Patent Application
20220335487
This application is based upon and claims the benefit of priority of the prior Japanese Patent Application No. 2021-070530, filed on Apr. 19, 2021, the entire contents of which are incorporated herein by reference.
The embodiments discussed herein are related to a non-transitory computer-readable recording medium, a data processing method, and a data processing apparatus.
The quadratic assignment problem (QAP) is one of the combinatorial optimization problems. The quadratic assignment problem is a problem for seeking for an assignment for assigning n elements (such as facilities) to n assignment locations with the aim of minimizing the total sum of products of costs between the elements (each being, for example, a flow amount such as an amount of supplies transported between two facilities) multiplied by the corresponding distances between the assignment locations to which the respective elements are assigned. For example, the quadratic assignment problem is a problem for searching for an assignment that satisfies the following Equation (1).
In Equation (1), fi,j denotes a cost between elements with identification numbers=i and j, dφ(i),φ(j) denotes a distance between assignment locations to which the elements with the identification numbers=i and j are assigned, and Sn denotes a set of n assignment locations.
As an apparatus for calculating a large scale discrete optimization problem which is not well calculated by a von Neumann architecture computer, there is an Ising apparatus (also referred to as a Boltzmann machine) using an Ising type evaluation function (also referred to as an energy function or the like).
The Ising apparatus converts a combinatorial optimization problem into an Ising model expressing behaviors of spins of magnetic elements. Based on the Markov chain Monte Carlo method such as a simulated annealing method or a replica exchange method (also referred to as a parallel tempering method or the like), the Ising apparatus searches for the state of the Ising model that minimizes the value of an Ising type evaluation function (equivalent to energy). The optimal solution is the state having the lowest minimum value among the minimum values of the evaluation function. By changing the sign of the evaluation function, the Ising apparatus is able to search for a state that maximizes the value of the evaluation function. The state of the Ising model may be expressed by a combination of the values of multiple state variables. Each of the state variables may use a value of 0 or 1.
The Ising type evaluation function is defined by, for example, the following Equation (2).
The first term on the right side is the total sum of products each obtained from values (each being 0 or 1) of two state variables and a weight value (indicating an intensity of interaction between the two state variables) in one of all the combinations, without omission and duplication, of all the state variables in the Ising model. Here, xi is a state variable with an identification number i, xj is a state variable with an identification number j, and Wij is a weight value indicating the intensity of interaction between the state variables with the identification numbers i and j. The second term on the right side is the total sum of products each obtained from a bias coefficient and a state variable for one of the identification numbers. Here, bi denotes a bias coefficient for the identification number=i. Then, c is a constant.
An energy change amount (ΔEi) due to a change in the value of xi is expressed by the following Equation (3).
In Equation (3), Δxi is −1 when the state variable xi changes from 1 to 0, whereas Δxi is 1 when the state variable xi changes from 0 to 1. Then, hi is referred to as a local field and ΔEi is the product of hi and a sign (+1 or −1) depending on Δxi.
For example, when ΔEi is smaller than a noise value (also referred to as thermal noise) obtained based on a random number and the value of a temperature parameter, a process of updating the value of xi to generate a state transition and updating the local fields is iterated.
Also, the above-described quadratic assignment problem may be calculated by using an Ising type evaluation function.
An Ising type evaluation function of the quadratic assignment problem may be expressed by the following Equation (4).
[Equation 4]
E=½xTWx (4)
In Equation (4), x is a vector of state variables and expresses assignment states of n elements to n assignment locations. Here, xT is expressed as (x1,1, x1,n, x2,1, x2,n, xn,1, . . . , xn,n) Here, xi,j=1 indicates that an element with an identification number=i is assigned to an assignment location with an identification number=j, and xi,j=0 indicates that the element with the identification number=i is not assigned to the assignment location with the identification number=j.
Then, W is a matrix of weight values, and may be expressed by the following Equation (5) by using a cost (fi,j) and a matrix D of distances between the n assignment locations as described above.
Japanese Laid-open Patent Publication Nos. 2019-185602 and 2020-194442 are disclosed as related art.
According to an aspect of the embodiments, a non-transitory computer-readable recording medium storing a program that causes a processor included in a computer which searches for a combination of n elements and n assignment locations to which the n elements are assigned, where n is an integer of 2 or more, to execute a process includes performing a processing that includes: specifying a first element and a second element among the n elements, reading n costs for each of the first element and the second element and n distances for each of a first assignment location of the first element and a second assignment location of the second element from a memory that stores costs between the n elements, distances between the n assignment locations, and assignment states of the n elements to the n assignment locations, calculating four local fields indicating a first change amount in a value of an evaluation function that is an Ising type evaluation function representing energy depending on the assignment states and including n2 state variables indicating the assignment states due to changes in values of four state variables in which the values change due to an exchange of the assignment locations between the first element and the second element among the n2 state variables, by using a total sum of products of the n costs and the respective n distances, obtaining a second change amount in the value of the evaluation function in a case where the exchange is made, based on the four local fields, a first cost between the first element and the second element among the n costs, and a first distance between the first assignment location and the second assignment location among the n distances, and determining whether or not to execute the exchange based on the second change amount and a predetermined value; search for the assignment states that minimize or maximize the value of the evaluation function by iterating the processing while changing an element pair serving as the first element and the second element.
The object and advantages of the invention will be realized and attained by means of the elements and combinations particularly pointed out in the claims.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory and are not restrictive of the invention.
According to the method of calculating a quadratic assignment problem using an Ising type evaluation function, the processes of calculating an energy change amount, updating the local fields, and so on are iterated many times. Therefore, it is desirable that a storage unit that stores information such as the weight values and the local fields to be used for these processes be accessible at high speed. However, in many cases, such a storage unit has a relatively small capacity. For this reason, when the scale of a problem is large, there is a possibility that the problem may not be calculated because such storage unit is incapable of storing all the information to be used for the calculation.
According to one aspect, an object of the present disclosure is to provide a program, a data processing method, and a data processing apparatus capable of calculating a large scale problem even when a storage unit having a relatively small capacity is used.
Hereinafter, the embodiments of the present disclosure will be described with reference to the drawings.
The quadratic assignment problem has a constraint that all elements are assigned to different assignment locations. This constraint may be considered as a constraint in which, when n2 state variables indicating assignment states of n elements to n assignment locations in an Ising type evaluation function indicating energy depending on the assignment states are arranged in a matrix with n rows and n columns, the sum of the values of the state variables included in each of the rows and the columns is 1.
In the method of updating the state variables one by one based on ΔEi in Equation (3), the calculation of the quadratic assignment problem having this constraint also involves a transition to states that do not satisfy the above constraint, and requests a long calculation time.
A conceivable way to shorten the calculation time is to exclude a transition to assignment states other than assignment states satisfying the above constraint. In this case, the values of four state variables are changed in one state transition. When n2 state variables are arranged in n rows and n columns, the values of four state variables are changed in one state transition so as to satisfy the constraint that the sum of the values of the state variables included in each of the rows and the columns is 1. This makes it possible to exclude a transition to assignment states other than the assignment states satisfying the above constraint.
When a state variable xb having a value of 0 is an update target candidate, state variables xa and xd each having a value of 1 are update target candidates among the state variables included in the same row and the same column as those of xb. In addition, xc having a value of 0 in the same column as that of xa and the same row as that of xd is an update target candidate.
When ΔEb denotes an energy change amount of the Ising model due to changes in the values of these four state variables, ΔEb is expressed by the following Equation (6). Here, ha denotes an energy change amount due to a change in xa, hb denotes an energy change amount due to a change in xb, he denotes an energy change amount due to a change in xc, and hd denotes an energy change amount due to a change in xd.
[Equation 6]
ΔEb=(ha+hd)−(hb+hc)−(Wad+Wbc) (6)
A local field change amount (Δhm (m=1, 2, . . . , n2)) due to the changes in xa, xb, xc, and xd is expressed by the following Equation (7).
[Equation 7]
Δhm=Wbm+Wcm−(Wam+Wdm) (7)
As presented in Equations (6) and (7), the weight values are used to calculate the energy change amount and update the local fields. Since the weight values are expressed by Equation (5), the storage unit may store the costs between the n elements and the distances between the n assignment locations instead of storing the n2×n2 weight values themselves. An energy change amount due to an exchange of the assignment locations between an element with an identification number=i and an element with an identification number=j is expressed by the following Equation (8) into which Equation (6) is transformed.
[Equation 8]
ΔE=ha+hd−hb−hc+2·fi,j·dϕ(i)·(j) (8)
In Equation (8), φ(i) is an identification number of an assignment location of the element with the identification number=i, and φ(j) is an identification number of an assignment location of the element with the identification number=j.
A change amount (ΔH) in the local fields (matrix H) in n rows and n columns is expressed by the following Equation (9).
[Equation 9]
ΔH=ΔFΔD=(f,j−fii)(dϕ(i).−dϕ(j),) (9)
In Equation (9), f,j denotes all the costs in a j-th column in costs (matrix F) in n rows and n columns, and f,i denotes all the costs in an i-th column in the matrix F. Then, dϕ(i), denotes all the distances in a φ(i)-th row in distances (matrix D) in n rows and n columns, and dφ(j), denotes all the distances in the φ(j)-th row in the matrix D.
When the calculation of the energy change amount and the update of the local fields are performed by using Equations (8) and (9), the storage unit does not have to store the n2×n2 weight values themselves. However, when the scale of a problem is increased, the data volumes of the matrix H, the matrix F, and the matrix D are also increased. For this reason, a storage unit being accessible at high speed but having a relatively small capacity, such as a storage unit including a static random-access memory (SRAM), a flip-flop, or the like, may be incapable of storing all of the matrix H, the matrix F, and the matrix D.
A data processing apparatus and a data processing method according to a first embodiment to be described below are capable of calculating a large scale problem even when a storage unit having a relatively small capacity is used.
A data processing apparatus 10 is, for example, a computer, and includes a storage unit 11 and a processing unit 12.
The storage unit 11 is, for example, a volatile storage device including an electronic circuit such as a dynamic random-access memory (DRAM) or a non-volatile storage device including an electronic circuit such as a hard disk drive (HDD) or a flash memory.
For example, the storage unit 11 stores a program for performing processing to be described later and also stores cost information, distance information, and arrangement information.
The cost information contains costs between n elements (n is an integer of 2 or more), and is expressed by a matrix F with n rows and n columns.
The distance information contains distances between the n assignment locations to which the n elements are assigned, and is expressed by a matrix D with n rows and n columns. Hereinafter, the assignment location is referred to as a location.
For example, the cost information and the distance information are input from outside of the data processing apparatus 10 and stored in the storage unit 11.
The arrangement information indicates the locations to which the n elements are assigned (arranged), and indicates assignment states of the n elements to the n locations. The arrangement information is expressed by a matrix X with n rows and n columns including n2 state variables. The initial values in the arrangement information are input from outside of the data processing apparatus 10 and stored in the storage unit 11, for example. The initial values in the arrangement information are set so as to satisfy the constraint that the sum of the values of the state variables included in each of the rows and the columns is 1. For example, this is intended to arrange each of the n elements at any one of the n locations.
In the example illustrated in
The arrangement information may be expressed by a row vector (one-dimensional array) including a sequence of location identification numbers (the column numbers in the example in
For example, the processing unit 12 may be implemented by a processor that is a piece of hardware such as a central processing unit (CPU), a graphics processing unit (GPU) or a digital signal processor (DSP). Instead, the processing unit 12 may be implemented by an electronic circuit such as an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA). The processing unit 12 executes a program stored in the storage unit 11 to cause the data processing apparatus 10 to perform the following processing. The processing unit 12 may be a set of multiple processors.
For example, the processing unit 12 searches for assignment states that minimize the value (energy) of the evaluation function expressed by Equation (2) by iterating the process of exchanging the arrangement locations between two elements. The optimal solution is the assignment states having the lowest minimum value among the minimum values of the evaluation function. By changing the sign of the evaluation function expressed by Equation (2), the processing unit 12 may search for the assignment states that maximize the value of the evaluation function (in this case, the greatest maximum value indicates the optimal solution).
Step S1: The processing unit 12 specifies two elements (the elements with element identification numbers=i and j in the example illustrated in
Step S2: From the storage unit 11, the processing unit 12 reads n costs for each of the two specified elements and n distances for each of the two arrangement locations where the two elements are arranged. By referring to the arrangement information, the processing unit 12 detects the arrangement locations of the two elements.
For example, in the example illustrated in
Although not illustrated, for the element with the element identification number=j, the n costs belonging to the row (or the column) corresponding to the element identification number=j are read from the matrix F. For the location of the location identification number=φi(j), the n distances belonging to the row (or the column) corresponding to the location identification number=φ(j) are read from the matrix D.
The rows (or columns) to which the costs or distances read as described above belong may also be expressed by using the identification numbers of the state variables whose values change due to an exchange between the two elements.
For example, as illustrated in
When xk (k=1, 2, . . . , n2) denotes one of xa, xb, xc, and xd, a row number A in the matrix F to which n costs (fA,) read for xk belong may be expressed as A=(k−1)/n+1 (digits to the right of the decimal point are discarded). The row number A corresponds to an element identification number indicating a row to which xk belongs in the matrix X.
A row number B in the matrix D to which the n distances (dB,) read for xk belong may be expressed as B=(k−1)% n+1. Here, (k−1)% n is a remainder of (k−1) divided by n. The row number B corresponds to a location identification number indicating a column to which xk belongs in the matrix X.
Step S3: The processing unit 12 calculates four local fields ha, hb, hc, and hd in Equation (8). Here, ha denotes an energy change amount due to a change in the value of xa in
In Equation (10), φ(j) is an array of column numbers in which the state variables each having a value of 1 are located at the respective rows in the matrix X. The reason why the local field may be calculated by such Formula (10) will be described below.
In the quadratic assignment problem, bi in Equation (3) is 0. Accordingly, a local field hk for xk may be expressed by the following Equation (11).
For example, the local field hk for xk may be expressed by a product of a k-th row in the matrix W of the weight values and a state vector including n2 state variables. The state vector is a vector in which each of the rows in the matrix X expressing the arrangement information illustrated in
The weight values to be used to calculate hk are at the k-th row in the matrix W. Thus, in a case where k=4, the weight values at the fourth row as illustrated in
According to the state vector, h4 may be expressed as follows, for example.
As a first example, the state vector is assumed to be xT=(1 0 0 0 1 0 0 0 1). In this case, h4 may be calculated as h4=1×f2,1d1,1+0×f2,1d1,2+0×f2,1d1,3+0×f2,2d1,1+1×f2,2d1,2+0×f2,2d1,3+0×f2,3d1,1+0×f2,3d1,2+1×f2,3d1,3=f2,1d1,1+f2,2d1,2+f2,3d1,3.
As a second example, the state vector is assumed to be xT=(0 1 0 0 0 1 1 0 0). In this case, the h4 may be calculated as h4=0×f2,1d1,1+1×f2,1d1,2+0×f2,1d1,3+0×f2,2d1,1+0×f2,2d1,2+1×f2,2d1,3+1×f2,3d1,1+0×f2,3d1,2+0×f2,3d1,3=f2,1d1,2+f2,2d1,3+f2,3d1,1.
Here, the above φ(j) is φ(j)=[1, 2, 3] in the first example and φ(j)=[2, 3, 1] in the second example, and is equal to the array of the column numbers of the distances that is the multipliers for the costs in the j-th column. For example, in the first example, the multiplier for the cost f2,1 in the first column is d1,1 in the first column, the multiplier for the cost f2,2 in the second column is 42 in the second column, and the multiplier for the cost f2,3 in the third column is d1,3 in the third column. In the second example, the multiplier for the cost f2,1 in the first column is d1,2 in the second column, the multiplier for the cost f2,2 in the second column is d1,3 in the third column, and the multiplier for the cost f2,3 in the third column is d1,1 in the first column.
The use of φ(j) as described above makes it possible to select a distance as a multiplier for each of the costs in the j-th column. From the above, hk may be calculated in accordance with Equation (10).
Here, in the processes at steps S2 and S3, the processing unit 12 may iterate the processing of reading fA,j and dBφ(j) from the storage unit 11 and calculating fA,jdB,φ(j) for a certain j in Equation (10), and then reading fA,j and dB,φ(j) from the storage unit 11 for the next j. In this case, the processing unit 12 calculates, as hk, a total sum of n fA,jdB,φ(j) obtained by n iterations of the processing.
Step S4: In accordance with Equation (8), the processing unit 12 calculates ΔE in a case where an exchange of the arrangement locations is made between two elements with the element identification numbers=i and j. The values calculated in the process at step S3 are used for ha, hb, hc, and hd in Equation (8). The values for fi,j and dφ(i),φ(j) in Equation (8) are included in the n costs and/or the n distances read in the process at step S2.
Step S5: Based on a result of comparison between ΔE and a predetermined value, the processing unit 12 determines whether or not to make an exchange of the arrangement locations between the two elements with the element identification numbers=i and j. For example, the predetermined value is a noise value obtained based on a random number and the value of a temperature parameter.
For example, the processing unit 12 determines to make the exchange when ΔE 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, both inclusive, and the temperature parameter (T).
Step S6: When determining to make the exchange in the process at step S5, the processing unit 12 executes the exchange by updating the arrangement information.
After the process at step S6 or when determining not to make the exchange in the process at step S5, the processing unit 12 returns to the process at step S1, updates the element identification numbers=i and j (changes an element pair for the exchange determination), and iterates the processes from step S2.
The element identification numbers=i and j may be specified randomly or according to a predetermined rule.
In a case of performing the simulated annealing method, the processing unit 12 decreases the value of the aforementioned temperature parameter (T) according to a predetermined schedule for temperature parameter change. The processing unit 12 outputs the arrangement information obtained by a predetermined number of iterations of the processes at steps S1 to S6, as a calculation result of the quadratic assignment problem (for example, displays the arrangement information on a display device not illustrated). The processing unit 12 may update the value of the evaluation function (energy) expressed by Equation (4) every time the arrangement locations of the elements are updated, and hold the energy and the arrangement information in the case where the energy becomes the minimum up to that time. In this case, the processing unit 12 may output, as the calculation result, the arrangement information associated with the minimum energy stored after the predetermined number of iterations of the processes at steps S1 to S6.
In the case of performing the replica exchange method, the processing unit 12 performs the processes at steps S1 to S6 illustrated in
According to the data processing apparatus 10 and the data processing method as described above, the local fields to be used for calculating LIE are not stored in the storage unit 11, but are calculated from the costs and the distances concerning two elements as targets for determining whether or not to exchange the arrangement locations, and the arrangement locations of the two elements. For this reason, the storage unit 11 does not have to store the local fields (also the weight values), and a large scale problem may be calculated even when the storage unit 11 is a small-capacity memory. For example, the storage capacity of the storage unit 11 may be reduced. In a case where the replica exchange method is performed, different local fields are used among the replicas. According to the data processing apparatus 10 and the data processing method described above, the storage unit 11 does not have to store the local fields themselves. For this reason, in the case where the replica exchange method is used, the effect of reducing the data volume stored in the storage unit 11 is higher than that in the case where the simulated annealing method is used.
If the local fields were stored, the n2 local fields would be updated in accordance with Formula (9) every time the arrangement information is updated. In contrast, the data processing apparatus 10 and the data processing method described above do not have to update of the local fields. This may reduce the amount of computation as compared with the case where the local fields are stored.
In a second embodiment to be described below, a quadratic assignment problem for assigning n elements to n assignment locations will be described by taking, as an example, a problem for arranging n facilities at n locations. Hereinafter, the above-described cost will be referred to as a flow amount. The flow amount indicates, for example, an amount of supplies transported between facilities or the like.
In the case of this example, a matrix F includes flow amounts in 3 rows and 3 columns, and a matrix D includes distances in 3 rows and 3 columns. In the example of
In the example of
For such a problem, an arrangement that minimizes the energy is searched for based on an energy change amount (expressed by Equation (8) described above) due to an exchange of the arrangement locations between two facilities, but the data volume to be stored increases as the number of facilities or the number of locations increases.
A data processing apparatus 20 is, for example, a computer, and includes a CPU 21, a random-access memory (RAM) 22, an HDD 23, a GPU 24, an input interface 25, a medium reader 26, and a communication interface 27. The above-described units are coupled to a bus.
The CPU 21 is a processor including an arithmetic circuit that executes program commands. The CPU 21 loads at least a part of a program and data stored in the HDD 23 into the RAM 22 and executes the program. The CPU 21 may include multiple processor cores, or the data processing apparatus 20 may include multiple processors. The processes to be described below may be executed in parallel by using the multiple processors or processor cores. A set of multiple processors (multiprocessor) may be referred to as a “processor”.
The RAM 22 is a volatile semiconductor memory that temporarily stores a program executed by the CPU 21 or data used for computation by the CPU 21. The data processing apparatus 20 may include a memory of a type other than the RAM 22, and may include multiple memories.
The HDD 23 is a non-volatile storage device that stores a software program such as an operating system (OS), middleware, and application software, and data. Examples of the program include a program for causing the data processing apparatus 20 to execute a process of searching for a solution to a quadratic assignment problem. The data processing apparatus 20 may include another type of storage device such as a flash memory or a solid-state drive (SSD), and may include multiple non-volatile storage devices.
The GPU 24 outputs images to a display 24a coupled to the data processing apparatus 20 in accordance with instructions from the CPU 21. As the display 24a, a cathode ray tube (CRT) display, a liquid crystal display (LCD), a plasma display panel (PDP), an organic electro-luminescence (OEL) display, or the like may be used.
The input interface 25 acquires an input signal from an input device 25a coupled to the data processing apparatus 20 and outputs the input signal to the CPU 21. As the input device 25a, a pointing device such as a mouse, a touch panel, a touchpad, and a trackball, as well as a keyboard, a remote controller, a button switch, or the like may be used. Multiple types of input devices may be coupled to the data processing apparatus 20.
The medium reader 26 is a reading device that reads a program or data recorded on a recording medium 26a. As the recording medium 26a, for example, a magnetic disk, an optical disk, a magneto-optical (MO) disk, a semiconductor memory, or the like may be used. The magnetic disks include a flexible disk (FD) and an HDD. The optical disks include a compact disc (CD) and a Digital Versatile Disc (DVD).
For example, the medium reader 26 copies a program or data read from the recording medium 26a to another recording medium such as the RAM 22 and the HDD 23. For example, the read program is executed by the CPU 21. The recording medium 26a may be a portable recording medium, and may be used to distribute a program or data. The recording medium 26a and the HDD 23 may be referred to as computer-readable recording media.
The communication interface 27 is an interface that is connected to a network 27a and that communicates with another information processing apparatus via the network 27a. The communication interface 27 may be a wired communication interface coupled to a communication device such as a switch via a cable, or may be a wireless communication interface coupled to a base station via a wireless link.
The data processing apparatus 20 includes an input unit 30, a control unit 31, a storage unit 32, a search unit 33, and an output unit 34.
For example, the input unit 30, the control unit 31, the search unit 33, and the output unit 34 may be implemented by using program modules executed by the CPU 21. The storage unit 32 may be implemented by using a storage area reserved in the RAM 22 or the HDD 23, for example.
For example, the input unit 30 receives input of problem information (such as flow amounts between facilities and distances between locations) of a quadratic assignment problem, and calculation conditions (such as, in a case where the replica exchange method is performed, for example, the number of replicas and a value of a temperature parameter set for each replica). These pieces of information may be input by a user operating the input device 25a, or may be input via the recording medium 26a or the network 27a.
The control unit 31 controls the units in the data processing apparatus 20 to cause the units to execute processing to be described later.
The storage unit 32 stores various kinds of information such as the problem information, the calculation conditions, each of the values of state variables, and the calculation results of energy.
For example, the search unit 33 searches for arrangement information that minimizes the value (energy) of the evaluation function expressed by Equation (2) by iterating the process of exchanging the arrangement locations between two elements.
For example, the output unit 34 outputs the arrangement information obtained by the search by the search unit 33 as a calculation result. For example, the output unit 34 may output and display the calculation result on the display 24a, may transmit the calculation result to another information processing apparatus via the network 27a, or may store the calculation result in an external storage device.
Next, a processing procedure of the data processing apparatus 20 will be described. A data processing method to which the replica exchange method is applied will be described in the following example.
Step S10: First, the input unit 30 receives input of the problem information and so on for a quadratic assignment problem. For example, the problem information may be input by the user operating the input device 25a, or may be input via the recording medium 26a or the network 27a. The input problem information is stored in the storage unit 32.
Step S11: The control unit 31 performs initialization. For example, the control unit 31 sets the initial values of the state variables so as to satisfy the constraint that the sum of the values of the state variables included in each of the n rows and the n columns is 1. The control unit 31 calculates an initial value of energy in accordance with Equations (4) and (5) based on the initial values of the state variables and the problem information. Further, the control unit 31 sets the initial value of the temperature parameter for each replica. As the value of the temperature parameter, different values are set among the replicas.
The control unit 31 initializes search locations (rA, rB) for specifying two facilities for determining whether or not to exchange the arrangement locations (hereafter referred to as flip determination). For example, the control unit 31 sets rA=0 and rB=1. Each of rA and rB specifies a facility identification number (a row number in the matrix X in the example of
Step S12: The control unit 31 selects one of the multiple replicas.
Step S13: In accordance with Equation (10), the search unit 33 calculates four local fields ha, hb, hc, and hd in a case where an exchange of the arrangement locations is made between the facilities with the facility identification numbers specified by rA and rB.
For example, the search unit 33 calculates ha, hb, hc, and hd by the following process. Hereinafter, state [i] denotes the column number of a column in which the value of the state variable is 1 at the (i+1)-th row (i=0, 1, . . . , n−1) in the matrix X. Then, F[y][x] denotes a flow amount at the (y+1)-th row in the (x+1)-th column in the matrix F, and D[y][x] denotes a distance at the (y+1)-th row in the (x+1)-th column in the matrix D. Here, x, y=0, 1, . . . , n−1. For each of ha, hb, hc, and hd, n flow amounts and n distances to be used for the calculation are read from the storage unit 32.
First, the search unit 33 sets ha=hb=hc=hd=0 and i=0.
By using temporary variables fA, fB, dA, and dB, the search unit 33 sets fA=F[rA][i], fB=f[rB][i], dA=D[state [i]][state [rA]], and dB=D[state [i]][state [rB]].
After that, the search unit 33 adds fA×dB to ha, adds fA×dA to hb, adds fB×dB to hc, and adds fB×dA to hd. The search unit 33 increments i by 1 and iterates the above-described process until i reaches n−1.
Step S14: Based on Equation (8), the search unit 33 calculates the change amount of the value of the evaluation function (the energy change amount (ΔE)) in a case where an exchange of the arrangement locations is made between the facilities with the facility identification numbers specified by rA and rB. The values calculated in the process at step S13 are used for ha, hb, hc, and hd in Equation (8). Then, F[rA][rB] is used for fi,j in Equation (8), and D[state [rA]][state [rB]] is used for dφ(i),φ(j) in Equation (8).
Step S15: Based on a result of comparison between ΔE and the noise value obtained based on the random number and the value of the temperature parameter set for the selected replica, the search unit 33 performs the flip determination as to whether or not to make the exchange of the arrangement locations between the facilities with the facility identification numbers specified by rA and rB.
For example, the search unit 33 determines to make the exchange when ΔE 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, both inclusive, and the temperature parameter (T). The search unit 33 performs a process at step S16 when determining to make the exchange, or performs a process at step S17 when determining not to make the exchange.
Step S16: The search unit 33 updates the arrangement information by exchanging the values of state [rA] and state [rB]. The search unit 33 updates the energy (E) by adding ΔE.
Step S17: The control unit 31 determines whether or not all the replicas have been selected. When determining that all the replicas have not been selected, the control unit 31 returns to the process at step S12, and causes the search unit 33 to perform the processes at steps S13 to S16 for an unselected replica. When determining that all the replicas have been selected, the control unit 31 performs a process at step S18.
Step S18: The control unit 31 updates the search locations. For example, in the case where rB=n−1 and rA=n−2, the control unit 31 initializes rA and rB to rA=0 and rB=1. In the case where rB=n−1 but rA≠n−2, the control unit 31 increments rA by 1 and sets rB=rA+1. In a case where rB≠n−1, the control unit 31 increments rB by 1.
Step S19: The control unit 31 determines whether or not a flip determination count indicates a replica exchange cycle. For example, when a remainder of the flip determination count divided by a value indicating the replica exchange cycle is 0, the control unit 31 determines that the flip determination count indicates the replica exchange cycle.
The control unit 31 performs a process at step S20 when determining that the flip determination count indicates the replica exchange cycle, or performs a process at step S21 when determining that the flip determination count does not indicate the replica exchange cycle.
Step S20: The control unit 31 performs a replica exchange process. For example, the control unit 31 randomly selects two of the multiple replicas, and exchanges the set values of the temperature parameter between the two selected replicas with a predetermined exchange probability based on an energy difference between the replicas or a difference in the value of the temperature parameter between the replicas.
Instead of exchanging the values of the temperature parameter between the two replicas, the replica exchange process may exchange the arrangement information.
Step S21: The control unit 31 determines whether or not the flip determination count reaches a predetermined end count. The control unit 31 performs a process at step S22 when determining that the flip determination count reaches the predetermined end count, or iterates the processes from step S12 when determining that the flip determination count does not reach the predetermined end count.
Step S22: The output unit 34 outputs a calculation result. For example, the output unit 34 outputs, as the calculation result, the arrangement information associated with the lowest minimum energy in all the replicas among the minimum energies stored for the respective replicas. For example, the output unit 34 may output and display the calculation result on the display 24a, may transmit the calculation result to another information processing apparatus via the network 27a, or may store the calculation result in an external storage device.
According to the data processing apparatus 20 and the data processing method as described above, the local fields to be used for calculating ΔE are not stored in the storage unit 32, but are calculated from the flow amounts and the distances concerning the two facilities as the targets for the flip determination and the arrangement locations thereof. Thus, the storage unit 32 does not have to hold the local fields (also the weight values), and a large scale problem may be calculated even when the storage unit 32 is a small-capacity memory.
For example, when n=1024 and the data volume of each of the local fields, the flow amounts, and the distances is 4 bytes, the data volume of the matrix W of the weight values is 1024×1024×1024×1024×4=4 TB. On the other hand, the total data volume of the matrix F of the flow amounts and the matrix D of the distances is 8 MB. Therefore, the data volume to be stored may be reduced by 99.9% as compared with the case where the matrix W of the weight values is stored as it is.
In the case where the matrix H of the local fields, the matrix F, and the matrix D each having n rows and n columns are stored without storing the matrix W, the data volume increases as the number of replicas increases.
For example, when n=1024, the number of replicas=32, and the data volume of each of the local fields, the flow amounts, and the distances is 4 bytes, the data volume to be stored is calculated as the data volume of the matrix H+the data volume of the matrix F+the data volume of the matrix D=34×4 MB=136 MB. On the other hand, the total data volume of the matrix F of the flow amounts and the matrix D of the distances is 8 MB as described above. Therefore, the data volume may be reduced by 94.1% as compared with the case where the matrix H is stored.
According to the data processing apparatus 20 and the data processing method of the second embodiment, the amount of computation may be reduced as compared with the case where the matrix H of the local fields is used. For example, in the case where the matrix H of the local fields is used, the number of multiplications for one flip determination and update of the matrix H along with an update of the arrangement information where n=1024 is calculated as 1024×1024+1=1048577 from Formulas (8) and (9). In contrast, in the data processing apparatus 20 and the data processing method according to the second embodiment, the matrix H is not updated and the number of multiplications for one flip determination where n=1024 is calculated as 1024×4+1=4097 from Formulas (8) and (10). For example, the number of multiplications may be reduced by 99.6% as compared with the case where the matrix H of the local fields is used.
As described above, the above-described processing may be implemented by causing the data processing apparatus 20 to execute a program.
The program may be recorded in a computer-readable recording medium (for example, the recording medium 26a). As the recording medium, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like may be used. The magnetic disks include an FD and an HDD. The optical disks include a CD, a CD-recordable (R)/rewritable (RW), a DVD, and a DVD-R/RW. The program may be recorded in a portable recording medium to be distributed. In this case, the program may be copied from the portable recording medium to another recording medium (for example, the HDD 23) to be executed.
A data processing apparatus 40 according to the third embodiment includes an accelerator card 41 coupled to a bus.
The accelerator card 41 is a hardware accelerator that searches for a solution to a quadratic assignment problem. The accelerator card 41 includes an FPGA 41a and a DRAM 41b.
In the data processing apparatus 40 according to the third embodiment, the FPGA 41a performs, for example, the processes by the control unit 31 and the search unit 33 illustrated in
The DRAM 41b functions as the storage unit 32 illustrated in
There may be provided multiple accelerator cards 41. In this case, for example, the processes (for example, the processes at steps S13 to S16 in
The data processing apparatus 40 according to the third embodiment described above also produces the same effect as that of the data processing apparatus 20 according to the second embodiment.
Although aspects of the program, the data processing method, and the data processing apparatus of the present disclosure have been described above based on the embodiments, the embodiments are merely examples and the present disclosure is not to be limited to the above description.
All examples and conditional language provided herein are intended for the pedagogical purposes of aiding the reader in understanding the invention and the concepts contributed by the inventor to further the art, and are not to be construed as limitations to such specifically recited examples and conditions, nor does the organization of such examples in the specification relate to a showing of the superiority and inferiority of the invention. Although one or more embodiments of the present invention have been described in detail, it should be understood that the various changes, substitutions, and alterations could be made hereto without departing from the spirit and scope of the invention.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2021-070530 | Apr 2021 | JP | national |
