Program, data processing method and data processing device
By calculating the rate of change in evaluation function index values, the method efficiently evaluates parameter values for combinatorial optimization problems, ensuring appropriate parameter selection and enhancing solution performance.
Patent Information
- Application Number
- JP2023100596
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-06-20
- Publication Date
- 2025-01-08
AI Technical Summary
Existing methods for solving combinatorial optimization problems often fail to efficiently evaluate whether parameter values are appropriate for the specific problem at hand, leading to suboptimal solution performance.
A program and data processing method that calculates the rate of change in the evaluation function's index values during a solution search to assess the performance of parameter values, allowing for the efficient evaluation of general-purpose or specialized parameter values based on this rate of change.
Enables efficient evaluation of parameter values, ensuring appropriate parameter selection for combinatorial optimization problems, thereby improving solution performance and reducing unnecessary specialized parameter searches.
Smart Images

Figure 2025001157000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a program, a data processing method, and a data processing device. [Background technology]
[0002] An Ising machine is sometimes used as an example of solving a combinatorial optimization problem using an evaluation function. A combinatorial optimization problem is converted into an energy function of an Ising model, which is a model that represents the behavior of spin in a magnetic material. An Ising machine searches for a combination of state variable values included in the energy function that minimizes the value of the energy function, for example. In this case, the combination of state variable values that minimizes the value of the energy function corresponds to the ground state or the optimal solution.
[0003] Methods for obtaining approximate solutions to combinatorial optimization problems in a practical amount of time include, for example, Simulated Annealing (SA), Genetic Algorithm (GA), Simulated Quantum Annealing (SQA), Tabu Search, etc. Each of these methods uses a specific parameter to perform a solution search.
[0004] For example, an optimization device has been proposed that searches for the ground state while gradually decreasing the value of the temperature parameter in the SA method. Also, an optimization device has been proposed that performs a solution search using the SA method using input parameters including the initial value of thermal noise (temperature) and the decay rate of the thermal noise (temperature).
[0005] There has also been proposed a computing system that generates a set of parameters, including a search period, a range, and a starting point, for a solver that executes a heuristic algorithm to search the solution space of the problem, based on the characteristics of the problem.
[0006] Furthermore, when performing quantum annealing, a computer system has been proposed that uses machine learning techniques to determine a set of annealing parameters for a problem. In addition, an optimization device has been proposed that performs a simulation using a provisional optimal solution to a combinatorial optimization problem converted into an Ising model, updates the Ising model by adding constraint terms based on the simulation results, and repeatedly searches for a provisional optimal solution. [Prior art documents] [Patent documents]
[0007] [Patent Document 1] JP 2020-194273 A [Patent Document 2] Patent Publication No. 2021-43787 [Patent Document 3] US Patent Application Publication No. 2011 / 0060710 [Patent Document 4] US Patent Application Publication No. 2021 / 0241143 [Patent Document 5] US Patent Application Publication No. 2020 / 0380065 Summary of the Invention [Problem to be solved by the invention]
[0008] In solving combinatorial optimization problems, it is important to determine appropriate parameter values to be used for the solution of the problem. Therefore, a problem arises as to how to evaluate whether the parameter values are appropriate for the problem.
[0009] In one aspect, the present invention aims to efficiently perform evaluation of parameter values. [Means for solving the problem]
[0010] In one embodiment, a program is provided. The program causes a computer to execute the following process. When a solution search is performed using a first parameter value based on an evaluation function, the computer calculates a rate of change indicating the degree of difference between the first index value and the second index value based on a first index value of the evaluation function value calculated by the evaluation function and a second index value of the evaluation function value obtained in a first period from a point after the start of the solution search period to the end of the period. The computer evaluates the solution-finding performance by the first parameter value for a problem represented by the evaluation function based on the rate of change.
[0011] Also, in one aspect, a computer-implemented data processing method is provided. Also, in one aspect, a data processing device is provided having a storage unit and a processing unit. Effect of the Invention
[0012] In one aspect, the evaluation of parameter values can be performed efficiently. [Brief description of the drawings]
[0013] [Figure 1] FIG. 1 is a diagram illustrating a data processing device according to a first embodiment. [Diagram 2] FIG. 11 illustrates an example of hardware of a data processing device according to a second embodiment. [Diagram 3] FIG. 2 illustrates an example of functions of a data processing device. [Figure 4] FIG. 13 is a diagram illustrating a first example of the relationship between time and energy in a solution search. [Diagram 5] FIG. 13 is a diagram illustrating a second example of the relationship between time and energy in a solution search. [Figure 6] 11 is a flowchart illustrating an example of processing performed by the data processing device. [Figure 7] 11 is a flowchart showing a comparative example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0014] The present embodiment will be described below with reference to the drawings. [First embodiment] A first embodiment will be described.
[0015] FIG. 1 is a diagram illustrating a data processing device according to a first embodiment. The data processing device 10 is used to solve a combinatorial optimization problem. The data processing device 10 includes a storage unit 11 and a processing unit 12.
[0016] The storage unit 11 may be a volatile semiconductor memory such as a random access memory (RAM), or may be a non-volatile storage such as a hard disk drive (HDD) or a flash memory. The processing unit 12 is a processor such as a central processing unit (CPU), a graphics processing unit (GPU), or a digital signal processor (DSP). The processing unit 12 may include an electronic circuit for a specific purpose such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The processing unit 12 may be a processor that executes a program stored in a memory such as a RAM (which may be the storage unit 11). A set of multiple processors may be called a "multiprocessor" or simply a "processor."
[0017] In the following description, an example will be shown in which the processing unit 12 searches for a solution. However, a search unit realized by an FPGA, a GPU, or the like may search for a solution in response to an instruction from the processing unit 12 and return the search result to the processing unit 12. In that case, the search unit may be included in the data processing device 10, or may be included in another device that communicates with the data processing device 10.
[0018] A combinatorial optimization problem is formulated by a predetermined evaluation function, and is replaced by a problem of minimizing, for example, the value of the evaluation function, that is, the evaluation value. The evaluation function may be called an objective function, or may be called an energy function when the evaluation function represents energy. The evaluation function includes multiple state variables. The state variables are binary variables that take the value of 0 or 1. The state variables may be called bits. A solution to a combinatorial optimization problem is represented by the values of multiple state variables. Here, the energy function of an Ising model is shown as an example of an evaluation function. A solution that minimizes the value of the energy function represents the ground state of the Ising model and corresponds to the optimal solution of the combinatorial optimization problem. The value of the energy function is called energy. Energy is an example of the value of the evaluation function, that is, the evaluation value. For this reason, the "energy function" in the following description can be read as the "evaluation function". Furthermore, the "energy" in the following description can be read as the "value of the evaluation function" or the "evaluation value".
[0019] The Ising-type energy function is expressed by equation (1).
[0020]
number
[0021] The state vector x has multiple state variables as elements and represents the state of the Ising model. Equation (1) is an energy function formulated in the QUBO (Quadratic Unconstrained Binary Optimization) format. In the case of a problem of maximizing energy, the sign of the energy function can be reversed.
[0022] The first term on the right side of equation (1) is the sum of the values of the two state variables and the weighting coefficients for all combinations of two state variables that can be selected from all state variables, without omissions or duplications. The subscripts i and j are the indices of the state variables. i is the i-th state variable. x j is the jth state variable. W ijis the weighting coefficient indicating the strength of the connection between the i-th state variable and the j-th state variable. W ij =W ji And W ii =0.
[0023] The second term on the right-hand side of equation (1) is the sum of the products of the biases of all state variables and the values of the state variables. i indicates the bias for the i-th state variable. Problem information including the weighting coefficient and bias included in the energy function is stored in the storage unit 11. State variable x i The change in energy due to the change in i is expressed by equation (2).
[0024]
number
[0025] ΔE i >0 state variable x i When the value of the energy function E(x) changes, the energy E(x) in formula (1) decreases. The processing unit 12 calculates, for each of the state variables, the amount of change in the value of the energy function caused by a change in the value of one of the state variables, and probabilistically accepts the change that reduces the value of the energy function with priority.
[0026] In this case, if the steepest descent method is used, it is impossible to escape if the method falls into a local solution. Therefore, the processing unit 12 uses the Metropolis method or the Gibbs method to determine the transition probability from one state to the next state by changing a certain state variable. That is, the processing unit 12 probabilistically allows changes that increase the value of the energy function according to a comparison between the amount of change in the value of the energy function and the thermal noise value. The thermal noise value is calculated based on a temperature value or a random number. The larger the temperature value, the larger the amplitude of the thermal noise value. The larger the amplitude of the thermal noise value, the more likely it is that a state transition with a large increase in the value of the energy function will be allowed.
[0027] For example, in the SA method, the processing unit 12 gradually changes the temperature value from the maximum temperature value to the minimum temperature value to reduce the amplitude of the thermal noise value, and causes the state of the Ising model to converge to the ground state. In this way, parameters such as the maximum temperature value and the minimum temperature value are used in the SA method.
[0028] For this reason, the processing unit 12 acquires in advance generic parameter values, i.e. generic parameter values, based on the solution results of the SA method or the like for a plurality of existing learning problems, and uses the generic parameter values to solve new problems. The learning problems are problems prepared for learning generic parameter values. In acquiring the generic parameter values, the processing unit 12 executes, for example, the SA method or the like using each of a plurality of candidate values of a parameter for a predetermined period of time, and evaluates the candidate values based on the energy obtained as a result of the execution. Then, the processing unit 12 selects the most highly evaluated candidate value as the generic parameter value, and stores it in the storage unit 11. However, generic parameter values are not necessarily useful for all problems.
[0029] Therefore, the processing unit 12 evaluates the solution-finding performance by the SA method or the like when using generic parameter values for a new problem, i.e., an evaluation problem. The processing unit 12 determines whether or not to acquire parameter values specialized for the evaluation problem, i.e., specialized parameter values, based on the evaluation result. Specifically, the processing unit 12 executes the following process. Hereinafter, generic parameter values are exemplified as parameter values to be evaluated for solution-finding performance, but the parameter values to be evaluated are not limited to generic parameter values and can be any parameter values.
[0030] First, when a solution search based on an energy function corresponding to an evaluation problem is executed using the generic parameter values, the processing unit 12 acquires history information of the energy obtained by the solution search. Graph 20 is an example of graphing history information of the energy, which is the value of the evaluation function obtained by the solution search. The horizontal axis of graph 20 is time. The vertical axis of graph 20 is the value of the evaluation function, which is energy in this example. The horizontal axis of graph 20 indicates the start time and the end time of the solution search.
[0031] The processing unit 12 also obtains a first index value of energy calculated by the energy function of the evaluation problem. Furthermore, the processing unit 12 obtains a second index value of energy obtained in a first period from a point after the start of the solution search period to the end of the period. For example, the first period is a period from time t to the end of the period. The first period is, for example, the latter n% of the entire period from the start to the end of the solution search.
[0032] For example, the processing unit 12 sets the median value of the energy obtained in the first period as the second index value. However, the processing unit 12 may set another statistical value, such as the average value or the minimum value of the energy obtained in the first period, as the second index value.
[0033] Examples of the first index value are as follows: In a first example, the processing unit 12 acquires, as the first index value, the median value of the energy obtained in a second period before the first period, starting from the start of the solution search. For example, the second period is the period from the start of the solution search to time τ. Time τ is a time before time t. The second period is, for example, the first half m% of the entire period of the solution search. m≦n. This is because the energy fluctuates more in the first half of the solution search than in the second half.
[0034] In a second example, the processing unit 12 uses, as the first index value, the energy of an initial solution used in a solution search based on the energy function of the evaluation problem. In a third example, the processing unit 12 uses the energy of a predetermined solution calculated in advance for the energy function of the evaluation problem as the first index value. For example, the energy of the predetermined solution is the energy of an optimal solution predicted in theory. However, the energy of the predetermined solution may be the energy of any solution other than the optimal solution, such as a local solution.
[0035] The processing unit 12 calculates a rate of change indicating the degree of difference between the first index value and the second index value based on the first index value and the second index value. For example, the processing unit 12 may calculate the ratio of the difference between the first index value and the second index value to the first index value as the rate of change. More specifically, the processing unit 12 may calculate the rate of change by dividing the difference between the first index value and the second index value by the first index value. The difference between the first index value and the second index value may be the absolute value of the value obtained by subtracting the second index value from the first index value. In addition, when the energy of a predetermined solution is used as the first index value, the rate of change may be referred to as a difference rate.
[0036] Then, the processing unit 12 evaluates the solution-finding performance of the generic parameter values for the evaluation problem based on the rate of change. For example, the processing unit 12 evaluates the solution-finding performance in accordance with a comparison between the rate of change and a predetermined reference value. Specifically, the processing unit 12 evaluates the solution-finding performance as insufficient when the rate of change is less than the reference value. The processing unit 12 evaluates the solution-finding performance as sufficient when the rate of change is equal to or greater than the reference value.
[0037] Insufficient solution-finding performance means that the generic parameter values are not appropriate for the evaluation problem. Therefore, when the solution-finding performance is insufficient, the processing unit 12 determines to acquire specialized parameter values specialized for the evaluation problem. Sufficient solution-finding performance means that the generic parameter values are appropriate for the evaluation problem. Therefore, when the solution-finding performance is sufficient, the processing unit 12 determines not to acquire the specialized parameter values.
[0038] Here, the reference value is determined in advance. The reference value may be specified in advance by a user. Alternatively, the processing unit 12 may acquire a history of energy obtained by a solution search performed to acquire a generic parameter value for each of a plurality of learning problems, and calculate the reference value based on the acquired energy history. More specifically, the processing unit 12 may obtain a change rate similar to that described above for each learning problem based on the history of energy obtained by a solution search for each learning problem with the generic parameter value, and may set the median of the change rates for each learning problem as the reference value. Furthermore, when comparing with the change rate, the processing unit 12 may multiply the reference value by a variable d (d<1) to set the level of performance to be tolerated.
[0039] When the processing unit 12 determines not to acquire the specialized parameter values, it outputs the first solution obtained by the solution search using the general-purpose parameter values, and the process of finding a solution for the evaluation problem is then completed. On the other hand, when the processing unit 12 determines to acquire the specialized parameter value, it performs a solution search using each of the multiple parameter candidate values, and acquires the specialized parameter value based on the execution result of the solution search for each of the multiple parameter candidate values.
[0040] The time required for solution search using each of the multiple parameter candidate values may be shorter than the time required for solution search using the specialized parameter value. For example, the processing unit 12 may select, as the specialized parameter value, the parameter candidate value that reaches the lowest energy as a result of a solution search using each of the multiple parameter candidate values for a certain period of time. Alternatively, the processing unit 12 may evaluate the parameter candidate value higher the shorter the time required to reach the lower energy, and select the most highly evaluated parameter candidate value as the specialized parameter value.
[0041] The processing unit 12 executes a solution search using the specialized parameter values, and outputs the second solution obtained by the solution search using the specialized parameter values, and then the solution of the evaluation problem is completed. In this way, according to the data processing device 10, a solution search is performed using a first parameter value based on the evaluation function. A first index value of the evaluation function value calculated by the evaluation function is acquired. A second index value of the evaluation function value obtained in a first period from a point after the start of the solution search period to the end of the period is acquired. A rate of change indicating the degree of difference between the first index value and the second index value is calculated based on the first index value and the second index value. The solution-finding performance using the first parameter value for the problem represented by the evaluation function is evaluated based on the rate of change.
[0042] This allows the data processing device 10 to efficiently evaluate the parameter values. The processing unit 12 may output an evaluation result of the solution performance according to the first parameter value for the problem. For example, this allows the data processing device 10 to assist the user in checking whether the first parameter value is appropriate for the problem. The above-mentioned generic parameter value is an example of the first parameter value.
[0043] Here, in solving a combinatorial optimization problem, a general-purpose parameter value, i.e., a general-purpose parameter value, may be acquired in advance based on the result of executing a solution search for an existing problem, and the general-purpose parameter value may be used to solve a new problem. However, the general-purpose parameter value is not necessarily useful for all problems. Depending on the problem, the general-purpose parameter value may not provide sufficient solution performance. On the other hand, it is also possible to acquire a parameter value appropriate for each problem. However, in acquiring the parameter value, for example, a process such as executing a solution search for a predetermined time using each of a plurality of candidate values of the parameter and evaluating each candidate value based on the execution result is performed. For this reason, it is inefficient to acquire a parameter value for the problem every time a new problem appears.
[0044] Therefore, the data processing device 10 can make it possible to efficiently find a solution as follows by using a general-purpose parameter value as the first parameter value to be evaluated. For example, if the data processing device 10 evaluates, based on the rate of change, that the solution-finding performance when using generic parameter values is sufficient, the solution searched for using the generic parameter values can be directly adopted as the solution to the problem in question, and the acquisition of specialized parameter values can be omitted.
[0045] Furthermore, the data processing device 10 can narrow down the problems for which specialized parameter values are to be acquired to those for which the solution-finding performance is evaluated to be insufficient when generic parameter values are used. This eliminates the need for the data processing device 10 to acquire specialized parameter values for each problem. Furthermore, the data processing device 10 can acquire specialized parameter values when the solution-finding performance is insufficient with generic parameter values, thereby enabling a solution search using the specialized parameter values instead of generic parameter values, and improving the solution-finding performance for the problem in question.
[0046] [Second embodiment] Next, a second embodiment will be described. FIG. 2 illustrates an example of hardware of a data processing device according to the second embodiment.
[0047] The data processing device 100 solves a combinatorial optimization problem. The combinatorial optimization problem is formulated by the energy function of equation (1). The data processing device 100 searches for a solution corresponding to the combinatorial optimization problem by executing an SA method or the like based on the energy function.
[0048] The data processing device 100 has a processor 101, a RAM 102, a HDD 103, a GPU 104, an input interface 105, a media reader 106, a communication interface 107, and an accelerator card 108. These units of the data processing device 100 are connected to a bus inside the data processing device 100. The processor 101 corresponds to the processing unit 12 of the first embodiment. The RAM 102 or the HDD 103 corresponds to the storage unit 11 of the first embodiment.
[0049] The processor 101 is an arithmetic device that executes program instructions. The processor 101 is, for example, a CPU. The processor 101 loads at least a part of the program and data stored in the HDD 103 into the RAM 102 and executes the program. The processor 101 may include multiple processor cores. The data processing device 100 may have multiple processors. The processing described below may be executed in parallel using multiple processors or processor cores. A set of multiple processors may also be called a "multiprocessor" or simply a "processor."
[0050] The RAM 102 is a volatile semiconductor memory that temporarily stores programs executed by the processor 101 and data used in calculations by the processor 101. Note that the data processing device 100 may include a type of memory other than a RAM, or may include multiple memories.
[0051] 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 data 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.
[0052] The GPU 104 outputs an image to a display 51 connected to the data processing device 100 in accordance with an instruction from the processor 101. As the display 51, 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.
[0053] The input interface 105 acquires an input signal from an input device 52 connected to the data processing device 100, and outputs the signal to the processor 101. As the input device 52, 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. In addition, a plurality of types of input devices may be connected to the data processing device 100.
[0054] The medium reader 106 is a reading device that reads programs and data recorded on the recording medium 53. 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 53. 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).
[0055] The medium reader 106 copies, for example, the program or data read from the recording medium 53 to another recording medium such as the RAM 102 or the HDD 103. The read program is executed, for example, by the processor 101. Note that the recording medium 53 may be a portable recording medium, and may be used for distributing the program or data. The recording medium 53 and the HDD 103 may also be referred to as computer-readable recording media.
[0056] The communication interface 107 is connected to the network 54 and communicates with other information processing devices via the network 54. The communication interface 107 may be a wired communication interface connected to a wired communication device such as a switch or a router, or a wireless communication interface connected to a wireless communication device such as a base station or an access point.
[0057] The accelerator card 108 is a hardware accelerator that executes the SA method based on the energy function of the Ising model. The accelerator card 108 may search for a solution using an algorithm other than the SA method. The accelerator card 108 has a processor 108a and a RAM 108b. The processor 108a is, for example, a CPU, a GPU, an FPGA, an ASIC, or the like. The processor 108a executes the SA method according to an instruction from the processor 101, and returns a solution obtained by the search. The RAM 108b is a memory that stores data used in the processing of the processor 108a. For example, parameter values such as a maximum temperature value and a minimum temperature value are used to execute the SA method. The parameter values are input to the accelerator card 108 by the CPU 101 and are held in the RAM 108b.
[0058] An apparatus that solves a problem expressed by an Ising-type energy function, that is, an Ising problem, such as the accelerator card 108 or the data processing apparatus 100, is called an Ising machine.
[0059] FIG. 3 is a diagram illustrating an example of functions of the data processing device. The data processing device 100 has a memory unit 110, a learning problem input unit 120, a general-purpose parameter search unit 130, a reference value calculation unit 135, an evaluation problem input unit 140, a difficulty calculation unit 145, solution search units 150, 180, a general-purpose parameter performance evaluation unit 160, a specialized parameter search unit 170 and a solution output unit 190.
[0060] The storage unit 110 uses the storage areas of the RAM 102 and the HDD 103. The learning problem input unit 120, the generic parameter search unit 130, the reference value calculation unit 135, the evaluation problem input unit 140, the difficulty calculation unit 145, the solution search units 150 and 180, the generic parameter performance evaluation unit 160, the specialized parameter search unit 170, and the solution output unit 190 are realized by the processor 101 executing the programs stored in the RAM 102.
[0061] The storage unit 110 stores data on learning problems and evaluation problems. Learning problems are combinatorial optimization problems used for general-purpose parameter search. A plurality of learning problems are prepared in advance. Evaluation problems are combinatorial optimization problems that the user wants to solve.
[0062] The learning problem input unit 120 acquires data of the learning problem stored in the memory unit 110, and inputs it to the general-purpose parameter search unit 130 and the reference value calculation unit 135. The learning problem data may include information such as an Ising-type energy function indicating the learning problem and an initial solution used in the solution search.
[0063] The generic parameter search unit 130 performs generic parameter search. The generic parameter search is a process of acquiring generic parameter values using a learning problem. In the SA method, the generic parameter values to be acquired are, for example, both or either of the maximum temperature value and the minimum temperature value. When there are multiple types of parameters, such as the maximum temperature and the minimum temperature, the generic parameter values to be acquired are a set of the values of these parameters. The same applies to the specialized parameter values shown below.
[0064] First, the general parameter search unit 130 acquires multiple candidate values for the parameter. For example, the general parameter search unit 130 may acquire candidate values from a predetermined candidate value range in a brute force manner by a grid search. Alternatively, the general parameter search unit 130 may acquire multiple candidate values randomly from a predetermined candidate value range in a random search.
[0065] The generic parameter search unit 130 inputs the parameter candidate values and the learning problem data into the accelerator card 108, executes a solution search for a certain period of time, and obtains an update history of the energy in the solution search. For a plurality of learning problems, the generic parameter search unit 130 obtains the energy reached for each candidate value and the time it took to reach that energy, and evaluates each candidate value so that the lower the energy reached in the shorter time, the higher the evaluation. The generic parameter search unit 130 then selects the most highly evaluated candidate value as the generic parameter value, and stores it in the memory unit 110.
[0066] The reference value calculation unit 135 calculates a reference value based on the energy update history in the solution search acquired in the generic parameter search. The reference value calculation is a process of obtaining a reference value for measuring the solution performance of the solution search when the generic parameter value is used for the evaluation problem. The reference value calculation unit 135 inputs the calculated reference value to the generic parameter performance evaluation unit 160.
[0067] The evaluation problem input unit 140 acquires the evaluation problem data and the generic parameter values stored in the storage unit 110, and inputs them to the solution search unit 150. The evaluation problem data may include information such as an Ising-type energy function indicating the evaluation problem and an initial solution used in the solution search. In addition, the evaluation problem input unit 140 inputs information on the number of state variables (number of bits) in the evaluation problem, the type of constraint, and the number of constraints to the difficulty calculation unit 145. Furthermore, the evaluation problem input unit 140 inputs the energy of a theoretical solution to the evaluation problem to the generic parameter performance evaluation unit 160.
[0068] Furthermore, when it is determined based on the performance evaluation by the generalized parameter performance evaluation unit 160 that a specialized parameter search is to be performed, the evaluation problem input unit 140 inputs the evaluation problem data to the specialized parameter search unit 170 and the solution search unit 180 .
[0069] The difficulty level calculation unit 145 calculates the difficulty level of an evaluation problem based on the number of bits in the evaluation problem, the type of constraint, and the number of constraints. Here, the constraints are expressed as constraint terms in an Ising-type energy function. The number of constraints corresponds to the number of constraint terms that represent the corresponding type of constraint. The types of constraints include, for example, 1way1hot constraints and 2way1hot constraints.
[0070] For example, the difficulty level calculation unit 145 increases the difficulty level of an evaluation question as the number of bits or the number of constraints increases. As an example, when there are two types of constraints, constraint 1 and constraint 2, the difficulty level calculation unit 145 may calculate the difficulty level of a question by the following formula (3).
[0071] Difficulty of problem = b1 × (number of bits of evaluation problem) + b2 × (number of constraints 1) + b3 × (number of constraints 2) (3) Alternatively, the difficulty level calculation unit 145 may calculate the difficulty level of a question using the following formula (4).
[0072] Difficulty of problem = b × (number of bits of evaluation problem) + (number of constraints 1) × (number of constraints 2) (4) Here, b1, b2, and b3 in equation (3) and b in equation (4) are coefficients given in advance. b1, b2, b3, and b are positive real numbers.
[0073] Alternatively, a table showing the correspondence relationship of difficulty for a set of the number of bits of a question, the number of constraints 1, and the number of constraints 2 may be stored in advance in the storage unit 110. For example, the table holds a pattern such that if the number of bits is c1 or less, the number of constraints 1 is c2 or less, and the number of constraints 2 is c3 or less, the difficulty of the question is v. In this case, the difficulty calculation unit 145 can search the table for a pattern that meets the conditions of the number of bits of the evaluation question, the number of constraints 1, and the number of constraints 2, and obtain the difficulty of the evaluation question.
[0074] Then, the difficulty level calculation unit 145 inputs the difficulty level of the evaluation question to the general-purpose parameter performance evaluation unit 160. The solution search unit 150 inputs the data of the evaluation problem and the generic parameter values to the accelerator card 108, and executes a solution search for the evaluation problem. The solution search unit 150 acquires the energy update history in the solution search and the searched solution from the accelerator card 108. The solution search unit 150 outputs the energy update history to the generic parameter performance evaluation unit 160.
[0075] The generic parameter performance evaluation unit 160 performs generic parameter performance evaluation. The generic parameter performance evaluation is a process for evaluating the solution-finding performance of the solution search when generic parameter values are used. First, the generic parameter performance evaluation unit 160 acquires a first index value of the energy calculated by the energy function of the evaluation problem. In addition, the generic parameter performance evaluation unit 160 acquires a second index value of the energy obtained in a first period from a point after the start point of the solution search period to the end point of the period based on the energy update history. Next, the generic parameter performance evaluation unit 160 calculates a rate of change indicating the degree of difference between the first index value and the second index value. Then, the generic parameter performance evaluation unit 160 judges whether the solution-finding performance when generic parameter values are used for the evaluation problem is sufficient according to a comparison between the rate of change and the reference value calculated by the reference value calculation unit 135.
[0076] If the generic parameter performance evaluation unit 160 determines that the solution-finding performance is sufficient, it outputs the determination result to the solution output unit 190. If the generic parameter performance evaluation unit 160 determines that the solution-finding performance is not sufficient, it instructs the generic parameter search unit 170 to execute a specialized parameter search.
[0077] The specialized parameter search unit 170 performs specialized parameter search. The specialized parameter search is a process of acquiring specialized parameter values for an evaluation problem. The specialized parameter search unit 170 inputs candidate values of specialized parameters and data of the evaluation problem to the accelerator card 108, executes a solution search for a certain period of time, and acquires an update history of energy in the solution search. The specialized parameter search unit 170 causes the accelerator card 108 to execute a solution search for the evaluation problem for each of a plurality of candidate values. The specialized parameter search unit 170 acquires the energy reached for each candidate value and the time at which the energy was reached, and evaluates each candidate value such that the lower the energy is reached in a shorter time, the higher the evaluation. Then, the specialized parameter search unit 170 selects the most highly evaluated candidate value as the specialized parameter value, and outputs it to the solution search unit 180.
[0078] The solution search unit 180 inputs the data of the evaluation problem and the specialized parameter values to the accelerator card 108, and executes a solution search for the evaluation problem. The solution search unit 180 acquires a solution acquired by the solution search from the accelerator card 108. The solution search unit 180 outputs the acquired solution to the solution output unit 190.
[0079] When the generic parameter performance evaluation unit 160 determines that the solution-finding performance using the generic parameter values is sufficient, the solution output unit 190 obtains a solution from the solution search unit 150 and outputs the solution. When the solution output unit 190 determines that the solution-finding performance using the generic parameter values is not sufficient, the solution output unit 190 obtains a solution from the solution search unit 180 and outputs the solution.
[0080] For example, the solution output section 190 may cause an image showing the solution to the evaluation problem to be displayed on the display 51, or may transmit data of the solution to another computer via the network .
[0081] Although the general-purpose parameter search unit 130, the solution search units 150 and 180, and the specialized parameter search unit 170 are described as causing the accelerator card 108 to execute the solution search, the function of the solution search may be realized by the processor 101 executing a program. In this case, the data processing device 100 may not have the accelerator card 108.
[0082] Next, an example of an update history of energy obtained in a solution search for an evaluation problem will be described. FIG. 4 is a diagram showing a first example of the relationship between time and energy in a solution search. Graph 200 illustrates the relationship between time and energy indicated by the energy update history. The horizontal axis of graph 200 is time. The vertical axis of graph 200 is the energy reached by the solution search. For example, the time of the solution search, that is, the time from the start to the end of the solution search, is specified in advance.
[0083] For example, the general-purpose parameter performance evaluation unit 160 evaluates that sufficient solution-finding performance has not been obtained if the rate of change in energy calculated from the first m% and the last n% of the energy update history using the following formula (5) is less than a reference value, where m≦n.
[0084] Rate of change of energy = (median of energy in the first m% - median of energy in the second n%) / (median of energy in the first m%) (5) However, when an initial solution for solution search has been obtained by preprocessing using another solution search method, the general-purpose parameter performance evaluation unit 160 may use the following equation (6).
[0085] Rate of change of energy = (energy of initial solution - median of energy of last n%) / (energy of initial solution) (6) Furthermore, when the energy of a theoretical solution is known in advance, the general-purpose parameter performance evaluation section 160 may use a difference rate expressed by the following equation (7) as the rate of change.
[0086] Energy difference rate = (energy of theoretical solution - median of the energies of the last n%) / energy of theoretical solution (7) The energy of the theoretical solution may be the energy of a local solution, the energy of an optimal solution, or the energy of a solution other than these.
[0087] The reference value is calculated in advance by reference value calculation unit 135 based on the energy update history of the generic parameter search using the learning problem, using a formula similar to the rate of change or rate of difference used by generic parameter performance evaluation unit 160. Specifically, reference value calculation unit 135 may obtain a rate of change similar to the above for each learning problem, based on the history of energy obtained in the solution search for each learning problem with the generic parameter value, and may set the median of the rates of change for each learning problem as the reference value.
[0088] Furthermore, when comparing with the rate of change, the general-purpose parameter performance evaluating section 160 may multiply the reference value by a variable d (d<1) to set the maximum level of performance that is acceptable. Here, the general-purpose parameter performance evaluation unit 160 can also evaluate the solution performance for the general-purpose parameter values based on the difficulty level of the evaluation problem. Next, an evaluation example of the solution performance using the difficulty level of the evaluation problem will be described.
[0089] FIG. 5 is a diagram showing a second example of the relationship between time and energy in solution search. Graph 300 illustrates the relationship between time and energy shown by the update history of energy. The horizontal axis of graph 300 is time. The vertical axis of graph 300 is the energy reached by solution search. The general-purpose parameter performance evaluation unit 160 obtains the time T0 from the start time of solution search to the final update time of energy based on the update history of energy.
[0090] And the general-purpose parameter performance evaluation unit 160 may evaluate that the solution performance is sufficient when the time T0 is short with respect to the difficulty level of the evaluation problem. Specifically, the general-purpose parameter performance evaluation unit 160 may evaluate that the solution performance is sufficient when the time T0 < a × (difficulty level of the problem) is satisfied. a is a coefficient given in advance. a is a positive real number. In this case, the general-purpose parameter performance evaluation unit 160 evaluates that the solution performance is not sufficient when the time T0 ≥ a × (difficulty level of the problem).
[0091] When it is not determined that the solution performance is sufficient by comparing the change rate with the reference value, the general-purpose parameter performance evaluation unit 160 may perform an evaluation of the solution performance based on the comparison between the difficulty level of the problem and the time T0. Alternatively, when it is not determined that the solution performance is sufficient by comparing the difficulty level of the problem and the time T0, the general-purpose parameter performance evaluation unit 160 may perform an evaluation of the solution performance based on the comparison between the change rate and the reference value.
[0092] Next, the processing procedure of the data processing apparatus 100 will be described. FIG. 6 is a flowchart showing a processing example of the data processing apparatus. (S10) The learning problem input unit 120 acquires the data of the learning problem from the storage unit 110 and inputs it to the general-purpose parameter search unit 130.
[0093] (S11) The generic parameter search unit 130 performs a generic parameter search. Specifically, the generic parameter search unit 130 inputs generic parameter candidate values and learning problem data to the accelerator card 108, executes a solution search for a certain period of time, and obtains an update history of energy in the solution search. For a plurality of learning problems, the generic parameter search unit 130 obtains the energy reached for each candidate value and the time it took to reach that energy, and evaluates each candidate value so that the lower the energy reached in the shorter time, the higher the evaluation. The generic parameter search unit 130 then selects the most highly evaluated candidate value as the generic parameter value, and stores it in the memory unit 110.
[0094] (S12) Reference value calculation unit 135 calculates a reference value based on the energy update history in the solution search acquired in the generic parameter search, and stores the reference value in storage unit 110. The reference value is used in evaluation of solution finding performance in generic parameter performance evaluation unit 160.
[0095] (S13) The evaluation question input unit 140 acquires the evaluation question data and the generic parameter values from the storage unit 110, and inputs them to the solution search unit 150. The evaluation question input unit 140 also inputs the evaluation question data to the difficulty level calculation unit 145.
[0096] (S14) The solution search unit 150 inputs the data of the evaluation problem and the generic parameter values to the accelerator card 108, and executes a solution search for the evaluation problem. The solution search unit 150 acquires, from the accelerator card 108, the energy update history in the solution search and the searched solution.
[0097] (S15) The difficulty level calculation unit 145 calculates the difficulty level of the evaluation problem based on the data of the evaluation problem. The difficulty level is used to evaluate the solution performance in the general-purpose parameter performance evaluation unit 160. For example, either equation (3) or (4) is used to calculate the difficulty level. However, equations (3) and (4) can be changed depending on the number of constraint terms for each type of constraint.
[0098] (S16) The general-purpose parameter performance evaluation unit 160 performs general-purpose parameter performance evaluation. In the general-purpose parameter performance evaluation, the general-purpose parameter performance evaluation unit 160 obtains a first index value of the energy calculated by the energy function of the evaluation problem. Also, the general-purpose parameter performance evaluation unit 160 obtains a second index value of the energy obtained in a first period from a time point after the start of the solution search period to the end of the period based on the update history of the energy. The general-purpose parameter performance evaluation unit 160 calculates a change rate by dividing the difference between the first index value and the second index value by the first index value. For example, any of equations (5), (6), and (7) is used for the calculation of the change rate.
[0099] Then, the general-purpose parameter performance evaluation unit 160 evaluates whether the solution performance is sufficient by comparing the change rate with the reference value calculated by the reference value calculation unit 135. For example, the general-purpose parameter performance evaluation unit 160 evaluates that the solution performance is sufficient when the change rate is greater than or equal to the reference value, and does not evaluate that the solution performance is sufficient when the change rate is less than the reference value.
[0100] Also, when the general-purpose parameter performance evaluation unit 160 does not evaluate that the solution performance is sufficient by the comparison, it evaluates the solution performance based on the comparison between the difficulty level of the problem and the aforementioned time T0. Specifically, the general-purpose parameter performance evaluation unit 160 evaluates that the solution performance is sufficient when the time T0 < a × (difficulty level of the problem) is satisfied. The general-purpose parameter performance evaluation unit 160 evaluates that the solution performance is not sufficient when the time T0 ≥ a × (difficulty level of the problem).
[0101] (S17) The general-purpose parameter performance evaluation unit 160 determines whether sufficient performance has been obtained by at least one of the methods of evaluating the solution performance by comparing the change rate with the reference value and evaluating the solution performance by comparing the difficulty level of the problem with the time T0. If it is evaluated that sufficient performance has been obtained by at least one of the methods, the process proceeds to step S20. If it is not evaluated that sufficient performance has been obtained by any of the methods, the process proceeds to step S18.
[0102] (S18) The specialized parameter search unit 170 performs a specialized parameter search. The specialized parameter search unit 170 inputs specialized parameter candidate values and evaluation problem data to the accelerator card 108, executes a solution search for a certain period of time, and obtains an update history of energy in the solution search. The specialized parameter search unit 170 causes the accelerator card 108 to execute a solution search for the evaluation problem for each of a plurality of candidate values. The specialized parameter search unit 170 obtains the energy reached for each candidate value and the time it took to reach that energy, and evaluates each candidate value such that the lower the energy reached in the shorter time, the higher the evaluation. The specialized parameter search unit 170 then selects the most highly evaluated candidate value as the specialized parameter value.
[0103] (S19) The solution search unit 180 inputs the data of the evaluation problem and the specialized parameter value to the accelerator card 108, and executes a solution search for the evaluation problem. The solution search unit 180 acquires the solution found by the solution search from the accelerator card 108.
[0104] (S20) The solution output unit 190 outputs the solution acquired by the solution search unit 150 or the solution acquired by the solution search unit 180 as the final solution to the evaluation problem. Then, the processing of the data processing device 100 ends.
[0105] In this way, the generic parameter performance evaluation unit 160 determines whether or not to perform a specialized parameter search for the evaluation problem based on the evaluation of the solution-finding performance by comparing the rate of change with the reference value. Note that in step S16, the generic parameter performance evaluation unit 160 does not need to evaluate the solution-finding performance based on the difficulty of the evaluation problem. In that case, in step S17, the generic parameter performance evaluation unit 160 evaluates that the solution-finding performance is sufficient, that is, sufficient performance has been obtained, when the rate of change is equal to or greater than the reference value. On the other hand, the generic parameter performance evaluation unit 160 evaluates that the solution-finding performance is insufficient, that is, sufficient performance has not been obtained, when the rate of change is less than the reference value.
[0106] Next, a comparative example will be described. In the comparative example, the functions of the reference value calculation unit 135, the difficulty calculation unit 145, the generalized parameter performance evaluation unit 160, the specialized parameter search unit 170, the solution search unit 180 and the solution output unit 190 are not used.
[0107] FIG. 7 is a flowchart showing a comparative example. (S30) The learning problem input unit 120 inputs learning problem data to the general-purpose parameter search unit 130.
[0108] (S31) The general-purpose parameter search unit 130 acquires general-purpose parameters. In acquiring the general-purpose parameters, the same process as the general-purpose parameter search described above is performed. However, in step S31, the general-purpose parameter search unit 130 executes a solution search for a certain period of time using the accelerator card 108 for one candidate value of the target parameter.
[0109] (S32) The general-purpose parameter search unit 130 acquires the energy update history from the accelerator card 108, and evaluates candidate parameter values based on the energy update history. (S33) The generic parameter search unit 130 determines whether or not to end the parameter search, i.e., whether or not to end the repetition of the processes of steps S31 and S32. If the parameter search is to be ended, processing proceeds to step S34. If the parameter search is not to be ended, processing proceeds to step S31. For example, the generic parameter search unit 130 determines to end the parameter search when evaluation of all candidate values is completed using all learning problems. If not, the generic parameter search unit 130 determines not to end the parameter search and proceeds to step S31, where it performs processing on combinations of unexecuted learning problems and candidate values.
[0110] (S34) The general-purpose parameter search unit 130 outputs the general-purpose parameter value. (S35) The evaluation problem input unit 140 inputs the evaluation problem data and the generic parameter values to the solution search unit 150.
[0111] (S36) The solution search unit 150 inputs the data of the evaluation problem and the general-purpose parameter values to the accelerator card 108, and executes a solution search for the evaluation problem. The solution search unit 150 obtains the obtained solution from the accelerator card 108.
[0112] (S37) The solution searching unit 150 outputs the acquired solution. Then, the process of the comparative example ends. It is possible to use generic parameter values for all evaluation problems, as in the comparative example. However, generic parameter values are not necessarily appropriate for all evaluation problems. Specifically, the following experiment was conducted.
[0113] In the experiment, all 134 problems of the standard benchmark "QAP-LIB" whose solutions were known were solved. Here, the following reference 1 serves as a reference for QAP-LIB. Reference 1: RE Burkard, SE Karisch, and F. Rendl, “Qaplib - a quadratic assignment problem library,” Journal of Global Optimization, vol. 10, no. 4, pp. 391-403, 1997. Ten problems from QAP-LIB were used as training problems, and the general-purpose parameter values were obtained by 1,000 general-purpose parameter searches with each search taking 30 seconds. Using these general-purpose parameter values, solutions were searched for a total of 134 problems with an evaluation time of 60 seconds, and performance was evaluated based on the rate of change in the obtained energy.
[0114] As a result, the gap between the solution and the 125 problems for which performance was judged to be sufficient was 0.047%, and the rate of reaching the solution was 93.6% (117 / 125). On the other hand, the gap between the solution and the 9 problems for which performance was judged to be insufficient was 0.303%, and the rate of reaching the solution was poor at 33.3% (3 / 9). This experiment confirmed that it is possible to correctly judge whether or not the solution-finding performance is sufficient based on the rate of change.
[0115] In addition, the method of the comparative example may not be able to obtain an appropriate solution for some problems. Therefore, a specialized parameter search was performed for 300 seconds (30 seconds x 10 times) for each of the nine problems that were determined to have insufficient performance, and a solution search was performed using the specialized parameter values obtained by the specialized parameter search. As a result, the gap with the solution for the nine problems improved to 0.216%, and the rate of reaching the solution improved to 77.8% (7 / 9). In particular, for Tai80a, one of the nine problems, the energy reached in 60 seconds of evaluation time using specialized parameter values was not reached even after one hour using general-purpose parameter values.
[0116] In this way, the data processing device 100 uses the rate of change and rate of difference of energy to evaluate the solution-finding performance. The rate of change and rate of difference can be easily acquired from the energy update history in the solution search. Therefore, the data processing device 100 can efficiently evaluate the solution-finding performance for the evaluation problem at the corresponding parameter value.
[0117] Furthermore, the data processing device 100 can reduce the execution time for finding a solution by narrowing down the problems for which the performance is insufficient with the general-purpose parameter values and performing specialized parameter search through performance evaluation using the rate of change or difference of energy or difficulty. In this way, the data processing device 100 can make it possible to find a solution efficiently.
[0118] As described above, the data processing device 100 executes, for example, the following processes. The accelerator card 108 executes a solution search using a first parameter value based on the evaluation function. The solution search may be executed by the processor 101. The processor 101 acquires a first index value of the evaluation function value calculated by the evaluation function. The processor 101 acquires a second index value of the evaluation function value obtained in a first period from a point after the start of the solution search period to the end of the period. The processor 101 calculates a rate of change indicating the degree of difference between the first index value and the second index value based on the first index value and the second index value. The processor 101 evaluates the solution-finding performance by the first parameter value for a problem (evaluation problem) expressed by the evaluation function based on the rate of change.
[0119] This allows the data processing device 100 to efficiently evaluate the parameter values. When multiple types of parameters are used in the solution search, the first parameter value is a set of multiple types of parameter values. For example, when the solution search is performed by the SA method, the first parameter value may be both or either one of the maximum temperature value and the minimum temperature value. When the solution search is performed by a method other than the SA method, the first parameter value may be a parameter value of a type other than the maximum temperature value and the minimum temperature value.
[0120] In addition, the general-purpose parameter performance evaluation unit 160 can use multiple reference values to evaluate the solution-seeking performance in three or more stages according to a comparison between the rate of change and each of the multiple reference values. For example, the general-purpose parameter performance evaluation unit 160 may evaluate the solution-seeking performance using a first reference value and a second reference value larger than the first reference value. For example, the general-purpose parameter performance evaluation unit 160 may evaluate the solution-seeking performance as higher than standard when the rate of change is larger than the second reference value. The general-purpose parameter performance evaluation unit 160 may evaluate the solution-seeking performance as standard when the rate of change is equal to or larger than the first reference value and smaller than the second reference value. The general-purpose parameter performance evaluation unit 160 may evaluate the solution-seeking performance as lower than standard when the rate of change is less than the first reference value. The general-purpose parameter performance evaluation unit 160 may output such an evaluation result to support the user in checking the evaluation result of the solution-seeking performance according to the first parameter value.
[0121] For example, the evaluation function may be an energy function of an Ising model, or may be any other type of function used in problem expression. For example, the processor 101 calculates the rate of change by dividing the difference between the first index value and the second index value by the first index value, which enables the data processing device 100 to easily obtain the rate of change.
[0122] The processor 101 may also obtain the median of the evaluation function values obtained in the first period as the second index value. This allows the data processing device 100 to easily obtain the second index value. However, the second index value may be another statistical value such as the average value, minimum value, or percentile value of the evaluation function values obtained in the first period.
[0123] The processor 101 may also obtain, as the first index value, the median value of the evaluation function obtained in a second period that is earlier than the first period and starts from the start point of the solution search. This allows the data processing device 100 to easily obtain the first index value.
[0124] Furthermore, the processor 101 may obtain, as the first index value, a value of the evaluation function corresponding to an initial solution in a solution search for the evaluation problem. This allows the data processing device 100 to easily obtain the first index value.
[0125] The processor 101 may also obtain a value of the evaluation function corresponding to a predetermined solution, which is calculated in advance for the evaluation function, as the first index value. This allows the data processing device 100 to easily obtain the first index value. The predetermined solution may be a local solution, an optimal solution, or any other solution.
[0126] Furthermore, in evaluating the solution-finding performance, the processor 101 compares, for example, the rate of change with a predetermined reference value. Then, when the rate of change is less than the reference value, the processor 101 determines that the solution-finding performance is insufficient. On the other hand, when the rate of change is equal to or greater than the reference value, the processor 101 determines that the solution-finding performance is sufficient. This allows the data processing device 100 to efficiently evaluate the parameter values. The reference value may be designated in advance by the user. As described above, when comparing the rate of change with the reference value, the processor 101 may compare a value obtained by multiplying the reference value by a variable value (coefficient value) less than 1 with the rate of change.
[0127] The processor 101 may calculate the rate of change by dividing the absolute value of the difference between the first index value and the second index value by the first index value. In this case, when the processor 101 uses the value of the evaluation function corresponding to the optimal solution as the first index value, the processor 101 may evaluate the solution-finding performance as follows. That is, the processor 101 may determine that the solution-finding performance is insufficient when the rate of change is greater than a reference value. On the other hand, the processor 101 may determine that the solution-finding performance is sufficient when the rate of change is equal to or less than the reference value.
[0128] The first parameter value may be a general-purpose parameter value that is used for multiple problems. In this case, when the solution-finding performance is insufficient, the processor 101 determines to acquire a second parameter value (specialized parameter value) specialized for the evaluation problem. When the solution-finding performance is sufficient, the processor 101 determines not to acquire the second parameter value. This allows the data processing device 100 to narrow down the problems from which the second parameter value is to be acquired. As a result, the data processing device 100 can efficiently execute the solution-finding. The second parameter value is a value of a parameter of the same type as the first parameter value. When multiple types of parameters are used in the solution search, the second parameter value is a set of values of the multiple types of parameters.
[0129] The processor 101 may acquire a history of values of an evaluation function obtained by a solution search performed to acquire a first parameter value for each of a plurality of learning problems, and calculate a reference value based on the history of the acquired values of the evaluation function. This allows the data processing device 100 to eliminate the need for a user to specify a reference value. Furthermore, by automatically setting the reference value, the data processing device 100 eliminates the need for a user to have knowledge of the problem, making it easier for the user to use the solution-finding function of the data processing device 100.
[0130] Furthermore, when the processor 101 determines not to acquire the second parameter value, it outputs a first solution obtained by a solution search using the first parameter value. On the other hand, when the processor 101 determines to acquire the second parameter value, it performs a solution search using each of a plurality of parameter candidate values, and acquires the second parameter value based on the execution result of the solution search for each of the plurality of parameter candidate values. The processor 101 performs a solution search using the second parameter value, and outputs a second solution obtained by the solution search using the second parameter value.
[0131] As a result, the data processing device 100 can omit acquisition of the second parameter value for a problem for which it is not necessary to acquire the second parameter value, and efficiently acquire a first solution. Also, for a problem for which the second parameter value is acquired, the data processing device 100 can efficiently acquire a second solution that is more appropriate than the first solution for the problem. As described above, the processor 101 may perform a solution search using a search unit such as the accelerator card 108.
[0132] Furthermore, the processor 101 may evaluate the solution performance of the problem (evaluation problem) based on the difficulty of the problem and the time T0 from the start of the solution search period to the time when the value of the evaluation function was last updated. This allows the data processing device 100 to more appropriately evaluate the parameter values.
[0133] The processor 101 may calculate the difficulty of a problem based on, for example, the number of state variables and the number of constraint terms representing constraints included in an evaluation function corresponding to the problem. This allows the data processing device 100 to efficiently calculate the difficulty of a problem. For example, the processor 101 increases the difficulty level as the number of state variables increases. Also, the processor 101 increases the difficulty level as the number of constraint terms increases.
[0134] The information processing of the first embodiment can be realized by causing the processing unit 12 to execute a program. The information processing of the second embodiment can be realized by causing the processor 101 to execute a program. The program can be recorded in a computer-readable recording medium 53.
[0135] For example, the program can be distributed by distributing recording medium 53 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 53 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. [Explanation of symbols]
[0136] 10 Data Processing Device 11 Storage section 12 Processing section 20 Graphs
Claims
1. On the computer, When a solution search is executed using a first parameter value based on an evaluation function, a rate of change indicating a degree of difference between the first index value and the second index value is calculated based on a first index value of the value of the evaluation function calculated by the evaluation function and a second index value of the value of the evaluation function obtained in a first period from a point after the start point of the period of the solution search to an end point of the period; evaluating the solution performance of the first parameter value for the problem represented by the evaluation function based on the rate of change; A program that executes a process.
2. The program according to claim 1 , wherein the evaluation function is an energy function of an Ising model.
3. In calculating the rate of change, the rate of change is calculated by dividing the difference between the first index value and the second index value by the first index value. The program according to claim 1, which causes the computer to execute a process.
4. A median value of the evaluation function values obtained in the first period is acquired as the second index value. The program according to claim 1, which causes the computer to execute a process.
5. A median value of the evaluation function obtained in a second time period that is earlier than the first time period and starts from the start time point is acquired as the first index value. The program according to claim 1, which causes the computer to execute a process.
6. The first index value is a value of the evaluation function corresponding to an initial solution in the solution search. The program according to claim 1.
7. The first index value is a value of the evaluation function corresponding to a predetermined solution that is pre-calculated for the evaluation function. The program according to claim 1.
8. In the evaluation of the solution finding performance, the rate of change is compared with a predetermined reference value, and if the rate of change is less than the reference value, it is determined that the solution finding performance is insufficient, and if the rate of change is equal to or greater than the reference value, it is determined that the solution finding performance is sufficient. The program according to claim 1, which causes the computer to execute a process.
9. The first parameter value is a general-purpose parameter value that is used for a plurality of problems; determining, when the solution-finding performance is insufficient, to obtain a second parameter value specialized for the problem, and, when the solution-finding performance is sufficient, to not obtain the second parameter value; The program according to claim 8, which causes the computer to execute a process.
10. acquiring a history of values of the evaluation function obtained by the solution search performed to acquire the first parameter value for each of a plurality of learning problems, and calculating the reference value based on the acquired history of values of the evaluation function. The program according to claim 9, which causes the computer to execute a process.
11. when it is determined that the second parameter value is not to be acquired, a first solution obtained by the solution search using the first parameter value is output; When it is determined that the second parameter value is to be acquired, the solution search is performed using each of a plurality of parameter candidate values, the second parameter value is acquired based on a result of the solution search performed for each of the plurality of parameter candidate values, the solution search is performed using the second parameter value, and a second solution obtained by the solution search using the second parameter value is output. The program according to claim 9, which causes the computer to execute a process.
12. evaluating a solution performance for the problem using the first parameter value based on the difficulty of the problem and the time from the start time to the time when the value of the evaluation function was last updated in the solution search; The program according to claim 1, which causes the computer to execute a process.
13. Calculating the difficulty level based on the number of state variables included in the evaluation function and the number of constraint terms representing constraints. The program according to claim 12, which causes the computer to execute a process.
14. The computer When a solution search is executed using a first parameter value based on an evaluation function, a rate of change indicating a degree of difference between the first index value and the second index value is calculated based on a first index value of the value of the evaluation function calculated by the evaluation function and a second index value of the value of the evaluation function obtained in a first period from a point after the start point of the period of the solution search to an end point of the period; evaluating the solution performance of the first parameter value for the problem represented by the evaluation function based on the rate of change; Data processing methods.
15. a storage unit that stores a first parameter value used in a solution search based on an evaluation function; a processing unit that, when the solution search is executed using the first parameter value, calculates a rate of change indicating a degree of difference between the first index value and the second index value based on a first index value of the value of the evaluation function calculated by the evaluation function and a second index value of the value of the evaluation function obtained in a first period from a point after the start point of the period of the solution search to an end point of the period, and evaluates a solution-finding performance using the first parameter value for the problem represented by the evaluation function based on the rate of change; A data processing device comprising:
Citation Information
Patent Citations
Optimization device, optimization method and optimization program
JP2020194273A
Optimization apparatus, optimization program, and optimization method
JP2021043787A
Quantum and digital processor hybrid systems and methods to solve problems
US20110060710A1
Optimization apparatus, optimization method, and recording medium
US20200380065A1
Systems and methods for optimizing annealing parameters
US20210241143A1
