Intelligent control limited precision quantum optimization hardware closed loop mapping maintenance evaluation method

Abstract

The application provides a limited precision quantum optimization hardware closed loop mapping maintaining evaluation method for intelligent control, comprising the following steps: S1, obtaining a QUBO coefficient matrix to be deployed from an optimization decision module of an intelligent control system; S2, determining an initial mapping parameter vector in three dimensions of coefficient dynamic range constraint, constraint-target level constraint and hardware precision compatibility constraint on the QUBO coefficient matrix; S3, determining the QUBO coefficient matrix according to the mapping parameter, and performing scaling, rounding and limiting operations on the QUBO coefficient matrix to generate an integer coefficient matrix; S4, uploading and sampling the integer coefficient matrix through a hardware interface module; S5, calculating a pair-wise order relation maintaining rate of a candidate solution sample and performing consistency threshold comparison, when the pair-wise order relation maintaining rate is less than the consistency threshold, entering step S6; S6, performing closed loop feedback adjustment and returning to step S3 to be re-performed until the pair-wise order relation maintaining rate is greater than or equal to the consistency threshold, and the output result is good in reliability and consistency.

Description

Technical Field

[0001] This invention relates to the fields of quantum computing and intelligent control technology, specifically to a finite-precision quantum optimization method for hardware closed-loop mapping preservation evaluation in intelligent control. Background Technology

[0002] With the rapid development of intelligent technologies, the application of finite-precision combinatorial optimization hardware in intelligent control fields such as power system dispatching, combinatorial optimization, and machine learning has gradually gained widespread attention. Quantum heuristic optimization hardware includes coherent Ising machines, quantum annealing machines, and digital annealing machines. When deploying quadratic unconstrained binary optimization (QUBO) models on finite-precision combinatorial optimization hardware (such as coherent Ising machines (CIM) and quantum annealing machines (QA), the combinatorial optimization problem is generally solved by encoding it into a QUBO model or an Ising model. However, the coefficient representation precision of quantum heuristic optimization hardware is limited, typically 4-8 bit integer quantization, while the original coefficients of QUBO models are usually floating-point numbers with a large dynamic range. Therefore, before submitting the QUBO model to the hardware, precision mapping must be performed to convert the floating-point coefficients into an integer format that the hardware can represent.

[0003] Traditional precision mapping methods mainly include fixed linear scaling, truncated quantization, adaptive quantization, and mixed precision schemes. Fixed linear scaling selects a single scaling factor and multiplies all coefficients by the single scaling factor and then rounds them. Truncated quantization directly truncates or rounds the coefficients at fixed points without scaling. Adaptive quantization selects quantization parameters based on the statistical characteristics (mean, variance) of the coefficient distribution. Mixed precision schemes use different quantization strategies for large and small coefficients.

[0004] However, existing precision mapping methods have the following drawbacks. First, all existing methods are open-loop one-time operations, meaning that the mapping parameters are determined based on the statistical characteristics of the coefficient matrix before hardware execution, and are not adjusted after execution. This paradigm has the following structural defects: in the case of 4-bit precision mapping, the consistency index R=1.0, which is considered perfect preservation, while R≈0.72 for 6-bit precision mapping is considered significant degradation, and R≈0.85 for 10-bit precision mapping is considered partial recovery. This means that the intuitive assumption that "increasing precision can improve the result" does not hold, and open-loop one-time mapping cannot guarantee finding the optimal precision configuration.

[0005] Second, in the QUBO coefficient matrix, the coefficients of the constraint penalty term and the coefficients of the economic objective term coexist. The choice of scaling factor not only affects the absolute representation accuracy of the coefficients, but also changes the energy hierarchy relationship between the constraint term and the objective term. Fixed linear scaling cannot perceive this structural coupling.

[0006] Third, existing methods do not perform consistency evaluation on the hardware output after mapping and submission. It is impossible to know whether the candidate solutions returned by the hardware maintain the energy ranking structure of the original QUBO model, nor can it be determined whether the mapping parameters need to be adjusted and the solution resolved.

[0007] Fourth, in the optimization decision-making module of the intelligent control system, since the solution results of the quantum optimization hardware directly drive the generation of control commands, if the output of the quantum optimization hardware is distorted due to improper precision mapping, the reliability of the control decision cannot be guaranteed.

[0008] Therefore, the above-mentioned technologies lack a technical solution for closed-loop feedback iterative adjustment of the accuracy mapping process.

[0009] For example, patent document with patent application number 202510296821.2 and publication date of December 2, 2025 discloses a feedback iterative method for allocating optical network resources by fusing quantum and classical computing. The method includes: applying a heuristic algorithm to preprocess optical network resources and establishing a QUBO model for the corresponding resource allocation problem; using the candidate optical paths and candidate resource information as binary logical variables and mapping the resource allocation scheme into Hamiltonians of different sizes according to constraints; solving the QUBO model in a dedicated quantum computer using a heuristic algorithm and obtaining the resource allocation scheme after post-processing.

[0010] The above literature describes a process where, when a resource allocation scheme satisfying the constraints is not obtained in a single solution using a dedicated quantum computer, a heuristic algorithm is applied to tune the parameters based on the solution results. The solution is then applied again using the dedicated quantum computer until a resource allocation scheme satisfying the constraints is obtained or the upper limit of the number of iterations is reached. When the constraints of optical path and wavelength / spectrum are not satisfied and the upper limit of the number of iterations has not been reached, the parameters of the QUBO coefficient matrix or the parameters of the dedicated quantum computer are adjusted, and the solution is solved again. The iterative mechanism mainly adjusts the constraint penalty term coefficients in the QUBO coefficient matrix or the operating parameters of the dedicated quantum computer. Its purpose is to ensure that the solution results satisfy physical constraints such as routing, wavelength / spectrum, and signal-to-noise ratio. The feedback iteration is merely an iteration at the solver parameter tuning level, rather than a closed-loop optimization at the precision mapping parameter level. The method does not consider the iterative optimization of the precision mapping process from floating-point representation to hardware integer representation of the QUBO coefficient matrix. For those skilled in the art, precision mapping is a deterministic format conversion operation. Given the coefficient range and hardware bit width, the mapping parameters can be uniquely determined by the formula without iteration. Increasing the hardware precision bit width will inevitably improve the solution quality, and there is no non-monotonic effect. The hardware output quality is determined by the hardware physical characteristics, and software-level adjustment of mapping parameters cannot substantially improve it. These factors make it impossible for those skilled in the art to think of iteratively optimizing the precision mapping process. It also does not involve preprocessing steps such as coefficient dynamic range compression, constraint-target hierarchy preservation, and consistency evaluation in the precision mapping process, thus leading to unreliable precision and poor control consistency. Summary of the Invention

[0011] This invention provides a finite-precision quantum optimization hardware closed-loop mapping maintenance evaluation method for intelligent control. By transforming the precision mapping from an open-loop format to a closed-loop iterative optimization process, the mapping parameters are adjusted based on the evaluation results until the consistency index meets the requirements, thus ensuring the reliability of intelligent control decisions.

[0012] To achieve the above objectives, the technical solution provided by this invention is: a finite-precision quantum optimization method for hardware closed-loop mapping preservation evaluation in intelligent control, comprising the following steps: S1. Obtain the QUBO coefficient matrix to be deployed and the coefficient bit width, matrix size and sampling number of the target finite precision combination optimization hardware from the optimization decision module of the intelligent control system. S2. For the QUBO coefficient matrix, the initial mapping parameter vector is determined jointly based on three dimensions: coefficient dynamic range constraint, constraint-target level constraint, and hardware precision compatibility constraint. The hardware precision compatibility constraint determines the feasible range of the scaling factor based on the coefficient bit width and matrix size. The initial mapping parameter vector includes the scaling factor, effective bit width selection, and dynamic range pruning threshold. The mapping parameters are set to the initial preset parameter vector. S3. Determine the QUBO coefficient matrix according to the mapping parameters, and perform scaling, rounding and limiting operations on the QUBO coefficient matrix to generate an integer coefficient matrix. Perform basic integrity verification on the mapped matrix. S4. Submit the integer coefficient matrix to the finite precision combinatorial optimization hardware through the hardware platform's native matrix upload interface, obtain the corresponding number of candidate solution samples and corresponding hardware energy values ​​according to the number of samplings, and maintain the original n-dimensional QUBO variable structure. S5. Calculate the hardware energy value of the original floating-point QUBO model and the difference between the corresponding hardware energy values ​​of the candidate sample to determine the pairwise order retention rate. Compare the pairwise order retention rate with the consistency threshold. If the pairwise order retention rate is less than the consistency threshold, proceed to step S6; otherwise, proceed to step S7. S6. When the pairwise order retention rate is less than the consistency threshold, perform closed-loop feedback adjustment, determine the diagnosis result based on the pairwise order relationship, adjust the mapping parameters based on the diagnosis result, and return to step S3 until the pairwise order retention rate is greater than or equal to the consistency threshold. S7. Output an integer coefficient matrix for use by the intelligent control system's optimization decision-making module.

[0013] The above method first obtains the QUBO coefficient matrix in the optimization decision module of the intelligent control system. Then, it determines an optimal initial mapping parameter vector by considering the intersection of three constraints: coefficient dynamic range, constraint-target level, and hardware precision compatibility. The hardware precision compatibility constraint requires determining a feasible range for a scaling factor based on the coefficient bit width and matrix size. This ensures that the scaling factor and effective bit width are the initial optimal solutions considering the coefficient dynamic range, constraint-target level, and hardware precision compatibility, avoiding problems caused by fixed linear scaling. Finally, it determines the QUBO coefficient matrix based on the mapping parameters to form integer QUBO coefficients. The matrix is ​​used to sample hardware energy values ​​after uploading the integer QUBO coefficient matrix. Then, the pairwise order retention rate is determined by comparing the energy values ​​corresponding to the integer QUBO coefficient matrix with the floating-point QUBO coefficient matrix corresponding to the non-integer values. The pairwise order retention rate is used as a consistency evaluation index to directly measure whether the hardware output maintains the ranking of candidate solutions of the original model. The physical meaning is clear. If the consistency requirements cannot be met, the mapping parameters are readjusted and the consistency evaluation is carried out again. Through closed-loop feedback iteration, the precision mapping is transformed from an open-loop format to a closed-loop optimization process. The non-monotonic precision effect is used to find the counterintuitive optimal configuration, ensuring the reliability of intelligent control decisions.

[0014] Furthermore, in step S1, the QUBO coefficient matrix Q is an n×n real symmetric matrix, and the coefficient bit width B is an integer range [-2]. (B-1) , 2 (B-1) -1].

[0015] The above settings clarify the mathematical form of the QUBO coefficient matrix and the range of hardware bit width, providing a clear definition of input parameters for subsequent precision mapping, thus giving the method a clear parameter basis and hardware compatibility prerequisite.

[0016] Furthermore, step S2 includes S2.1-S2.3: S2.1 Coefficient Dynamic Range Constraints: Calculate the maximum absolute value of the coefficient matrix max|Q_ij|, the minimum non-zero absolute value min|Q_ij|≠0, and the dynamic range DR as shown in formula (1). DR=2log2(max|Q_ij| / min|Q_ij|)(1), min|Q_ij|≠0, where Q_ij is the value of the QUBO coefficient matrix at position (i,j), which determines the proportion of coefficients that can be retained under the B-bit representation of coefficients; S2.2 Constraints – Target Level Constraints: Identify the constraint penalty component Q_constraint and the target component Q_objective in the QUBO coefficient matrix, ensure that the energy level of the constraint penalty component is higher than the energy level of the target component after mapping, and determine the mapped coefficients that meet the conditions; S2.3 Hardware precision compatibility constraint: Based on the hardware bit width B and matrix size n, determine the feasible interval [s_min, s_max] of the scaling factor s, where s_min and s_max are the minimum and maximum values ​​of the scaling factor s, respectively, so that the coefficients after mapping neither overflow nor all return to zero.

[0017] The above settings, by jointly determining the mapping parameters from three dimensions—dynamic range of coefficients, constraint-target hierarchy, and hardware precision compatibility—incorporate the numerical characteristics of the coefficient matrix and the constraint structure into the decision-making process of the mapping parameters. This ensures that the initial mapping parameters maintain the constraint-target hierarchy while satisfying hardware physical constraints, avoiding the problems of "constraints retained but targets lost" or "targets retained but constraints collapsed" in fixed linear scaling.

[0018] Furthermore, the integer coefficient matrix Q'_ij in step S3 is shown in formula (2). Q'_ij=clip(round(s) Q_ij), -2 (B-1) , 2 (B-1) -1)(2), round() is for rounding to the nearest integer, and clip(·,lo,hi) is for clipping the input, where lo is the lower threshold and hi is the upper threshold.

[0019] The above settings, through a unified quantization mapping formula, perform scaling, rounding, and limiting operations on all coefficient matrices using scaling factors, ensuring the reliability of the certificate coefficient matrix in terms of rounding.

[0020] Furthermore, step S3 also includes: performing a basic integrity check on the integer matrix. If the conditions for not meeting the basic integrity check are found, the integer coefficient matrix is ​​re-obtained through step S2. The basic integrity check includes: the number of non-zero coefficients, whether the constraint penalty term coefficients maintain the hierarchical relationship, and whether there are rows or columns with all zeros.

[0021] The above settings perform an integrity check after mapping to ensure that the mapped integer coefficient matrix retains sufficient coefficient information and constraint structure, thus avoiding the matrix from degenerating into an invalid matrix due to excessive compression.

[0022] Further, step S4 includes: setting the number of samplings to N_sample, submitting the integer coefficient matrix to the finite precision combinatorial optimization hardware through the hardware platform's native matrix upload interface to obtain N_sample candidate solution samples {x(1),x(2), ... x(N_sample)} and the corresponding hardware energy values ​​{E_hw(1), E_hw(2), ... E_hw(N_sample)}, without performing variable compression or subgraph embedding, and maintaining the original n-dimensional QUBO variable structure.

[0023] The above settings obtain multiple candidate solution samples and their energy values ​​by configuring the number of samplings, providing sufficient statistical samples for the subsequent calculation of the pairwise order preservation rate, ensuring the statistical significance and reliability of the consistency evaluation index, while also avoiding the compression of variables and affecting the QUBO variable structure, thus ensuring the reliability of the QUBO coefficient matrix variables.

[0024] Furthermore, the pairwise order preservation rate R in step S5 is shown in formula (3). R = Number of samples that maintain the order / Total number of samples (3) The definition of the sample logarithm that preserves the order relation is as follows: From N_sample candidate solutions, randomly select two different samples (x(a), x(b)) and calculate their energy difference ΔE_float on the original floating-point QUBO model, as shown in Equation (3.1). ΔE_float=E_float(x^(a))-E_float(x^(b)(3.1),E_float represents the energy value of the original floating-point coefficient matrix; The energy difference ΔE_hw calculated on the hardware integer coefficient matrix is ​​shown in equation (3.2). ΔE_hw=E_hw(x(a))-E_hw(x(b)) (3.2), where E_hw represents the energy value of the integer coefficient matrix; If sign(ΔE_float)=sign(ΔE_hw), the sample pair is regarded as maintaining the order relation; sign() is used to determine the direction of energy change, such as increase, decrease or no change; The total number of sample pairs is shown in formula (3.3) Total number of sample pairs=C(N_sample,2)=N_sample×(N_sample-1) / 2 (3.3), where C(N_sample, 2) is the number of combinations for selecting 2 elements from N_sample different elements, R∈[0,1], where R=1 indicates that the hardware output perfectly maintains the energy ranking structure of the original model; R<1 indicates that there is ranking distortion.

[0025] With the above arrangement, by defining the pairwise order relation maintenance rate, the energy ranking consistency between the hardware output and the original floating-point model is quantified as a comparable value between 0 and 1, so that the evaluation of hardware output quality is upgraded from qualitative judgment to quantitative evaluation, with clear physical meaning and no need to rely on additional calibration data.

[0026] Further, the step of "comparing the pairwise order relation maintenance rate with the consistency threshold" in step S5 includes: setting a consistency threshold R_th, when R≥R_th, determining that the hardware output consistency under the current mapping parameters meets the requirements, and proceeding to step S7; when R<R_th, determining that the consistency does not meet the requirements, and proceeding to step S6.

[0027] With the above arrangement, by introducing the consistency threshold, a clear quantitative boundary is provided for the judgment of mapping quality, so that the system can automatically identify scenarios where the mapping quality does not meet the requirements, thereby triggering closed-loop feedback iteration and avoiding the subjectivity of relying on manual experience judgment.

[0028] Further, step S6 includes S6.1 to S6.4: S6.1 Distortion mode diagnosis: analyze the features of sample pairs that violate the maintained order relation, and identify the distortion source: if the violations are concentrated in sample pairs with small energy difference |ΔE_float|, the diagnosis result is determined as "insufficient precision distortion", and the adjustment direction is to change the scaling strategy or dynamic range clipping point; if the features of the sample pairs that violate the maintained order relation are scattered in each energy difference interval, the diagnosis result is determined as "structure-coupled distortion", the adjustment direction is to change the scaling strategy or dynamic range clipping point, and then proceed to step S6.2; S6.2 Mapping parameter adjustment: According to the diagnosis result, generate the adjusted mapping parameter θ_{k+1}, which specifically includes: the adjustment of the scaling factor s_{k+1} is shown in formula (4); s_{k+1}=s_k×(1±δ_s) (4), δ_s is the preset step size; The adjustment of the effective bit width B_{k+1} is shown in Formula (4.1), B_{k+1}=B_k±1 (4.1); The effective bit width is within the hardware support range, The dynamic range clipping adjustment R_dynamic_{k+1} is expanded or narrowed according to the distortion condition by the preset adjustment value.

[0029] S6.3 Iterative execution: Return to step S3 with θ_{k+1} to re-execute precision mapping until the termination condition is met to terminate the iteration. The termination condition includes any one of the following conditions: R≥R_th; the number of iterations k≥k_max, where k_max is the preset maximum number of iterations; the variation of R in two consecutive iterations |ΔR|<ε_R, where ε_R is the preset variation threshold.

[0030] S6.4 Iteration record: When the iteration is terminated, record the mapping parameter θ_k, consistency index R_k, and distortion pattern diagnosis result of each iteration to form a mapping process file for reference by the decision reliability assessment module of the intelligent control system.

[0031] With the above arrangement, the order relation violation phenomenon is classified into the insufficient precision type and the structural coupling type through distortion pattern diagnosis, and differentiated mapping parameter adjustment strategies are adopted for different distortion patterns, so that the closed-loop iteration has directionality and purposiveness, avoiding blind parameter adjustment. At the same time, a mapping process file is formed through iteration records, which provides traceable data support for the decision reliability assessment of the intelligent control system.

[0032] Further, step S7 includes: outputting the optimal mapping parameter θ and the corresponding integer coefficient matrix Q'; the hardware candidate solution sample set and the corresponding original floating-point energy re-assessment value; the final consistency index R and the mapping process file; R ≥R_th is marked as "trusted", R <R_th is marked as "for reference only", and the result output is for use by the optimization decision module of the intelligent control system as a basis for generating control instructions.

[0033] With the above arrangement, by attaching a reliability label of "trusted or for reference only" to the final output result, the optimization decision module of the intelligent control system can select or adjust the subsequent control strategy according to the trust degree of the solution result, avoiding generating control instructions based on untrusted results when the mapping quality is poor, and providing a quantitative basis for the result credibility of intelligent control decisions. Brief Description of the Drawings

[0034] Figure 1 This is a flowchart of the present invention.

[0035] Figure 2 This is a schematic diagram illustrating the determination of the joint constraint mapping scheme of the present invention.

[0036] Figure 3 The figure shows the experimental results of the non-monotonic precision effect of the present invention.

[0037] Figure 4 This is a schematic diagram of the closed-loop iterative convergence process of the present invention.

[0038] Figure 5 This is a system architecture diagram of the present invention. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0040] Example 1.

[0041] like Figures 1 to 5 As shown, a finite-precision quantum-optimized hardware closed-loop mapping preservation evaluation method for intelligent control includes the following steps: S1. Obtain the QUBO coefficient matrix to be deployed and the coefficient bit width, matrix size and sampling number of the target finite precision combination optimization hardware from the optimization decision module of the intelligent control system. S2. For the QUBO coefficient matrix, the initial mapping parameter vector is determined jointly based on three dimensions: coefficient dynamic range constraint, constraint-target level constraint, and hardware precision compatibility constraint. The hardware precision compatibility constraint determines the feasible range of the scaling factor based on the coefficient bit width and matrix size. The initial mapping parameter vector includes the scaling factor, effective bit width selection, and dynamic range pruning threshold. The mapping parameters are set to the initial preset parameter vector. S3. Determine the QUBO coefficient matrix according to the mapping parameters, and perform scaling, rounding and limiting operations on the QUBO coefficient matrix to generate an integer coefficient matrix. Perform basic integrity verification on the mapped matrix. S4. Submit the integer coefficient matrix to the finite precision combinatorial optimization hardware through the native matrix upload interface of the hardware platform, obtain the corresponding number of candidate solution samples and corresponding hardware energy values ​​according to the preset number of samplings, maintain the original n-dimensional QUBO variable structure, and do not perform variable compression or subgraph embedding. S5. Calculate the hardware energy value of the original floating-point QUBO model and the difference between the corresponding hardware energy values ​​of the candidate solutions to determine the pairwise order retention rate. Use the pairwise order retention rate as a consistency evaluation index and compare it with the consistency threshold. When the pairwise order retention rate is less than the consistency threshold, proceed to step S6. When the pairwise order retention rate is greater than or equal to the consistency threshold, proceed to step S7. S6. Perform closed-loop feedback adjustment, determine the diagnosis result based on the pairwise order relationship and adjust the mapping parameters based on the diagnosis result, and return to step S3 until the pairwise order relationship retention rate is greater than or equal to the consistency threshold. S7 outputs an integer coefficient matrix and mapping parameters for use by the intelligent control system's optimization decision-making module.

[0042] In step S1, the QUBO coefficient matrix Q is an n×n real symmetric matrix, and the coefficient bit width B can represent the integer range [-2]. (B-1) ,2 (B-1) -1].

[0043] In one embodiment, the QUBO coefficient matrix is ​​derived from the power system unit combination and scheduling problem, such as n=363, and the hardware is the Coherent Ising Machine (CIM). The B=4, 6, 8, 10, and 12-bit coefficient representations are tested respectively.

[0044] Step S2 includes S2.1-S2.3: S2.1 Coefficient Dynamic Range Constraints: Calculate the maximum absolute value max|Q_ij|, the minimum non-zero absolute value min|Q_ij|≠0, and the dynamic range DR of the QUBO coefficient matrix as shown in formula (1). DR = 2log2(max|Q_ij| / min|Q_ij|)(1), min|Q_ij|≠0, where Q_ij is the value of the QUBO coefficient matrix at position (i,j), determines the proportion of coefficients that can be retained under the B-bit representation of coefficients; and a dynamic range cutoff point is selected based on the dynamic range so that the cutoff region is within the dynamic range, thus determining the dynamic range cutoff point; S2.2 Constraints – Target Level Constraints: Identify the constraint penalty component Q_constraint and the target component Q_objective in the QUBO coefficient matrix. Ensure that the energy level of the constraint penalty component remains higher than that of the target component after mapping, and determine the coefficients that satisfy the conditions after mapping. Specifically, this can be achieved by adjusting the penalty coefficients of the penalty components. By selecting different penalty coefficients, the energy level of the constraint penalty component after mapping can be higher than that of the target component. If the penalty coefficient is too small, the solver may choose a solution that violates the constraints in order to reduce the value of the objective function; if it is too large, it may overwhelm the differences in the objective function, making it difficult to find the optimal feasible solution. That is, the mapping does not destroy the energy landscape structure of "constraint satisfaction takes precedence over objective optimization". S2.3 Hardware Precision Compatibility Constraints: Based on the hardware bit width B and matrix size n, determine the feasible interval [s_min, s_max] of the scaling factor s, where s_min and s_max are the minimum and maximum values ​​of the scaling factor s, respectively, ensuring that the mapped coefficients do not overflow (|s·Q_ij| ≤ 2). (B-1) -1) are not all zero, but overall, the coefficients after mapping can satisfy |s·Q_ij| ≥1.

[0045] In step S3, the integer coefficient matrix Q'_ij is shown in formula (2). Q'_ij = clip(round(s) Q_ij), -2 (B-1) , 2 (B-1) -1)(2), `round()` rounds the input to the nearest integer, and `clip(·, lo, hi)` clips the input, where `lo` is the lower threshold and `hi` is the upper threshold. Then, a basic integrity check is performed on the integer matrix. If the basic integrity check conditions are not met, the matrix is ​​deemed invalid, and the process returns to step S2 to re-obtain the coefficient matrix. If no conditions are met, step S4 is performed. The basic integrity check includes: the number of non-zero coefficients, whether the constraint penalty coefficients maintain a hierarchical relationship, and whether there are rows or columns with all zeros. The number of non-zero coefficients can be obtained by counting the number of non-zero values ​​in the coefficients. Whether the constraint penalty coefficients maintain a hierarchical relationship can be determined by comparing the existing constraint penalty coefficients with the preset hierarchical relationship. Whether there are rows or columns with all zeros can be determined by detecting whether there is a row or column with all zero coefficients in the integer coefficient matrix.

[0046] In step S4, the number of samples is set to N_sample. The integer coefficient matrix is ​​submitted to the finite-precision combinatorial optimization hardware through the hardware platform's native matrix upload interface to obtain N_sample candidate solution samples {x(1), x(2), ... x(N_sample)} and their corresponding hardware energy values ​​{E_hw(1), E_hw(2), ... E_hw(N_sample)}. Variable compression or subgraph embedding is not performed, and the original n-dimensional QUBO variable structure is maintained. In this embodiment, the number of samples N_sample = 1000.

[0047] The pairwise order preservation rate R in step S5 is shown in formula (3). R = Number of sample pairs that maintain order / Total number of sample pairs (3) The definition of the number of sample logs that maintain the order relationship is as follows: Take two different samples (x^(a), x^(b)) from N_sample candidate solutions and calculate their energy difference ΔE_float on the original floating-point QUBO model as shown in formula (3.1). The original floating-point QUBO model is the QUBO coefficient matrix to be deployed. ΔE_float= E_float(x^(a)) - E_float(x^(b))(3.1), E_float represents the energy value of the original floating-point coefficient matrix; The energy difference ΔE_hw calculated on the hardware integer coefficient matrix model is shown in equation (3.2). ΔE_hw=E_hw(x^(a))-E_hw(x^(b))(3.2), E_hw represents the energy value of the integer coefficient matrix; If sign(ΔE_float) = sign(ΔE_hw) (i.e., the two models agree on the merits of this pair of candidate solutions), then the sample pair is said to maintain an order relationship. sign() is used to determine the direction of energy change, such as increase, decrease or no change. The total sample logarithm is shown in formula (3.3); Total number of sample pairs = C(N_sample, 2) = N_sample × (N_sample - 1) / 2 (3.3), where C(N_sample, 2) is the number of combinations of selecting 2 elements from N_sample distinct elements. Where R ∈ [0, 1], R = 1 indicates that the hardware output perfectly preserves the energy ranking structure of the original model; R<1 indicates that there is ranking distortion, that is, the solution that the hardware considers "good" is not necessarily better in the original model.

[0048] "Consistency determination" in step S5 includes: setting a consistency threshold R_th (typical value 0.90-0.95). When R ≥ R_th, it is determined that the hardware output consistency under the current mapping parameters meets the requirements, and the process proceeds to result output; when R < R_th, it is determined that the consistency does not meet the requirements, and the closed-loop feedback iteration of step S6 is triggered.

[0049] Step S6 includes S6.1 to S6.4: S6.1 Distortion pattern diagnosis: analyze the features of sample pairs that violate the order preservation relationship, and identify the distortion source: if the features of the sample pairs that violate the order preservation relationship are concentrated in sample pairs with a small energy difference |ΔE_float|, i.e., near-degenerate pairs, the diagnosis result is determined as "insufficient precision distortion", and the adjustment direction is to change the scaling strategy or the dynamic range clipping point; perform increment processing on the basis of the current effective hardware precision according to a preset increment amount each time; if the features of the sample pairs that violate the order preservation relationship are scattered in each energy difference interval, the diagnosis result is determined as "structure-coupled distortion", and the adjustment direction is to change the scaling strategy or the dynamic range clipping point, and changing the scaling strategy or the dynamic range clipping point is performed according to step S6.2; S6.2 Mapping parameter adjustment: According to the diagnosis result, generating the adjusted mapping parameter θ_{k+1} specifically includes: adjusting the scaling factor s_{k+1} as shown in formula (4); s_{k+1}= s_k × (1 ± δ_s) (4), δ_s is a preset step size, with a value range of 0.1-0.3; The effective bit width B_{k+1} is adjusted as shown in formula (4.1), B_{k+1} = B_k ± 1 (4.1); where the effective bit width is within the hardware support range, The dynamic range clipping adjustment R_dynamic_{k+1} is expanded or narrowed according to the distortion situation by a preset adjustment value.

[0050] In this embodiment, for the B0=8-bit configuration, s0=128 / max|Q|, and R_dynamic0 is initially set to no clipping.

[0051] S6.3 Iterative execution: return to step S3 with θ_{k+1} to re-execute precision mapping until the iteration terminates when a termination condition is satisfied. The termination condition includes any one of the following conditions: R ≥ R_th; the number of iterations k ≥ k_max, where k_max is the maximum number of iterations, typically 3-5 times; the variation |ΔR| of R between two consecutive iterations is less than ε_R, where ε_R is a preset variation threshold, and the current optimal result is output when convergence is reached but the standard is not met. In this embodiment, k_max=5; when the variation |ΔR| of R between two consecutive iterations is less than ε_R, it is determined that convergence is reached but the standard is not met, and the current optimal result is output.

[0052] S6.4 Iteration record: when the iteration terminates, the mapping parameter θ_k, consistency index R_k, and distortion pattern diagnosis result of each iteration are recorded to form a mapping process file for reference by the decision reliability assessment module of the intelligent control system.

[0053] Step S7 includes: when R≥R_th in step S5, output the optimal mapping parameter θ and the corresponding integer coefficient matrix Q'_ij; the hardware candidate solution sample set and the corresponding original floating-point energy re-evaluation value; the final consistency index R and the mapping process file; R ≥ R_th is marked as "trusted", R <R_th is marked as "for reference only", and the output result is provided to the optimization decision module of the intelligent control system as a basis for generating control instructions. In this embodiment, the final R =1.0, marked as "trusted".

[0054] The present invention can be applied to the optimization decision module in an intelligent control system, and targets the upstream calculation link of the optimization decision module in the intelligent control system, performs consistency evaluation and reliability marking on the quantum optimization hardware solution result, so that the control system can select or adjust the subsequent control strategy according to the trust degree of the solution result, and provide reliability guarantee for the hardware solution result in intelligent control scenarios requiring real-time optimization decision-making such as power dispatching and energy management.

[0055] The working principle of this invention is as follows: First, the QUBO coefficient matrix in the optimization decision module of the intelligent control system is obtained. Then, an optimal initial mapping parameter vector is determined by the intersection of three aspects: coefficient dynamic range constraint, constraint-target level constraint, and hardware precision compatibility constraint. The hardware precision compatibility constraint requires determining a feasible range of scaling factors through the coefficient bit width and matrix size. This ensures that the scaling factor and effective bit width are the initial optimal solutions after considering the coefficient dynamic range, constraint-target level, and hardware precision compatibility, avoiding problems caused by fixed linear scaling. Then, the QUBO coefficient matrix is ​​determined according to the mapping parameters to form an integer QUBO. The coefficient matrix is ​​generated, and after uploading the integer QUBO coefficient matrix, the hardware energy value is sampled. Then, the pairwise order retention rate is determined by comparing the energy value corresponding to the integer QUBO coefficient matrix with the floating-point QUBO coefficient matrix corresponding to the non-integer values. The pairwise order retention rate is used as a consistency evaluation index to directly measure whether the hardware output maintains the ranking of candidate solutions of the original model. The physical meaning is clear. When the consistency requirement cannot be met, the mapping parameters are readjusted and the consistency evaluation is carried out again. Through closed-loop feedback iteration, the precision mapping is transformed from an open-loop format to a closed-loop optimization process. The non-monotonic precision effect is used to find the counterintuitive optimal configuration, ensuring the reliability of intelligent control decisions.

[0056] Example 2.

[0057] This embodiment uses the power system unit combination scheduling problem to verify that improving accuracy necessarily improves consistency, thus demonstrating the necessity of closed-loop iteration. QUBO source: Power system unit combination scheduling problem, W3 coding structure (symbolic state coding), n=363 variables; Hardware: Coherent Ising Machine (CIM), supporting 4 / 6 / 8 / 10 / 12-bit coefficient representation; Sampling: N_sample=1000 candidate solutions are collected for each bit width configuration.

[0058] like Figure 4 As shown, the experimental data are as follows: The bit width is 12-bit, the scaling factor is 2048 / max||Q_ij|, and R=0.870; The bit width is 10-bit, the scaling factor is 512 / max||Q_ij|, and R=0.850; The bit width is 8-bit, the scaling factor is 128 / max||Q_ij|, and R=0.780; The bit width is 6-bit, the scaling factor is 32 / max||Q_ij|, and R=0.720; With a bit width of 4 bits, a scaling factor of 8 / max||Q_ij|, and R=1.000, consistency is achieved. This is demonstrated by the above experiments, where R=1.0 ≥ R_th=0.90 is achieved in a 4-bit configuration. However, this condition is not met with 6-bit and 8-bit bit widths. In an open-loop method (fixed 8 bits), R=0.78, and the user cannot know that the result is unreliable. This embodiment proves that maintaining a constant pairwise order retention rate cannot reliably find a deployable configuration that meets consistency requirements.

[0059] Example 3.

[0060] QUBO originates from the IEEE 33-node power distribution system scheduling problem. W3 encoding is symbolic state encoding, n=363; the hardware is CIM, with an initial bit width B0=8; the threshold R_th=0.90, k_max=5; and the sampling N_sample=1000.

[0061] In this embodiment, as Figure 5 As shown, the complete iterative process is as follows: The iteration 0-bit width B0=8-bit, scaling factor s0=128 / max||Q_ij|, R=0.780, diagnosed as precision-deficient distortion; Iteration 1 - Bit width B1 = 6-bit, scaling factor s1 = 32 / max||Q_ij|, R = 0.720, diagnosed as precision-deficient distortion; The iteration uses a 2-bit width B2 = 4 bits, a scaling factor s2 = 8 / max||Q_ij|, and R = 1.000. Consistency is achieved, meaning the above experiments demonstrate that under a 4-bit configuration, R = 1.0 ≥ R_th = 0.90, the loop terminates, and the reliability label is output as "trustworthy". Thus, the optimal configuration can be found after 3 iterations. If in the open-loop method (fixed 8 bits), R = 0.78, the user cannot know that the result is untrustworthy. This invention can find the most suitable deployable configuration that meets consistency requirements through closed-loop iteration.

[0062] Example 4.

[0063] This embodiment is used to verify the technical effectiveness of the joint constraint analysis in step S2. The QUBO coefficient matrix includes constraint penalty components (penalty coefficient λ_PB=100, coefficient range [50, 800]) and economic target components (coefficient range [1, 45]); the hardware bit width B=8 (representation range [-128, 127]). The comparison schemes are shown in the table below: Scheme A (fixed linear scaling): scaling factor s=127 / 800=0.159, after mapping constraint term [8,127], uniform scaling of all coefficients, the minimum coefficient of the target term 1×0.159=0.159→0 (return to zero), the maximum coefficient of the target term 45×0.159=7.155→7, R=0.65.

[0064] Scheme B (S2 Joint Constraint Analysis of the Invention): Scaling factor s = 127 / 800 = 0.159, constraint condition is s·max(Q_objective) ≥ 1, thus the maximum coefficient of the objective term is selected as 45 × 0.159 = 7.155 → 7, that is, s ∈ [127 / 800, 1 / 1] = [0.159, 1]. Select s = 0.159 and perform dynamic range clipping (cut the maximum coefficient of the constraint term 800 to 600). After mapping, the constraint term is [8, 127], the objective term is [1, 7], and R = 0.92.

[0065] Option A scales all coefficients uniformly, reducing the minimum coefficient of the target item to zero and thus losing its economic differentiation capability. Option B identifies the minimum coefficient of the target item through joint constraint analysis, performs dynamic range pruning after scaling to release the representation space for the target item, and maintains the economic differentiation capability.

[0066] Example 5.

[0067] This embodiment is used to verify the convergence efficiency of closed-loop iteration on large-scale problems. The experiment uses 300 different encoding-precision configurations (W1 / W2 / W3 encoding × 4-12 bits × 6 types of penalty weights × 2 types of scaling × 2 types of random seeds). Only 4 configurations achieve R ≥ 0.90 (reserved configurations), and 296 configurations have R < 0.90 (degenerate configurations).

[0068] Statistical results show that the average number of convergences in the closed-loop iteration is 2.3 (maximum 5). Starting from the degenerate configuration, the success rate of finding the retained configuration through bit-width adjustment is: 100% for W3 encoding, 12% for W1 encoding, and 8% for W2 encoding. The average hardware solution time cost per iteration is the same as that of a single solution. The additional time cost of closed-loop iteration is only 2-3 hardware repetitions, but it can improve the deployment success rate from 4 / 300 = 1.3% (blind selection) to 100% under W3 encoding (closed-loop search).

Claims

1. A finite-precision quantum-optimized hardware closed-loop mapping-preserving evaluation method for intelligent control, characterized in that: Includes the following steps: S1. Obtain the QUBO coefficient matrix to be deployed and the coefficient bit width, matrix size and sampling number of the target finite precision combination optimization hardware from the optimization decision module of the intelligent control system. S2. For the QUBO coefficient matrix, the initial mapping parameter vector is determined jointly based on three dimensions: coefficient dynamic range constraint, constraint-target level constraint, and hardware precision compatibility constraint. The hardware precision compatibility constraint determines the feasible range of the scaling factor based on the coefficient bit width and matrix size. The initial mapping parameter vector includes the scaling factor, effective bit width selection, and dynamic range pruning threshold. The mapping parameters are set to the initial preset parameter vector; S3. Determine the QUBO coefficient matrix according to the mapping parameters, and perform scaling, rounding and limiting operations on the QUBO coefficient matrix to generate an integer coefficient matrix; S4. Submit the integer coefficient matrix to the finite precision combinatorial optimization hardware through the hardware platform's native matrix upload interface, obtain the corresponding number of candidate solution samples and corresponding hardware energy values ​​according to the number of samplings, and maintain the original n-dimensional QUBO variable structure. S5. Calculate the hardware energy value of the original floating-point QUBO model and the difference between the corresponding hardware energy values ​​of the candidate solutions to determine the pairwise order retention rate. Compare the pairwise order retention rate with the consistency threshold. When the pairwise order retention rate is less than the consistency threshold, proceed to step S6. Otherwise, proceed to step S7; S6. Perform closed-loop feedback adjustment, determine the diagnosis result based on the pairwise order relationship and adjust the mapping parameters based on the diagnosis result, and return to step S3 until the pairwise order relationship retention rate is greater than or equal to the consistency threshold. S7. Output an integer coefficient matrix for use by the intelligent control system's optimization decision-making module.

2. The finite-precision quantum-optimized hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 1, characterized in that: In step S1, the QUBO coefficient matrix Q is an n×n real symmetric matrix, and the coefficient bit width B is an integer range [-2]. (B-1) , 2 (B-1) -1].

3. The finite-precision quantum optimization hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 1, characterized in that: Step S2 includes steps S2.1-S2.3; S2.1 Coefficient Dynamic Range Constraints: Calculate the maximum absolute value max|Q_ij|, the minimum non-zero absolute value min|Q_ij|≠0, and the dynamic range DR of the QUBO coefficient matrix as shown in formula (1). DR = 2log2(max|Q_ij| / min|Q_ij|)(1), min|Q_ij|≠0, where Q_ij is the value of the QUBO coefficient matrix at position (i,j), which determines the proportion of coefficients that can be retained with a coefficient bit width of B. S2.2 Constraints – Target Level Constraints: Identify the constraint penalty component Q_constraint and the target component Q_objective in the QUBO coefficient matrix, ensure that the energy level of the constraint penalty component is higher than the energy level of the target component after mapping, and determine the mapped coefficients that meet the conditions; S2.3 Hardware precision compatibility constraint: Based on the hardware bit width B and matrix size n, determine the feasible interval [s_min, s_max] of the scaling factor s, where s_min and s_max are the minimum and maximum values ​​of the scaling factor s, respectively, so that the coefficients after mapping neither overflow nor all return to zero.

4. The finite-precision quantum optimization hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 1, characterized in that: In step S3, the integer coefficient matrix Q'_ij is generated as shown in formula (2). Q'_ij = clip(round(s Q_ij), -2 (B-1) , 2 (B-1) -1)(2), round(·) is the rounding operation for integer taking, clip(·, lo, hi) is the clipping operation on the input, where lo is the lower limit threshold and hi is the upper limit threshold.

5. The finite-precision quantum-optimized hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 4, characterized in that: Step S3 further comprises: performing a basic integrity check on the integer matrix; if a condition that does not satisfy the basic integrity check occurs, re-acquiring the integer coefficient matrix via step S2, wherein the basic integrity check comprises: the number of non-zero coefficients, whether the constraint penalty term coefficients maintain a hierarchical relationship, and whether all-zero rows or columns exist.

6. The finite-precision quantum optimization hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 1, characterized in that: Step S4 comprises: setting the number of sampling times as N_sample, submitting the integer coefficient matrix to the finite-precision combinatorial optimization hardware through the native matrix upload interface of the hardware platform to obtain N_sample candidate solution samples {x(1), x(2),.. x(N_sample)} and corresponding hardware energy values {E_hw(1), E_hw(2), ..E_hw(N_sample)}, without performing variable compression or subgraph embedding, and maintaining the original n-dimensional QUBO variable structure.

7. The finite-precision quantum optimization hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 1, characterized in that: In step S5, the pairwise order relation retention rate R is shown as Formula (3), R = Number of sample pairs retaining the order relation / Total number of sample pairs (3), The number of sample pairs retaining the order relation is defined as: taking any two different samples (x(a), x(b)) from the N_sample candidate solutions, calculating their energy difference ΔE_float on the original floating-point QUBO model as shown in Formula (3.1), ΔE_float= E_float(x(a)) - E_float(x(b)) (3.1), E_float represents the energy value of the original floating-point coefficient matrix; Calculating the energy difference ΔE_hw on the hardware integer coefficient matrix as shown in Formula (3.2), ΔE_hw=E_hw(x(a))-E_hw(x(b)) (3.2), E_hw represents the energy value of the integer coefficient matrix; If sign(ΔE_float) = sign(ΔE_hw), the sample pair is regarded as retaining the order relation, wherein sign() is used to determine the direction of energy change, such as increase, decrease or unchanged; The total number of sample pairs is shown as Formula (3.3), Total number of sample pairs= C(N_sample, 2) = N_sample × (N_sample - 1) / 2 (3.3), where C(N_sample, 2) is the number of combinations of selecting 2 elements from N_sample different elements, R ∈ [0, 1], if R = 1, it indicates that the hardware output perfectly retains the energy sorting structure of the original model; if R<1, it indicates that sorting distortion exists.

8. The finite-precision quantum-optimized hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 7, characterized in that: The step of "comparing the pairwise order relation retention rate with a consistency threshold" in step S5 comprises: setting a consistency threshold R_th, when R ≥ R_th, determining that the hardware output consistency under the current mapping parameters meets the requirement, and proceeding to result output; when R<R_th, determining that the consistency does not meet the requirement, and proceeding to step S6.

9. The finite-precision quantum-optimized hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 7, characterized in that: Step S6 comprises: steps S6.1 to S6.4; S6.1 Distortion Mode Diagnosis: Analyze the features of sample pairs that violate the order-preserving relation and identify the source of distortion: If the features of sample pairs that violate the order-preserving relation are concentrated in sample pairs with small energy differences |ΔE_float|, the diagnosis result is determined to be "insufficient precision distortion", and the adjustment direction is to increase the effective precision of the hardware; if the features of sample pairs that violate the order-preserving relation are scattered in various energy difference intervals, the diagnosis result is determined to be "structural coupling distortion", and the adjustment direction is to change the scaling strategy or dynamic range clipping point and proceed to step S6.2; S6.2 Mapping parameter adjustment: Based on the diagnostic results, the adjusted mapping parameter θ_{k+1} is generated, specifically including: the scaling factor s_{k+1} is adjusted as shown in formula (4); s_{k+1}= s_k × (1 ± δ_s) (4), δ_s is the preset step size, s_k is the scaling factor at k times, and s_{k+1} is the scaling factor at k+1 times. The effective bit width B_{k+1} is adjusted as shown in formula (4.1). B_{k+1} = B_k ± 1 (4.1); The effective bit width is within the range supported by the hardware. The dynamic range clipping adjustment R_dynamic_{k+1} expands or narrows according to the preset adjustment value based on the distortion situation. S6.3 Iterative Execution: Return to step S3 with θ_{k+1} and re-execute the precision mapping until the termination condition is met, at which point the iteration terminates. The termination condition includes any of the following conditions: R ≥ R_th; the number of iterations k ≥ k_max, where k_max is the preset maximum number of iterations; the change in R between two consecutive iterations |ΔR| < ε_R, where ε_R is a preset change threshold. S6.4 Iteration Record: When the iteration terminates, the mapping parameter θ_k, consistency index R_k, and distortion mode diagnosis results of each iteration are recorded to form a mapping process archive for reference by the decision reliability assessment module of the intelligent control system.

10. The finite-precision quantum-optimized hardware closed-loop mapping-preserving evaluation method for intelligent control according to claim 1, characterized in that: Step S7 includes: outputting the optimal mapping parameter θ and the corresponding integer coefficient matrix Q'_ij; the hardware candidate solution sample set and the corresponding original floating-point energy intermediate evaluation value; the final consistency indicator R and the mapping process archive; R ≥ R_th is marked as "trusted", R <R_th is marked as "for reference only", and the output result is provided to the optimization decision module of the intelligent control system as a basis for generating control instructions.

Images