Optimization device, optimization method and procedure
By combining variable transformation, inverse transformation, and evaluation value calculation components, the problem of deterioration in the quantization accuracy of continuous variables is solved, and high-precision quantization is achieved when there is no consistent discretization point.
Patent Information
- Application Number
- CN202480090903.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-23
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies suffer from degraded quantization accuracy when continuous variables are quantized and there is no consistent discretization point.
The variable transformation unit converts continuous variables into candidate solutions of binary variables, and the variable inverse transformation unit converts them back into continuous variables. The evaluation value is calculated by the evaluation value calculation unit, and the transformation and inverse transformation rules are updated by the transformation rule update unit to prevent the degradation of quantization accuracy.
It effectively prevents the degradation of quantization accuracy when there is no discretization point consistent with the continuous variable, and improves quantization accuracy.
Smart Images

Figure CN123125610A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to an optimization apparatus, optimization method, and procedure. Background Technology
[0002] There exists an optimization device that quantizes continuous variables, which are represented by either real or integer values.
[0003] As an example of such an optimization device, Patent Document 1 discloses an optimization device that quantizes a continuous variable by determining the discretization point closest to the continuous variable among multiple discretization points. The multiple discretization points are points where the domain of the continuous variable is equally divided.
[0004] Patent Document 1: Japanese Patent Application Publication No. 2020-113190 Summary of the Invention
[0005] The optimization device disclosed in Patent Document 1 has the following problem: when there is no discretization point that is consistent with the continuous variable among multiple discretization points, the quantization accuracy will deteriorate.
[0006] This disclosure was made to solve the problems described above, and its purpose is to obtain an optimization device that can prevent the degradation of quantization accuracy when there is no discretization point consistent with the continuous variable.
[0007] The optimization apparatus disclosed herein comprises: a variable transformation unit that acquires a first continuous variable as a variable represented by either a real value or an integer value, and transforms the first continuous variable into a candidate solution of a binary variable; a variable inverse transformation unit that inversely transforms the candidate solution into a second continuous variable; an evaluation value calculation unit that calculates an evaluation value for the second continuous variable; and a transformation rule updating unit that updates, based on the evaluation value, the transformation rule performed by the variable transformation unit to transform into a candidate solution and the inverse transformation rule performed by the variable inverse transformation unit to inversely transform into a second continuous variable.
[0008] According to this disclosure, it is possible to prevent degradation of quantization accuracy when there is no discretization point consistent with the continuous variable. Attached Figure Description
[0009] Figure 1 This is a structural diagram showing the optimization device involved in Implementation Method 1.
[0010] Figure 2 This is a hardware structure diagram showing the hardware of the optimization device involved in Implementation Method 1.
[0011] Figure 3 It is a hardware structure diagram of a computer when the optimization device is implemented by software or firmware.
[0012] Figure 4 It means Figure 1 The flowchart of the optimization method is shown as the processing steps of the optimization device.
[0013] Figure 5 This is an explanatory diagram showing the multiple discretization processes performed by the variable transformation unit 1.
[0014] Figure 6 This indicates that the discretization point s is processed by the update step 4 of the transformation rule. k An illustration of the processing of (k = 1, ..., K) moves. Detailed Implementation
[0015] The following description, in order to illustrate the present disclosure in more detail, describes the form used to implement the present disclosure in accordance with the accompanying drawings.
[0016] Implementation Method 1
[0017] Figure 1 This is a structural diagram showing the optimization device involved in Implementation Method 1.
[0018] Figure 2 This is a hardware structure diagram showing the hardware of the optimization device involved in Implementation Method 1.
[0019] Figure 1 The optimization device shown includes a variable conversion unit 1, a variable inverse conversion unit 2, an evaluation value calculation unit 3, and a conversion rule update unit 4.
[0020] Variable transformation unit 1, for example, is composed of Figure 2 The variable conversion circuit 11 shown is implemented.
[0021] The variable conversion unit 1 includes an encoder 1a, a model estimator 1b, and a sampler 1c.
[0022] The variable conversion unit 1 obtains the first continuous variable x as a variable represented by either a real value or an integer value.
[0023] Variable transformation unit 1 transforms the first continuous variable x into a candidate solution z of a binary variable. t The process of converting the first continuous variable x into a candidate solution zt of a binary variable is the first process.
[0024] Variable transformation unit 1 transforms the candidate solutions z of the binary variable t Output to variable inverse conversion section 2.
[0025] Encoder 1a retrieves the first continuous variable x from storage unit 3b, which will be described later.
[0026] Encoder 1a discretizes the first continuous variable x and outputs the discrete value z, which is the discretization result of the first continuous variable x, to model estimator 1b.
[0027] Specifically, encoder 1a calculates the distances between multiple discretized points and the first continuous variable x.
[0028] Encoder 1a represents the distances between multiple discretized points and the first continuous variable x, and denotes the shortest distance L. k The relevant discretization point s k The solution z is output to the model estimator 1b.
[0029] Model estimator 1b obtains the representation and shortest distance L from encoder 1a. k The relevant discretization point s k The solution z.
[0030] Model estimator 1b estimates the Q matrix of the surrogate model based on the solution z and outputs the Q matrix to sampler 1c.
[0031] Sampler 1c obtains the Q matrix from model estimator 1b.
[0032] Sampler 1c performs the Bayesian Optimization (BO) algorithm using the Q matrix to compute candidate solutions z for the binary variable. t .
[0033] Sampler 1c will select candidate solutions z for the binary variable t Output to variable inverse conversion section 2.
[0034] Inverse transformation part 2, for example, is from Figure 2 The variable inverse conversion circuit 12 shown is implemented.
[0035] Inverse transformation unit 2 obtains candidate solutions z for the binary variable from variable transformation unit 1. t .
[0036] Inverse transformation part 2 will convert candidate solutions z t Inverse transformation to the second continuous variable x t and the second continuous variable x t Output to the evaluation value calculation unit 3. The candidate solution z... t Inverse transformation to the second continuous variable x t The second processing step is to process the data.
[0037] Evaluation value calculation unit 3, for example, is composed of Figure 2 The evaluation value calculation circuit 13 shown is implemented.
[0038] The evaluation value calculation unit 3 includes an evaluator 3a and a storage unit 3b.
[0039] Evaluation value calculation unit 3 obtains the second continuous variable x from variable inverse transformation unit 2. t .
[0040] Evaluation value calculation unit 3 calculates the second continuous variable x t Evaluation value y t Calculate the second continuous variable x. t Evaluation value y t The processing is the third process.
[0041] Evaluator 3a obtains the second continuous variable x from the variable inverse transformation unit 2. t .
[0042] Evaluator 3a uses an evaluation function to calculate the second continuous variable x. t Evaluation value y t .
[0043] Evaluator 3a will use the second continuous variable x t and evaluation value y t The data is output to storage unit 3b respectively.
[0044] Storage unit 3b stores the second continuous variable x. t and evaluation value y t .
[0045] Transformation Law Update Section 4, for example, by Figure 2 The conversion rule shown is updated in circuit 14.
[0046] The conversion rule update unit 4 is based on the evaluation value y calculated by the evaluation value calculation unit 3. t The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule. Update the transformation to candidate solution z performed by variable transformation part 1 respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule is the fourth step.
[0047] Figure 1 In the envisioned optimization device, the components namely the variable transformation unit 1, the variable inverse transformation unit 2, the evaluation value calculation unit 3, and the transformation rule updating unit 4 are respectively composed of... Figure 2The example shown is a dedicated hardware implementation. That is, in the optimization device, the variable transformation circuit 11, the variable inverse transformation circuit 12, the evaluation value calculation circuit 13, and the transformation rule update circuit 14 are respectively equivalent to, for example, a single circuit, a composite circuit, a processor that executes processing according to program instructions, a processor that executes processing according to parallel program instructions, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.
[0048] The components of an optimization device are not limited to being implemented by dedicated hardware; the optimization device can also be implemented by software, firmware, or a combination of software and firmware.
[0049] Software or firmware is stored as a program in a computer's memory. A computer refers to the hardware that executes programs, such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor).
[0050] Figure 3 It is a hardware structure diagram of a computer when the optimization device is implemented by software or firmware.
[0051] When the optimization device is implemented by software or firmware, the program for executing each processing step in the variable conversion unit 1, the variable inverse conversion unit 2, the evaluation value calculation unit 3, and the conversion rule update unit 4 is stored in the memory 21. Then, the computer's processor 22 executes the program stored in the memory 21.
[0052] in addition, Figure 2 Examples are shown where the components of the optimization device are implemented by dedicated hardware. Figure 3 An example is shown where the optimization device is implemented by software or firmware. However, this is only one example; some components of the optimization device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware.
[0053] Next, an explanation Figure 1 The operation of the optimization device shown.
[0054] Figure 4 It means Figure 1 The flowchart of the optimization method is shown as the processing steps of the optimization device.
[0055] First, variable transformation unit 1 obtains the first continuous variable x from the outside, for example.
[0056] Variable transformation unit 1 transforms the first continuous variable x into a candidate solution z of a binary variable. t .
[0057] Variable transformation unit 1 transforms the candidate solutions z of the binary variable t Output to variable inverse conversion section 2.
[0058] The following details the transformation of the variable transformation unit 1 into the candidate solution z. t The conversion process.
[0059] Figure 5 This is an explanatory diagram showing the multiple discretization processes performed by the variable transformation unit 1.
[0060] The encoder 1a of the variable conversion unit 1, for example, obtains the first continuous variable x from an external source. Figure 4 Step ST1).
[0061] The encoder 1a has K discretization points s stored in its internal memory. k k = 1, ..., K, where K is an integer greater than or equal to 2.
[0062] like Figure 5 As shown, there are K discretization points s k The points are obtained by equally dividing the domain W of the first continuous variable x. If the domain W of the first continuous variable x is, for example, 0.0 to 1.0, then the discretized points s are... k For example, it can be represented by the following formula (1).
[0063] s k = [0.0, 0.2, 0.4, 0.6, 0.8, 1.0] (1)
[0064] Encoder 1a calculates K discretization points s as shown in equation (2). k The distance L between (k = 1, ..., K) and the first continuous variable x k .
[0065] L k =|s k -x| (2)
[0066] Encoder 1a at K distances L1 to L K Determine the shortest distance L in the middle k The relevant discretization point s k ( Figure 4 Step ST2).
[0067] exist Figure 5 In the example, since the first continuous variable x is 0.35, the discretized point s3 (=0.4) is the shortest distance L. k The relevant discretization points.
[0068] Encoder 1a will represent the shortest distance L k The relevant discretization point s k The solution z is output to the model estimator 1b.
[0069] Model estimator 1b obtains the representation and shortest distance L from encoder 1a. k The relevant discretization point s k The solution z.
[0070] Model estimator 1b estimates the Q-matrix of the surrogate model based on the solution z. Figure 4 Step ST3). Since the processing of the Q matrix of the presumed surrogate model is a well-known technique, detailed explanation is omitted.
[0071] Model estimator 1b outputs the Q matrix to sampler 1c.
[0072] Sampler 1c obtains the Q matrix from model estimator 1b.
[0073] Sampler 1c performs the BO algorithm using the Q matrix to compute candidate solutions z for the binary variable. t ( Figure 4 Step ST4). Since candidate solutions z are calculated by performing the BO algorithm. t The processing itself is a well-known technique, so detailed explanation is omitted.
[0074] exist Figure 5 In the example, since the discretized point s3 (=0.4) is the shortest distance L k The relevant discretization points, therefore the candidate solutions z t It is represented by the following formula (3).
[0075] z t = [0, 0, 1, 0, 0, 0] (3)
[0076] In equation (3), 1 represents the shortest distance L. k The relevant discretized points, where 0 represents the distance from the point L to the shortest distance L. k Discretized points other than those related to distance.
[0077] Sampler 1c will select candidate solutions z for the binary variable t Output to variable inverse conversion section 2.
[0078] Inverse transformation unit 2 obtains candidate solutions z for the binary variable from variable transformation unit 1. t .
[0079] The internal memory of the variable inverse conversion unit 2 stores K discretization points s. k .
[0080] The variable inverse transformation part 2 is shown in equation (4) as follows: by transforming the candidate solution z t Multiply by the discretization point s k , will the candidate solution z t Inverse transformation to the second continuous variable x t ( Figure 4 Step ST5).
[0081] x t =z t ×s k (4)
[0082] exist Figure 5 In the example, due to the candidate solution z t As expressed in equation (3), the discretization point s k As expressed in equation (1), therefore the second continuous variable x t It is represented by the following formula (5).
[0083] x t =0.4 (5)
[0084] Inverse transformation part 2 transforms the second continuous variable x t Output to the evaluation value calculation unit 3.
[0085] The evaluator 3a of the evaluation value calculation unit 3 obtains the second continuous variable x from the variable inverse transformation unit 2. t .
[0086] Evaluator 3a uses the second continuous variable x t Substitute into the evaluation function f(·), and calculate the second continuous variable x. t Evaluation value y t ( Figure 4 Step ST6). t =f(x) t ).
[0087] The evaluation function f(·) is, for example, the following function: if the second continuous variable x t If it is consistent with the first continuous variable x, then it is used as the evaluation value y. t Returns the maximum value; the second continuous variable x t The further away from the first continuous variable x, the more it is used as the evaluation value y. t Returns a value smaller than the maximum value.
[0088] Evaluator 3a will use the second continuous variable x t and evaluation value y t The data is output to storage unit 3b respectively.
[0089] Storage unit 3b stores the second continuous variable x. t and evaluation value y t .
[0090] Transformation Rule Update Section 4 Based on Evaluation Value y t The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule ( Figure 4 Step ST7).
[0091] The following details the update process of the conversion rule and the update process of the reverse conversion rule performed by the conversion rule update unit 4.
[0092] Figure 6 This indicates that the discretization point s is processed by the update step 4 of the transformation rule. k An illustration of the processing of (k = 1, ..., K) moves.
[0093] exist Figure 6 In the example, since the first continuous variable x exists between discretization points s3 and s4, the method to narrow the interval between discretization points s3 and s4 is to make the K discretization points s1 to s4... K Move. Thus, by making K discretized points s1~s K Moving the discretization points increases the granularity of the discretization points where the optimal solution exists, thus improving the discretization accuracy.
[0094] The transformation rule update unit 4 retrieves the second continuous variable x from the storage unit 3b. t and evaluation value y t .
[0095] The transformation rule update part 4 is shown in equation (6) based on the second continuous variable x. t With evaluation value y t Calculate the discretization point s k The movement ds of (k = 1, ..., K) k .
[0096]
[0097] In equation (6), sng(s k -x t ) is a function as follows: if (s k -x t If (s) is 0 or higher, then return 1; if (s) is 0 or higher, then return 1. k -x t If σ is less than 0, then -1 is returned. σ is the evaluation value y. t The standard deviation is given by η, which is a constant.
[0098] The transformation rule update section 4 makes the discretization point s k(k = 1, ..., K) Moving discretization point s k The amount of movement ds k The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule.
[0099] exist Figure 6 In the example, at the step of t=0 before the optimization device starts the optimization process, since the first continuous variable x exists between discretization points s3 and s4, the K discretization points s1 to s4 are narrowed by making the interval between discretization points s3 and s4 smaller. K Move. Therefore, the interval between discretization point s3 and discretization point s4 in the next step t=1 is narrower than the interval between discretization point s3 and discretization point s4 in the step t=0.
[0100] The encoder 1a of the variable conversion unit 1 increments step t.
[0101] If step t has reached N max ( Figure 4 If step ST8 is the case (in the case of yes), then the series of processes performed by the optimization device ends. N max It is an integer greater than or equal to 1.
[0102] If step t has not yet reached N max ( Figure 4 Step ST8: If not, then encoder 1a will store the second consecutive variable x in storage unit 3b. t As the first continuous variable x, obtain ( Figure 4 Step ST9).
[0103] Encoder 1a calculates K discretization points s as shown in equation (2). k The distance L between (k = 1, ..., K) and the first continuous variable x k .
[0104] Encoder 1a at K distances L1 to L K Determine the shortest distance L in the middle k The relevant discretization point s k ( Figure 4 Step ST2).
[0105] Encoder 1a will represent the shortest distance L k The relevant discretization point s k The solution z is output to the model estimator 1b.
[0106] Model estimator 1b obtains the representation and shortest distance L from encoder 1a.k The relevant discretization point s k The solution z.
[0107] Model estimator 1b estimates the Q-matrix of the surrogate model based on the solution z. Figure 4 Step ST3).
[0108] Model estimator 1b outputs the Q matrix to sampler 1c.
[0109] Sampler 1c obtains the Q matrix from model estimator 1b.
[0110] Sampler 1c performs the BO algorithm using the Q matrix to compute candidate solutions z for the binary variable. t ( Figure 4 Step ST4).
[0111] Sampler 1c will select candidate solutions z for the binary variable t Output to variable inverse conversion section 2.
[0112] Inverse transformation unit 2 obtains candidate solutions z for the binary variable from variable transformation unit 1. t .
[0113] The variable inverse transformation part 2, as shown in equation (4), involves the candidate solution z t Multiply by the discretization point s k , will the candidate solution z t Inverse transformation to the second continuous variable x t ( Figure 4 Step ST5).
[0114] Inverse transformation part 2 transforms the second continuous variable x t Output to the evaluation value calculation unit 3.
[0115] The evaluator 3a of the evaluation value calculation unit 3 obtains the second continuous variable x from the variable inverse transformation unit 2. t .
[0116] Evaluator 3a uses the second continuous variable x t Substitute into the evaluation function f(·), and calculate the second continuous variable x. t Evaluation value y t ( Figure 4 Step ST6). t =f(x) t ).
[0117] The evaluation function f(·) is, for example, the following function: if the second continuous variable x t If it is consistent with the first continuous variable x, then it is used as the evaluation value y. t Returns the maximum value; the second continuous variable x t The further away from the first continuous variable x, the more it is used as the evaluation value y.t Returns a value smaller than the maximum value.
[0118] Evaluator 3a will use the second continuous variable x t and evaluation value y t The data is output to storage unit 3b respectively.
[0119] Storage unit 3b stores the second continuous variable x. t and evaluation value y t .
[0120] Transformation Rule Update Section 4 Based on Evaluation Value y t The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule ( Figure 4 Step ST7).
[0121] exist Figure 6 In the example, at step t=1, since the first continuous variable x exists between discretization points s3 and s4, the K discretization points s1 to s4 are narrowed by making the interval between discretization points s3 and s4 smaller. K Move. Therefore, the interval between discretization point s3 and discretization point s4 in the next step t=2 is narrower than the interval between discretization point s3 and discretization point s4 in the step t=1.
[0122] The encoder 1a of the variable conversion unit 1 increments step t.
[0123] If step t has reached N max ( Figure 4 If step ST8 is the case, then the series of processes performed by the optimization device will end.
[0124] If step t has not yet reached N max ( Figure 4 Step ST8: If not, then encoder 1a will store the second consecutive variable x in storage unit 3b. t As the first continuous variable x, obtain ( Figure 4 Step ST9).
[0125] Next, continue with steps ST2 to ST8. That is, repeat steps 1, 2, 3, and 4.
[0126] In Embodiment 1 described above, the optimization apparatus is configured to include: a variable conversion unit 1, which acquires a first continuous variable that is represented by either a real value or an integer value, and converts the first continuous variable into a candidate solution of a binary variable; a variable inverse conversion unit 2, which inversely converts the candidate solution into a second continuous variable; an evaluation value calculation unit 3, which calculates an evaluation value for the second continuous variable; and a conversion rule update unit 4, which updates the conversion rule for converting the variable into a candidate solution performed by the variable conversion unit 1 and the inverse conversion rule for inverse conversion into the second continuous variable performed by the variable inverse conversion unit 2, respectively, based on the evaluation value. Therefore, the optimization apparatus can prevent the degradation of quantization accuracy when there is no discretization point consistent with the continuous variable.
[0127] In Embodiment 1, the optimization device is configured to repeatedly perform transformation processing by the variable transformation unit 1, inverse transformation processing by the variable inverse transformation unit 2, evaluation value calculation processing by the evaluation value calculation unit 3, and update processing by the transformation rule update unit 4. Therefore, the optimization device can improve quantization accuracy when there is no discretization point consistent with the continuous variable.
[0128] In Embodiment 1, the optimization device is configured such that the variable transformation unit 1 determines any one of the multiple discretization points as a candidate solution based on the distance between the multiple discretization points and the first continuous variable. Therefore, the optimization device can convert the first continuous variable into a candidate solution for a binary variable even when there is no discretization point that matches the continuous variable.
[0129] In Implementation 1, the optimization device is configured such that the transformation rule update unit 4 updates the transformation rule, causing the discretization point to move. Therefore, the optimization device can prevent the degradation of quantization accuracy when there is no discretization point consistent with the continuous variable.
[0130] In Embodiment 1, the optimization device is configured such that the transformation rule update unit 4 updates the transformation rule and the inverse transformation rule based on the second continuous variable and the evaluation value, respectively. Therefore, the optimization device can easily update the transformation rule and the inverse transformation rule separately.
[0131] exist Figure 1 In the optimization device shown, the transformation rule update unit 4 calculates the discretization point s as shown in equation (6). k The movement ds of (k = 1, ..., K) k By making the discretization point s k Moving amount ds k The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x tThe inverse transformation rule. However, this is only one example; the transformation rule update part 4 can also be calculated as shown in equation (7) to determine the discretization point s. k The movement ds of (k = 1, ..., K) k By making the discretization point s k Moving amount ds k The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule.
[0132]
[0133] In equation (7), g(s) tk ) is the surrogate model used for model estimator 1b.
[0134] Implementation Method 2
[0135] In Embodiment 2, the optimization apparatus described below will be explained: the transformation rule update unit 4 uses a method based on candidate solution z t and evaluation value y t The surrogate function g(B) calculated separately is used to update the candidate solution z transformed by the variable transformation unit 1. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule.
[0136] The structure of the optimization device according to Embodiment 2 is the same as that of the optimization device according to Embodiment 1. The structural diagram of the optimization device according to Embodiment 2 is shown below. Figure 1 .
[0137] In the optimization apparatus according to Embodiment 2, the transformation rule update unit 4 uses the following formula (8) based on the candidate solution z t and evaluation value y t The surrogate function g(B) is calculated separately, and the discretization point s is calculated. k The movement ds of (k = 1, ..., K) k .
[0138]
[0139] In equation (8), ▽ is the mathematical symbol representing the gradient of the surrogate function g(B).
[0140] The transformation rule update section 4 makes the discretization point s k Moving amount ds k The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. tThe transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule.
[0141] Except for the conversion rule update unit 4, it is the same as the optimization device involved in Embodiment 1.
[0142] In the above-described embodiment 2, the optimization device is configured such that the transformation rule update unit 4 uses a method based on candidate solution z. t and evaluation value y t The surrogate function g(B) calculated separately is used to update the candidate solution z transformed by the variable transformation unit 1. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t Therefore, the optimization device can not only prevent the degradation of quantization accuracy when there is no discretization point consistent with the continuous variable, but also easily update the transformation rule and the inverse transformation rule separately.
[0143] In the optimization apparatus according to Embodiment 1, the conversion rule update unit 4 updates the evaluation value y. t The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule.
[0144] Furthermore, in the optimization apparatus according to Embodiment 2, the transformation rule update unit 4 updates the discretization point s by making the discretization point s k Moving amount ds k The transformation of z into candidate solutions by the variable transformation unit 1 is updated respectively. t The transformation rule and the inverse transformation performed by the variable inverse transformation part 2 to the second continuous variable x t The inverse transformation rule.
[0145] In any update, if the number of partitions of the discretized point sk is large, the search will take time even with annealing due to the large solution space.
[0146] As shown, it can also be configured by starting with a small group of discretized points and adding the required number of discretized points.
[0147] The additional discretization point positions utilize the position information r shown in equation (9). t The location information r t This is equivalent to moving the amount of movement ds. k Discretization point s k The location.
[0148]
[0149] The discretization points are appended to the candidate solution z calculated by sampler 1c. t The discretized point position s in the next step of the nearest neighbor is shown in equation (10). t+1 Identify the discretized point s k The symbol k is rearranged from the beginning to k = 0, ..., k, ... M.
[0150]
[0151] The transformation rule update section 4 calculates the discretization point s using the surrogate function g(B). k The movement ds of (k = 1, ..., K) k In the case of selecting additional discretization points, k can also be calculated as shown in equation (12). t new .
[0152]
[0153] Furthermore, this disclosure allows for free combination of various embodiments, modification of any constituent elements in various embodiments, or omission of any constituent elements in various embodiments.
[0154] Industrial utilization potential
[0155] This disclosure applies to optimization apparatus, optimization methods, and procedures.
[0156] Explanation of reference numerals in the attached figures
[0157] 1...Variable conversion unit; 1a...Encoder; 1b...Model estimator; 1c...Sampler; 2...Inverse variable conversion unit; 3...Evaluation value calculation unit; 3a...Evaluator; 3b...Storage unit; 4...Conversion rule update unit; 11...Variable conversion circuit; 12...Inverse variable conversion circuit; 13...Evaluation value calculation circuit; 14...Conversion rule update circuit; 21...Memory; 22...Processor.
Claims
1. An optimization device, characterized in that, have: The variable transformation unit acquires the first continuous variable as a variable represented by either a real value or an integer value, and transforms the first continuous variable into a candidate solution of a binary variable; The variable inverse transformation unit converts the candidate solution into a second continuous variable. The evaluation value calculation unit calculates the evaluation value of the second continuous variable; as well as The transformation rule updating unit updates the transformation rule for converting to the candidate solution performed by the variable transformation unit and the inverse transformation rule for converting to the second continuous variable performed by the variable inverse transformation unit, based on the evaluation value.
2. The optimization device according to claim 1, characterized in that, The process involves repeated transformation processing by the variable transformation unit, inverse transformation processing by the variable inverse transformation unit, evaluation value calculation processing by the evaluation value calculation unit, and update processing by the transformation rule update unit.
3. The optimization device according to claim 1, characterized in that, The variable transformation unit determines any one of the multiple discretization points as the candidate solution based on the distance between the multiple discretization points and the first continuous variable.
4. The optimization device according to claim 3, characterized in that, The transformation rule update unit updates the respective transformation rule and the inverse transformation rule, causing the discretization point to move.
5. The optimization device according to claim 1, characterized in that, The transformation rule update unit updates the transformation rule and the inverse transformation rule based on the second continuous variable and the evaluation value, respectively.
6. The optimization device according to claim 1, characterized in that, The transformation rule update unit uses surrogate functions calculated based on the candidate solutions and the evaluation values to update the transformation rule and the inverse transformation rule respectively.
7. An optimization method, characterized in that, The variable transformation unit acquires the first continuous variable as a variable represented by either a real value or an integer value, and transforms the first continuous variable into a candidate solution of a binary variable. The variable inverse transformation unit inversely transforms the candidate solution into a second continuous variable. The evaluation value calculation unit calculates the evaluation value of the second continuous variable. Based on the evaluation value, the transformation rule update unit updates the transformation rule for converting to the candidate solution performed by the variable transformation unit and the inverse transformation rule for converting to the second continuous variable performed by the variable inverse transformation unit.
8. A program for causing a computer of an optimization apparatus to execute, the optimization apparatus converting a first continuous variable, which is represented as a variable by either a real value or an integer value, into a binary variable. Its features are, The program causes the computer to execute: The first process involves obtaining the first continuous variable and converting it into a candidate solution for a binary variable. The second process involves converting the candidate inverse solution into a second continuous variable. The third step is to calculate the evaluation value of the second continuous variable; as well as The fourth process involves updating, based on the evaluation value, the transformation rule from the first process to the candidate solution and the inverse transformation rule from the second process to the second continuous variable.
Citation Information
Patent Citations
Optimization device, optimization system, method for optimization, and program
JP2020113190A
