Layered multi-objective optimization method and system for approximate calculation parameter optimization

By constructing a hierarchical multi-objective optimization method and combining multi-armed slot machine and multi-objective Bayesian optimization, the problems of low efficiency and high cost in hierarchical and mixed variable search spaces in approximate calculations are solved, and efficient and low-cost approximate calculation configuration optimization is achieved.

CN121480618APending Publication Date: 2026-02-06NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610020952.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-08
Publication Date
2026-02-06

AI Technical Summary

Technical Problem

Existing tuning techniques are ineffective in handling hierarchical and mixed variable search spaces in approximation computation, resulting in low efficiency and high costs. Furthermore, the existing MAB+BO architecture fails to solve the hierarchical decision-making problem of 'selecting a strategy first and then tuning parameters' in approximation computation.

Method used

A hierarchical multi-objective optimization method is constructed. The multi-armed slot machine algorithm is used to select approximate strategies in the categorical decision layer, and multi-objective Bayesian optimization is combined to optimize parameters in the numerical decision layer. Evaluation points are efficiently generated through Beta distribution and qNEHVI acquisition function, and sample utilization efficiency is improved by Latin hypercube sampling and iterative update mechanism.

Benefits of technology

It significantly shortens the tuning cycle, reduces costs, and provides an approximate computational configuration that balances performance and quality. By accurately matching the hierarchical parameter space features, it reduces the number of invalid evaluations and improves sample utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121480618A_ABST
    Figure CN121480618A_ABST
Patent Text Reader

Abstract

The invention relates to a hierarchical multi-objective optimization method and system for approximate calculation parameter optimization. The method comprises the following steps: constructing a hierarchical search model for approximate calculation commitment optimization; and in the category-type decision-making layer, taking each approximation strategy as an arm, maintaining the arms by adopting a multi-arm tiger machine algorithm, obtaining random sampling values in each round of iteration according to the obtained Beta distribution, and evaluating the target arm corresponding to the maximum value so as to obtain an arm to be optimized. And inputting the parameter point to be evaluated into a numerical decision-making layer to perform parameter iterative optimization, and generating a next parameter point to be evaluated by maximizing a qNEHVI acquisition function, so that a program, corresponding to the parameter point, of the arm to be optimized is executed through a multi-target Bayesian optimization engine, and an optimization evaluation result is obtained. And obtaining the most approximate calculation configuration according to the result and the preset evaluation budget. The method can face mixed and hierarchically dependent variable space, and can improve the tuning efficiency and reduce the cost.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of parallel application approximate computing, in particular to a hierarchical multi-objective optimization method and system for approximate computing parameter optimization. BACKGROUND

[0002] With the end of Amdahl's law and the slowdown of Moore's law, approximate computing has become an important paradigm to improve the performance of high-performance computing applications. This technology introduces controlled errors within the tolerance range of the application program, in exchange for significant performance improvement. In order to find the best approximate configuration, autotuning technology is widely used.

[0003] Currently, the mainstream autotuning technology, especially Bayesian Optimization (BO), performs well in handling "black box" function optimization problems with expensive evaluation costs. It learns the shape of the target function by building a proxy model (such as Gaussian process), and uses a collection function to intelligently select the next evaluation point, so as to find the optimal solution with fewer evaluation times. Some advanced tuning frameworks also begin to explore the combination of Bayesian optimization and multi-armed bandit (MAB) technology to handle more complex problems.

[0004] Although existing tuning technologies have made some progress, when applied to approximate computing, there are still the following key defects: 1. Difficulty in handling hierarchical and mixed variable complex search space: The parameter space of approximate computing has a unique hierarchical structure. The high level is a categorical decision, such as choosing which approximate strategy to use (for example, loop perforation or precision scaling); the bottom layer is a numerical decision, such as setting the specific parameters of the strategy (for example, perforation step or data type bit width). Traditional Bayesian optimization methods are inefficient in directly handling such mixed and hierarchical dependent variable spaces, making it difficult to effectively navigate.

[0005] 2. Low sample efficiency leads to high tuning cost: In high-performance computing, the cost of evaluating a single approximate configuration is extremely high, requiring complete program compilation and running. Existing tuning methods fail to fully utilize the internal structure of the approximate computing parameter space, resulting in slow convergence and requiring a large number of evaluation times to find a better solution, which is difficult to bear in practical applications.

[0006] 3. Existing MAB+BO architecture is not applicable: Some existing work combines MAB and BO, but the MAB is usually used for budget allocation of multi-fidelity evaluation, i.e., deciding how much computing resources to allocate to more promising configurations. This use does not solve the hierarchical decision problem of "first choosing strategy, then tuning parameters" in approximate computing. Therefore, there is a lack of an optimization architecture specifically tailored for the mixed variable search space of approximate computing. SUMMARY

[0007] Based on this, it is necessary to provide a hierarchical multi-objective optimization method and system for approximate computing parameter optimization, which can face mixed and hierarchical dependent variable space, improve optimization efficiency and reduce cost.

[0008] A hierarchical multi-objective optimization method for approximate computing parameter optimization, the method comprising: Constructing a hierarchical search model for approximate computing commitment optimization, the hierarchical search model comprising a categorical decision layer and a numerical decision layer.

[0009] In the categorical decision layer, each approximate strategy is taken as an arm, a multi-arm bandit algorithm is used to maintain the arm, a Beta distribution is obtained, a random sampling value is randomly obtained in each iteration according to the Beta distribution, the target arm corresponding to the maximum value of the random sampling value is evaluated to obtain the to-be-optimized arm.

[0010] The to-be-optimized arm is input to the numerical decision layer for parameter iterative optimization, the next parameter point to be evaluated is generated by maximizing the qNEHVI acquisition function, and the program corresponding to the to-be-optimized arm and the next parameter point to be evaluated is executed by the multi-objective Bayesian optimization engine to obtain the optimization evaluation result.

[0011] According to the optimization evaluation result and the preset evaluation budget, the best approximate computing configuration is obtained.

[0012] A hierarchical multi-objective optimization system for approximate computing parameter optimization, the system comprising: A model construction module for constructing a hierarchical search model for approximate computing commitment optimization, the hierarchical search model comprising a categorical decision layer and a numerical decision layer.

[0013] A to-be-optimized arm acquisition module for, in the categorical decision layer, taking each approximate strategy as an arm, using a multi-arm bandit algorithm to maintain the arm, obtaining a Beta distribution, randomly obtaining a random sampling value in each iteration according to the Beta distribution, evaluating the target arm corresponding to the maximum value of the random sampling value to obtain the to-be-optimized arm.

[0014] A parameter configuration optimization module for inputting the to-be-optimized arm to the numerical decision layer for parameter iterative optimization, generating the next parameter point to be evaluated by maximizing the qNEHVI acquisition function, and executing the program corresponding to the to-be-optimized arm and the next parameter point to be evaluated by the multi-objective Bayesian optimization engine to obtain the optimization evaluation result.

[0015] An approximate computing configuration module for obtaining the best approximate computing configuration according to the optimization evaluation result and the preset evaluation budget.

[0016] The aforementioned hierarchical multi-objective optimization method and system for approximate computation parameter optimization constructs a hierarchical search model with categorical and numerical decision layers, accurately matching the hierarchical parameter space characteristics of approximate computation where strategies are selected first and parameters are then adjusted. In the categorical decision layer, various approximate strategies are treated as arms of a multi-armed machine. The strategy evaluation state is maintained using a Beta distribution, and Thompson sampling is used to quickly select the arm with the best potential for optimization, avoiding the inefficiency waste caused by blindly exploring all strategies in traditional methods. Entering the numerical decision layer, for the numerical parameter space of the selected strategy, a double Gaussian process surrogate model is used to learn the mapping relationship between execution time and QoS loss, and the qNEHVI acquisition function is combined to efficiently generate the next parameter point to be evaluated, solving the dependency and adaptation problem between numerical parameters and categorical strategies. Simultaneously, the sample utilization efficiency is significantly improved through Latin hypercube sampling initialization and iterative update mechanisms. Each iteration only needs to focus on the parameter optimization of the optimal strategy, reducing the number of invalid evaluations and lowering the high cost of program compilation and execution. Ultimately, the optimal Pareto front configuration is selected based on a preset evaluation budget. While ensuring controllable QoS loss, the tuning cycle is significantly shortened. This not only solves the pain point of inefficient navigation in mixed hierarchical spaces by traditional methods, but also reduces tuning costs through efficient use of samples, providing users with a better approximate computation configuration that balances performance and quality. Attached Figure Description

[0017] Figure 1 This is a system architecture diagram of a hierarchical multi-objective optimization method for approximate computational parameter optimization in one embodiment; Figure 2 This is a flowchart of a hierarchical multi-objective optimization method for approximate computational parameter optimization in one embodiment; Figure 3 This is a flowchart illustrating a hierarchical multi-objective optimization process for approximate computational parameter optimization in one embodiment. Figure 4 This is a block diagram of a hierarchical multi-objective optimization system for approximate computational parameter optimization in one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] This invention provides a hierarchical multi-objective optimization method for approximate computational parameter optimization, which can be applied to, for example... Figure 1The system architecture shown consists of a high-level layer and a low-level layer. The high-level layer is for categorical strategy exploration (using Multi-Armed Slot Machine (MAB)), and the low-level layer is for numerical parameter optimization (using Multi-Objective Bayesian Optimization (MOBO)).

[0020] In one embodiment, such as Figure 2 As shown, a hierarchical multi-objective optimization method for approximate computational parameter optimization is provided, which is then applied to... Figure 1 Taking the system architecture in [the document] as an example, the following steps are included: Step 202: Construct a hierarchical search model for approximate computation commitment optimization. The hierarchical search model includes a categorical decision layer and a numerical decision layer.

[0021] Step 204: In the categorical decision layer, each approximation strategy is treated as an arm. The multi-armed slot machine algorithm is used to maintain the arms and obtain the Beta distribution. Based on the Beta distribution, random sample values ​​are randomly obtained in each iteration. The target arm corresponding to the maximum value of the random sample value is evaluated to obtain the arm to be optimized.

[0022] Step 206: Input the arm to be optimized into the numerical decision layer for parameter iterative optimization. By maximizing the qNEHVI acquisition function, generate the next parameter point to be evaluated, so that the program corresponding to the arm to be optimized and the next parameter point to be evaluated is executed through the multi-objective Bayesian optimization engine to obtain the optimization evaluation result.

[0023] Step 208: Based on the optimization evaluation results and the preset evaluation budget, obtain the most approximate calculation configuration.

[0024] In the aforementioned hierarchical multi-objective optimization method for approximate computation parameter optimization, a hierarchical search model is constructed, consisting of a categorical decision layer and a numerical decision layer. This model accurately matches the hierarchical parameter space characteristics of approximate computation, where a strategy is selected first, and then parameters are adjusted. In the categorical decision layer, various approximate strategies are treated as arms of a multi-armed machine. The strategy evaluation state is maintained using a Beta distribution, and Thompson sampling is used to quickly select the arm with the best potential for optimization, avoiding the inefficiency waste caused by blindly exploring all strategies in traditional methods. Upon entering the numerical decision layer, for the numerical parameter space of the selected strategy, a double Gaussian process surrogate model is used to learn the mapping relationship between execution time and QoS loss. Combined with the qNEHVI sampling function, the next parameter point to be evaluated is efficiently generated, solving the dependency and adaptation problem between numerical parameters and categorical strategies. Simultaneously, the sample utilization efficiency is significantly improved through Latin hypercube sampling initialization and iterative update mechanisms. Each iteration only needs to focus on the parameter optimization of the optimal strategy, reducing the number of invalid evaluations and lowering the high cost of program compilation and execution. Ultimately, the optimal Pareto front configuration is selected based on a preset evaluation budget. While ensuring controllable QoS loss, the tuning cycle is significantly shortened. This not only solves the pain point of inefficient navigation in mixed hierarchical spaces by traditional methods, but also reduces tuning costs through efficient use of samples, providing users with a better approximate computation configuration that balances performance and quality.

[0025] In one embodiment, in the categorical decision layer, each approximate strategy is treated as an arm. A multi-armed slot machine algorithm is used to maintain the arm, initialize the Beta distribution parameters for the arm, and record the historical evaluation results corresponding to the arm. If the historical evaluation results simultaneously meet a preset performance threshold and a QoS loss threshold, the arm is marked as successfully evaluated; otherwise, it is marked as failed. The corresponding Beta distribution parameters are updated based on the number of times the evaluation is marked as successful to obtain the Beta distribution.

[0026] In one embodiment, random sample values ​​are generated based on the current Beta distribution of the arm, the random sample values ​​are sorted, and the maximum value of the random sample values ​​is taken as a candidate target arm. It is verified whether the historical evaluation count of the candidate target arm exceeds the maximum initial exploration threshold for a single arm. If it does not exceed the threshold, the candidate target arm in the current iteration is the arm to be optimized. If it exceeds the threshold, a second sampling and re-selection are performed until the evaluation of all the arms is completed, and all arms to be optimized are obtained.

[0027] In one embodiment, the arm to be optimized is input into the numerical decision layer. Two Gaussian process surrogate models are used for iterative training on the execution time and QoS loss objectives of the arm to be optimized, respectively, to obtain prediction results. The boundary constraints and parameter types of the numerical parameter space are determined based on the prediction results. A legitimate parameter search domain is constructed based on the determined boundary constraints and parameter types. The qNEHVI values ​​of each candidate parameter point within the legitimate parameter search domain are calculated based on the prediction results of the Gaussian process surrogate model in the current iteration and the global Pareto front. By maximizing the qNEHVI acquisition function, the parameter point that maximizes the qNEHVI value is output as the next parameter point to be evaluated. .

[0028] In one embodiment, the parameter type includes continuous and discrete types. For the next parameter point to be evaluated... Perform a validity check; if the parameter point... If a point is a non-integer point or an invalid value in the discrete parameter space, then the nearest neighbor interpolation method or the constrained projection method is used to correct it into a valid parameter point, so that the parameter point... All dimensional parameters satisfy the parameter constraints of the arm to be optimized; Check the parameter points If the Euclidean distance between the Euclidean distance and the historical evaluation parameter points of the arm to be optimized is less than a preset minimum distance threshold, then the qNEHVI acquisition function is re-optimized to generate new parameter points.

[0029] In one embodiment, the approximate policy corresponding to the arm to be optimized is compared with the next parameter point to be evaluated using a multi-objective Bayesian optimization engine. The corresponding application source code is embedded to generate an executable file. The executable file is run under preset hardware environment and load conditions. Performance monitoring tools are used to collect time data for the entire program execution cycle, which is used as the execution time evaluation result. A dedicated detection tool is used to compare the deviation between the approximate calculation output and the precise calculation result, and the QoS loss value is calculated according to a preset QoS loss quantification formula. The execution time evaluation result and the QoS loss value are integrated to obtain an optimized evaluation result that includes parameter configuration and dual-objective evaluation data.

[0030] In one embodiment, the preset evaluation budget is the maximum number of evaluations or the longest evaluation duration, and the total number of evaluations and cumulative time of the current iteration are counted in real time. If the evaluation budget is not reached, the iteration optimization phase continues to execute the strategy selection of the categorical decision layer and the parameter optimization process of the numerical decision layer. If the evaluation budget has been reached, the iteration terminates and enters the optimal configuration screening, summarizing all valid evaluation data, using a non-dominated sorting algorithm to screen out the global Pareto front set, calculating the hypervolume contribution value of each configuration in the global Pareto front set, and selecting the configuration corresponding to the maximum hypervolume contribution value as the optimal approximate calculation configuration; if there are multiple configurations with the same hypervolume contribution value, the configuration corresponding to the minimum QoS loss is preferentially selected as the optimal approximate calculation configuration.

[0031] In one embodiment, such as Figure 3 As shown, a flowchart of a hierarchical multi-objective optimization process for approximate computational parameter optimization is provided. The specific steps are as follows: Step 1: Problem Modeling and Hierarchical Decomposition The parameter tuning problem of approximate calculation is modeled as a hierarchical search space.

[0032] High-level (categorical decision-making level): Defines all possible combinations of approximation strategies. Each unique combination of approximation strategies (e.g., using loop perforation for function A and precision scaling for function B) is considered an independent "arm," constituting a multi-armed slot machine (MAB) problem.

[0033] The bottom layer (numerical decision layer): For each "arm" (i.e., a strategy combination) selected by the top layer, its corresponding specific parameters (such as loop step size, data bit width) constitute a continuous or discrete numerical parameter space. The optimization objective of this layer is to find a set of parameters that simultaneously minimizes the program's execution time and QoS penalty, constituting a multi-objective optimization problem.

[0034] Step 2: Exploring High-Level Category-Based Strategies (Based on MAB) This invention employs the Thompson Sampling algorithm to efficiently explore high-level approximation strategies.

[0035] 1. Model Establishment: For each "arm" k (representing an approximate policy) in the MAB, maintain a Beta distribution, denoted as Beta(alpha_k, beta_k). This distribution is used to model the probability of the policy succeeding. The parameters alpha_k and beta_k record the number of successful and failed evaluations of the policy in history, respectively. A single evaluation is defined as a "success" if and only if the program can complete normally and its QoS loss meets the user-preset budget.

[0036] 2. Strategy Selection: In each iteration, the system randomly samples a value theta_k from the current Beta(alpha_k, beta_k) distribution of each "arm" k. Then, it selects the "arm" with the largest sampled value theta_k. This serves as an approximate strategy to be evaluated in this round. This mechanism can intelligently balance the "utilization" of known excellent strategies and the "exploration" of unknown strategies.

[0037] Step 3: Optimization of underlying numerical parameters (based on MOBO) When the senior MAB selects a strategy "arm" Then, the system starts the multi-objective Bayesian optimization (MOBO) engine to find the optimal numerical parameters.

[0038] 1. Proxy model construction: for the current strategy Two independent Gaussian process (GP) models are maintained, serving as proxy models for program execution time and QoS loss, respectively. Based on existing evaluation data, these two GP models learn the mapping relationship from the numerical parameter space of the policy to the two optimization objectives (performance and quality), and can provide predicted values ​​and uncertainty measures for unknown parameter points.

[0039] 2. Acquisition Function: To guide the search process, this invention employs q-NoisyExpected Hypervolume Improvement (qNEHVI) as the acquisition function. This function is used to find one or a set of new parameter points in the parameter space that are expected to maximize the hypervolume increment of the current Pareto front. qNEHVI is particularly suitable for handling noisy objective function values ​​derived from measurements of real systems.

[0040] 3. Parameter point generation: By optimizing the acquisition function, find the next parameter point x_next that maximizes the qNEHVI value, and use it as the specific parameter configuration to be evaluated in this round.

[0041] Step 4: Integration and Optimization Algorithm Flow The process of integrating MAB and MOBO into a unified iterative optimization framework is as follows: 1. Initialization Phase: To provide initial data for MAB and MOBO, the system uses Latin Hypercube Sampling to generate a small number of initial parameter configurations for each policy "arm" for evaluation. The evaluation results are used to initialize the Beta distribution parameters (alpha_k, beta_k) for each arm and the training dataset for the GP model.

[0042] 2. Iterative Optimization Phase: Within the given total evaluation budget, repeat the following loop: a. MAB Strategy Selection: Select an optimal strategy "arm" based on the Thompson sampling method described in step 2. b. MOBO Proposal Parameters: Based on the GP model associated with k_, and following the method described in step three, the next parameter point x_next to be evaluated is generated by maximizing the qNEHVI acquisition function. c. Evaluation and Update: Compile and execute the program applying policy k_ and parameter x_next, and measure its execution time and QoS loss. d. Update the corresponding "arm" in MAB based on whether the evaluation is "successful". The Beta(alpha_k, beta_k) distribution. e. If the evaluation is successful, add the new (parameters, target value) data points to the training set of the corresponding GP model in MOBO, and update the global Pareto front.

[0043] 3. Output: Once the evaluation budget is reached, the algorithm terminates and returns the global Pareto front P discovered so far. This set contains a series of approximate computational configurations that achieve the optimal trade-off between performance and quality for the user to choose from.

[0044] It's worth noting that through the collaborative work of MAB and MOBO, the system can intelligently explore and utilize a vast search space, prioritizing evaluation resources for the most promising strategies and parameter regions. This allows for rapid convergence to a high-quality solution set with far fewer evaluations than existing technologies, significantly reducing the time cost of automated tuning. The system aims to find Pareto optimal solutions, providing users with a series of optimal configurations that strike different trade-offs between performance and quality, rather than a single optimal point. This empowers users to flexibly choose the best solution based on their specific needs.

[0045] It should be understood that, although Figures 2-3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figures 2-3At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0046] In one embodiment, such as Figure 4 As shown, a hierarchical multi-objective optimization system for approximate calculation parameter optimization is provided, including: a model building module 402, a target arm acquisition module 404, a parameter configuration optimization module 406, and an approximate calculation configuration module 408, wherein: The model building module 402 is used to build a hierarchical search model for approximate computation commitment tuning. The hierarchical search model includes a categorical decision layer and a numerical decision layer.

[0047] The arm acquisition module 404 is used in the categorical decision layer to treat each approximate strategy as an arm, maintain the arms using a multi-armed slot machine algorithm to obtain a Beta distribution, randomly acquire random sample values ​​in each iteration based on the Beta distribution, and evaluate the target arm corresponding to the maximum value of the random sample value to obtain the arm to be optimized.

[0048] The parameter configuration optimization module 406 is used to input the arm to be optimized into the numerical decision layer for parameter iterative optimization. By maximizing the qNEHVI acquisition function, the next parameter point to be evaluated is generated, so that the program corresponding to the arm to be optimized and the next parameter point to be evaluated is executed by the multi-objective Bayesian optimization engine to obtain the optimization evaluation result.

[0049] The approximate calculation configuration module 408 is used to obtain the most approximate calculation configuration based on the optimization evaluation results and the preset evaluation budget.

[0050] Specific limitations regarding a hierarchical multi-objective optimization system for approximate computational parameter optimization can be found in the limitations of a hierarchical multi-objective optimization method for approximate computational parameter optimization described above, and will not be repeated here. Each module in the aforementioned hierarchical multi-objective optimization system for approximate computational parameter optimization can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0051] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0052] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A hierarchical multi-objective optimization method for approximate computational parameter optimization, characterized in that, The method includes: A hierarchical search model for approximate computational commitment optimization is constructed, the hierarchical search model including a categorical decision layer and a numerical decision layer; In the categorical decision layer, each approximation strategy is treated as an arm, and the multi-armed slot machine algorithm is used to maintain the arm to obtain a Beta distribution. Based on the Beta distribution, random sample values ​​are randomly obtained in each iteration, and the target arm corresponding to the maximum value of the random sample value is evaluated to obtain the arm to be optimized. The arm to be optimized is input into the numerical decision layer for parameter iterative optimization. By maximizing the qNEHVI acquisition function, the next parameter point to be evaluated is generated, so that the program corresponding to the arm to be optimized and the next parameter point to be evaluated is executed by the multi-objective Bayesian optimization engine to obtain the optimization evaluation result. Based on the optimized evaluation results and the preset evaluation budget, the most approximate calculation configuration is obtained.

2. The method according to claim 1, characterized in that, In the categorical decision layer, each approximation strategy is treated as an arm, and a multi-armed slot machine algorithm is used to maintain the arms, resulting in a Beta distribution, including: In the categorical decision layer, each approximate strategy is treated as an arm. A multi-armed slot machine algorithm is used to maintain the arm, initialize the Beta distribution parameters for the arm, and record the historical evaluation results corresponding to the arm. If the historical evaluation results simultaneously meet the preset performance threshold and QoS loss threshold, the evaluation of the arm is marked as successful; otherwise, it is marked as failed. The Beta distribution is obtained by updating the corresponding Beta distribution parameters based on the number of times the evaluation is marked as successful.

3. The method according to claim 1, characterized in that, In each iteration, random sample values ​​are obtained based on the Beta distribution. The target arm corresponding to the maximum value of the random sample values ​​is evaluated to obtain the arm to be optimized, including: Random sample values ​​are generated based on the current Beta distribution of the arm. The random sample values ​​are sorted, and the maximum value of the random sample values ​​is taken as the candidate target arm. It is verified whether the number of historical evaluations of the candidate target arm exceeds the maximum initial exploration threshold of a single arm. If it does not exceed the threshold, the candidate target arm in the current iteration is the arm to be optimized. If it exceeds the threshold, a second sampling and re-selection are performed until the evaluation of all the arms is completed, and all arms to be optimized are obtained.

4. The method according to claim 3, characterized in that, The arm to be optimized is input into the numerical decision layer for iterative parameter optimization. By maximizing the qNEHVI acquisition function, the next parameter point to be evaluated is generated, including: The arm to be optimized is input into the numerical decision layer, and two Gaussian process proxy models are used to iteratively train the execution time and QoS loss target of the arm to be optimized, respectively, to obtain the prediction results. Based on the prediction results, the boundary constraints and parameter types of the numerical parameter space are determined. Construct a valid parameter search domain based on the determined boundary constraints and parameter types; Based on the prediction results of the Gaussian process surrogate model in the current iteration round and the global Pareto front, calculate the qNEHVI value of each candidate parameter point in the legal parameter search domain; By maximizing the qNEHVI acquisition function, the parameter point that maximizes the qNEHVI value is output as the next parameter point to be evaluated. .

5. The method according to claim 4, characterized in that, The parameter types include continuous and discrete types; After inputting the arm to be optimized into the numerical decision layer for parameter iterative optimization, and generating the next parameter point to be evaluated by maximizing the qNEHVI acquisition function, the method further includes: For the next parameter point to be evaluated Perform a validity check; if the parameter point... If a point is a non-integer point or an invalid value in the discrete parameter space, then the nearest neighbor interpolation method or the constrained projection method is used to correct it into a valid parameter point, so that the parameter point... All dimensional parameters satisfy the parameter constraints of the arm to be optimized; Check the parameter points If the Euclidean distance between the Euclidean distance and the historical evaluation parameter points of the arm to be optimized is less than a preset minimum distance threshold, then the qNEHVI acquisition function is re-optimized to generate new parameter points.

6. The method according to claim 5, characterized in that, To enable the execution of a program corresponding to the arm to be optimized and the next parameter point to be evaluated through a multi-objective Bayesian optimization engine, and to obtain optimization evaluation results, including: This allows the approximate policy corresponding to the arm to be optimized to be compared with the next parameter point to be evaluated using a multi-objective Bayesian optimization engine. Embed the corresponding application source code to generate an executable file; The executable file is run under preset hardware environment and load conditions. Time data of the complete execution cycle of the program is collected by performance monitoring tools as the execution time evaluation result. And by comparing the deviation between the approximate calculation output and the precise calculation result using a dedicated detection tool, the QoS loss value is calculated according to the preset QoS loss quantification formula; By integrating the execution time evaluation results with the QoS loss values, an optimized evaluation result containing parameter configuration and dual-objective evaluation data is obtained.

7. The method according to claim 6, characterized in that, Based on the optimized evaluation results and the preset evaluation budget, the optimal approximate calculation configuration is obtained, including: The preset evaluation budget is the maximum number of evaluations or the longest evaluation time, and the total number of evaluations and cumulative time in the current iteration are counted in real time. If the assessment budget is not reached, return to the iterative optimization phase and continue executing the strategy selection process of the categorical decision layer and the parameter optimization process of the numerical decision layer. If the evaluation budget has been reached, the iteration terminates and enters the optimal configuration screening. All valid evaluation data are summarized, and a global Pareto front set is selected using a non-dominated sorting algorithm. The hypervolume contribution value of each configuration in the global Pareto front set is calculated, and the configuration corresponding to the maximum hypervolume contribution value is selected as the optimal approximate calculation configuration. If there are multiple configurations with the same hypervolume contribution value, the configuration corresponding to the minimum QoS loss is selected as the optimal approximate calculation configuration.

8. A hierarchical multi-objective optimization system for approximate computational parameter optimization, characterized in that, The system includes: The model building module is used to build a hierarchical search model for approximate computation commitment optimization. The hierarchical search model includes a categorical decision layer and a numerical decision layer. The module for obtaining the arm to be optimized is used in the categorical decision layer to treat each approximate strategy as an arm, maintain the arm using a multi-armed slot machine algorithm to obtain a Beta distribution, randomly obtain random sample values ​​in each iteration based on the Beta distribution, and evaluate the target arm corresponding to the maximum value of the random sample values ​​to obtain the arm to be optimized. The parameter configuration optimization module is used to input the arm to be optimized into the numerical decision layer for parameter iterative optimization. By maximizing the qNEHVI acquisition function, the next parameter point to be evaluated is generated, so that the program corresponding to the arm to be optimized and the next parameter point to be evaluated is executed by the multi-objective Bayesian optimization engine to obtain the optimization evaluation result. The approximate calculation configuration module is used to obtain the most approximate calculation configuration based on the optimization evaluation results and the preset evaluation budget.