfying solution. kl Each element f of F ij d kl The operation of adding or subtracting an arbitrary identical constant from also corresponds to modulation of the matrix Q.

[0178] These matrix modulations result in f ij d kl By reducing the deviation of the value of f and smoothing the solution space, the value of the constraint term can be reduced and the efficiency of solution search can be improved. From a statistical point of view, this can be said as follows. ij d kl However, by reducing the deviation using matrix modulation, the instances are transformed into a different instance sequence that has similar statistical properties regardless of the original instance sequence and has the same solution as the original instance.

[0179] In this way, when determining the constraint coefficients, the statistical properties can be made uniform by applying matrix modulation to the instances to be calculated. This makes it possible to make the adjustment coefficients used in the second embodiment to correct statistical differences due to instance sequences almost constant while reducing the sequence dependency of the instances.

[0180] FIG. 20 is a flowchart showing an example of the flow of matrix modulation. The control unit 130 calculates the average value of the elements of the matrix Q obtained by the Kronecker product of the flow matrix F and the distance matrix D (step S60). Also, the control unit 130 calculates the average value of the elements of each small matrix f ij D(i<j) (step S61). Then, for each small matrix f ij D(i<j), the control unit 130 subtracts (the average value of the elements of the small matrix f ij D(i<j) - the average value of the elements of the matrix Q) from each element of the small matrix f ij D(i<j) (step S62). Note that when (the average value of the elements of the small matrix f ij D(i<j) - the average value of the elements of the matrix Q) is negative, the control unit 130 adds this difference value to each element of the small matrix f ij D(i<j).

[0181] Next, the control unit 130 calculates the average value of the elements of each small matrix d kl F(k<l) (step S63). Then, for each small matrix d kl F(k<l), the control unit 130 subtracts (the average value of the elements of the small matrix d kl F(k<l) - the average value of the elements of the matrix Q) from each element of the small matrix d kl F(k<l) (step S64). Note that when (the average value of the elements of the small matrix d kl F(k<l) - the average value of the elements of the matrix Q) is negative, the control unit 130 adds this difference value to each element of the small matrix d kl F(k<l).

[0182] Finally, the control unit 130 subtracts the minimum value in the matrix Q from all the elements of Q (step S65). In the above matrix modulation operation, the process of step S62 corresponds to adding or subtracting a value from each element of each small matrix so that the average value of the elements of each small matrix f ij D(i<j) matches the average value of all the elements of the matrix Q. Also, the process of step S64 corresponds to adding or subtracting a value from each small matrix d kl F(k<l) so that the average value of the elements of each small matrix d kl F(k<l) matches the average value of all the elements of the matrix Q.

[0183] The above processing order is an example, and the processing order may be changed as appropriate. Fig. 21 is a diagram showing the relationship between the value of the adjustment coefficient and the calculation result of the instance when matrix modulation is used. Fig. 21 shows the expected value Δ - After finding C(x), an example is shown in which the value of the adjustment coefficient is changed and calculation is performed using the replica exchange method. The instances to be calculated are tai12a, scr12, rou12, nug12, and had12, selected from QAPLIB. The horizontal axis represents the value of the adjustment coefficient, and the vertical axis represents the number of replicas that obtained a constraint satisfying solution out of 128 replicas.

[0184] When the matrix Q modulated by the above method was used, it was possible to make the adjustment coefficients for five instances almost 1.0, as shown in Fig. 21. In other words, as shown in Fig. 18, it was possible to make the adjustment coefficients Δ - C(x) can be used as the constraint coefficient α as is.

[0185] However, the adjustment coefficient becomes 1.0 only when a matrix modulation operation such as that shown in FIG. 20 is performed; when other matrix modulation operations are performed, the adjustment coefficient may become a value other than 1.0.

[0186] It is also possible to aim for a better value of the cost term by adjusting the constraint coefficient α again before obtaining a solution without adjusting the constraint coefficient α, or after obtaining a solution. The information processing in the first embodiment can be realized by causing the processing unit 12 to execute a program. The information processing in the second and third embodiments can be realized by causing the CPU 101 to execute a program. The program can be recorded in a computer-readable recording medium 113.

[0187] For example, the program can be distributed by distributing recording medium 113 on which the program is recorded. The program may also be stored in another computer and distributed via a network. The computer may store (install) the program recorded in recording medium 113 or a program received from another computer in a storage device such as RAM 102 or HDD 103, and read and execute the program from the storage device.

[0188] While one aspect of the information processing system, information processing method, and program 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]

[0189] 10. Information Processing Systems 11 Storage section 12 Processing section 13 Exploration Department

Claims

1. An information processing system that calculates a combinatorial optimization problem including constraint conditions based on an energy function of an Ising model expressed by a sum of a cost term and a constraint term corresponding to the constraint conditions, a processing unit that acquires problem information of the combinatorial optimization problem, calculates an expected value of an amount of change in the cost term when a transition occurs from a solution that satisfies the constraint condition to a solution that does not satisfy the constraint condition based on the problem information, determines a coefficient of the constraint term based on the expected value, and outputs information of the energy function using the determined coefficient; A search unit that searches for a ground state of the Ising model based on the output information; An information processing system having the above configuration.

2. the combinatorial optimization problem is a quadratic assignment problem; 2. The information processing system according to claim 1, wherein the processing unit calculates a third matrix obtained by a Kronecker product of a first matrix and a second matrix representing the problem information of the quadratic assignment problem, and calculates the expected value based on a first average value of elements included in the third matrix.

3. The information processing system according to claim 1 , wherein the processing unit calculates the coefficient by multiplying the expected value by an adjustment coefficient.

4. The processing unit outputs the information on the energy function formulated using the coefficient obtained by multiplying the expected value by the adjustment coefficient; the search unit outputs a result of the search for the ground state performed based on the information; The processing unit increases or decreases the adjustment factor. repeating the process until the adjustment factor reaches an upper or lower limit; When the adjustment coefficient reaches an upper limit or a lower limit, the processing unit determines the adjustment coefficient based on a result of the search. The information processing system according to claim 3 .

5. the third matrix includes a plurality of submatrices obtained by multiplying elements included in one of the first matrix and the second matrix by the other matrix, the processing unit adds or subtracts a difference between a second average value of elements included in each of the plurality of sub-matrices and the first average value to or from a value of the element included in each of the plurality of sub-matrices, thereby making the first average value equal to the second average value. The information processing system according to claim 2 .

6. 1. An information processing method for calculating a combinatorial optimization problem including constraint conditions based on an energy function of an Ising model expressed by a sum of a cost term and a constraint term corresponding to the constraint conditions, comprising: a processing unit acquires problem information of the combinatorial optimization problem, calculates an expected value of a change in the cost term when a transition occurs from a solution that satisfies the constraint condition to a solution that does not satisfy the constraint condition based on the problem information, determines a coefficient of the constraint term based on the expected value, and outputs information of the energy function using the determined coefficient; A search unit searches for a ground state of the Ising model based on the output information. Information processing methods.

7. A computer used in an information processing system that calculates a combinatorial optimization problem including constraint conditions based on an energy function of an Ising model expressed as a sum of a cost term and a constraint term corresponding to the constraint conditions, acquiring problem information of the combinatorial optimization problem, and calculating, based on the problem information, an expected value of a change in the cost term when a transition occurs from a solution that satisfies the constraint condition to a solution that does not satisfy the constraint condition; determining a coefficient of the constraint term based on the expected value; outputting information of the energy function using the determined coefficient to a search unit that searches for a ground state of the Ising model based on the information; A program that executes a process.

Citation Information

Patent Citations

  • Method and device for adaptively determining weighting coefficient in relation to target function

    JP2004110831A

  • Calculation device, calculation method, and non-transitory computer-readable medium storing program

    WO2021044516A1