An OS Energy Consumption Adaptive Scoring Method and System Based on Multi-Fidelity Bayesian Optimization
By employing a multi-fidelity Bayesian optimization method, the problems of device differences and scene interference in OS energy consumption evaluation are solved, achieving standardization of energy consumption data and accuracy of comprehensive scoring, and adapting to the energy consumption evaluation needs of various types of terminals.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INFORMATION SCI & TECH UNIV
- Filing Date
- 2026-02-02
- Publication Date
- 2026-07-31
AI Technical Summary
Existing OS energy consumption assessment methods lack standardized procedures, cannot effectively eliminate interference from individual device differences and scenario load changes, resulting in distorted energy consumption data. Furthermore, the determination of weights is highly subjective, making it difficult to adapt to the assessment needs of different application scenarios. Optimization methods suffer from an imbalance between efficiency and accuracy.
A multi-fidelity Bayesian optimization method is adopted. By collecting and transforming energy consumption data, a mixed-effect model is constructed, and the optimal weights of each level are adaptively solved. A comprehensive score is performed based on the consistency of energy consumption ranking, and the energy consumption level is defined by combining the preset interval.
It achieves standardization and accuracy of energy consumption data, improves the consistency and reliability of energy consumption evaluation, adapts to the evaluation needs of different application scenarios, and reduces computing costs.
Smart Images
Figure CN122489392A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of operating system energy consumption evaluation, and in particular to an OS energy consumption adaptive scoring method and system based on multi-fidelity Bayesian optimization. Background Technology
[0002] With the widespread adoption of various computing devices such as embedded systems, mobile terminals, and servers, the operating system (OS), as the core management carrier of software and hardware resources, directly impacts device battery life, operational stability, and energy efficiency, making it one of the key indicators for measuring overall OS performance. Whether it's mobile device users' demand for long battery life or data centers' need to control operational energy costs, both place higher demands on the accuracy and comprehensiveness of OS energy consumption assessments.
[0003] In the current field of OS energy consumption evaluation, there are three main types of technologies: First, direct energy consumption measurement methods, which collect system operating energy consumption data through hardware devices such as power meters, or indirectly estimate energy consumption based on CPU performance counters; second, single index scoring methods, which focus on evaluating a single dimension such as global energy consumption average or energy consumption in a specific scenario; and third, comprehensive scoring methods, which select multiple energy consumption-related indicators, assign fixed weights using methods such as expert scoring and hierarchical analysis, and then sum the results by weighting. Some solutions introduce ordinary Bayesian optimization or linear model-assisted optimization.
[0004] However, these existing technologies still have significant shortcomings. Data preprocessing lacks standardized procedures and fails to effectively eliminate interference from non-OS factors such as individual device differences, experimental batch fluctuations, and scenario load changes, leading to distorted energy consumption data and affecting the objectivity of the evaluation. Furthermore, the indicator design is one-sided, or only focuses on a single energy consumption value, failing to decompose the multi-dimensional energy consumption characteristics of the OS and thus failing to fully reflect the true performance of the OS. At the same time, the determination of weights is highly subjective, and traditional manual weighting or fixed weight methods are difficult to adapt to the evaluation needs of different application scenarios. The optimization methods suffer from an imbalance between efficiency and accuracy, and the computational cost of single-fidelity optimization is too high / the prediction accuracy is insufficient, making it difficult to achieve the optimal weight solution under budget constraints.
[0005] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.
[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0007] The purpose of this disclosure is to provide an OS power consumption adaptive scoring method and system based on multi-fidelity Bayesian optimization, thereby overcoming at least to some extent one or more problems caused by the limitations and defects of related technologies.
[0008] Firstly, this application provides an OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization, including: Collect energy consumption data of the tested operating system and the benchmark operating system, and perform a natural logarithmic transformation after calculating the energy consumption ratio to obtain standard energy consumption data; A mixed-effect model is constructed based on the standard energy consumption data, and the parameters of the mixed-effect model are converted into hierarchical index parameters. With energy consumption ranking consistency as the optimization objective, an optimization problem is constructed based on the hierarchical index parameters, and the optimal weights of each level are adaptively solved through multi-fidelity Bayesian optimization iterative search. Based on the optimal weights of the hierarchical index parameters and the scores of each level, a comprehensive energy consumption score is obtained through weighted summation, and energy consumption levels are defined according to a preset range.
[0009] In one possible implementation, the step of collecting energy consumption data of the tested operating system and the benchmark operating system, and performing a natural logarithmic transformation after calculating the energy consumption ratio to obtain standard energy consumption data includes: Collect energy consumption data of the tested operating system and the benchmark operating system under the same device, in the same scenario, and in the same batch; Calculate the energy consumption ratio between the tested operating system's energy consumption data and the benchmark operating system's energy consumption data, and remove background noise; The energy consumption ratio is transformed by natural logarithm to improve the normality of the data.
[0010] In one possible implementation, the energy consumption ratio is calculated using the following formula: in, This is the energy consumption ratio. This refers to the power consumption data of the operating system under test. This is the power consumption data for the benchmark operating system; The formula for calculating the standard energy consumption data is: in, This is standard energy consumption data.
[0011] In one possible implementation, the step of constructing a mixed-effect model based on the standard energy consumption data and converting the parameters of the mixed-effect model into hierarchical index parameters includes: A structured data table is constructed based on the standard energy consumption data, and a mixed-effect model is built. The parameters of the mixed-effects model are estimated by means of the maximum likelihood method. The mixed-effects model parameters include global average effect, scene fixed effect, interaction term between operating system and scene, device volume difference, batch difference and residual. The parameters of the mixed-effects model are converted into hierarchical index parameters; the hierarchical index parameters are equipment variance variance, batch variance variance, and other noise during measurement.
[0012] In one possible implementation, the mixed-effects model is: in, , , , For standard energy consumption data, For global average effect, For scene fixed effect, For the interaction between the operating system and the scene, Due to equipment differences, This represents the variance due to equipment differences. Due to batch differences, For batch variation variance, For residuals, Other noise during measurement.
[0013] In one possible implementation, the step of constructing an optimization problem based on the hierarchical index parameters with energy consumption ranking consistency as the optimization objective, and adaptively solving for the optimal weights at each level through multi-fidelity Bayesian optimization iterative search, includes: Based on the hierarchical index parameters and using labels based on actual energy consumption data, an optimization problem is constructed with the goal of energy consumption ranking consistency. A multi-fidelity Bayesian optimization model is constructed by a joint Gaussian process; the multi-fidelity Bayesian optimization model performs coarse screening of the weight space through a low-fidelity approximation index, and then performs local precise optimization through a high-fidelity energy consumption ranking consistency. Based on the aforementioned multi-fidelity Bayesian optimization model, the optimal weights for each level are adaptively solved through iterative search using a cost-weighted expectation improvement function under budget constraints.
[0014] In one possible implementation, the multi-fidelity Bayesian optimization model is: in, For low-fidelity optimization model, To optimize the model for high fidelity, For scaling / offset relationships from low-fidelity to high-fidelity, This is the first independent Gaussian process in the joint Gaussian process. It is the second independent Gaussian process in the joint Gaussian process; The cost-weighted expected improvement function is: in, To improve the a posteriori expectation of the current joint Gaussian process for this fidelity, To assess costs.
[0015] In one possible implementation, the step of obtaining a comprehensive energy consumption score through weighted summation based on the optimal weights of the hierarchical index parameters and the scores of each level, and defining the energy consumption level according to a preset range, includes: The hierarchical score is calculated based on the hierarchical index parameters; wherein, the hierarchical score is a global score, a scene score, and a stability score; The comprehensive energy consumption score is calculated using the hierarchical scores and the optimal weights for each level. Based on a preset energy consumption level range, the energy consumption level is determined according to the comprehensive energy consumption score.
[0016] In one possible implementation, the expression for the global score is: in, For overall energy efficiency, This is the global average effect estimate obtained through the mixed-effects model; The expression for the scene rating is: in, , As a consistency indicator, To account for the sensitivity of energy consumption in different scenarios, These are the fixed effects estimates for the scenario obtained through a mixed-effects model. These are the estimated interaction terms between the operating system and the scene obtained through a mixed-effects model; The expression for the stability score is: in, , , The proportion of the first variance. This represents the proportion of the second variance. Assess stability score; The overall energy consumption score of the operating system is: in, The operating system is given a comprehensive energy consumption score. As the weight of the global scoring level, The weights for the scene rating levels. The weights for the stability scoring levels.
[0017] Secondly, this application provides an OS energy consumption adaptive scoring system based on multi-fidelity Bayesian optimization, the system being used to execute the above-described method, the system comprising: The data acquisition and processing module is used to collect energy consumption data of the tested operating system and the benchmark operating system, and after calculating the energy consumption ratio, it performs a natural logarithmic transformation to obtain standard energy consumption data. The hierarchical indicator conversion module is used to construct a hybrid effect model based on the standard energy consumption data and convert the parameters of the hybrid effect model into hierarchical indicator parameters. The scoring weight solution module is used to construct a multi-fidelity Bayesian optimization problem based on the hierarchical index parameters with energy consumption ranking consistency as the optimization objective, and adaptively solve the optimal weight of each level of scoring through iterative search. The energy consumption level classification module is used to obtain a comprehensive energy consumption score by weighted summation based on the hierarchical indicator parameters and the optimal weight of the scores of each level, and to classify the energy consumption level through a preset range.
[0018] The technical solution provided in this application may include the following beneficial effects: This application presents an OS energy consumption adaptive scoring method and system based on multi-fidelity Bayesian optimization. It eliminates interference from non-operating system factors such as devices and scenarios through a standardized energy consumption data preprocessing process. Utilizing a mixed-effects model, it accurately decomposes the energy consumption impact mechanism, transforming statistical parameters into hierarchical index parameters reflecting global energy efficiency, scenario adaptability, and stability. Furthermore, with energy consumption ranking consistency as the core optimization objective, it adaptively solves for the optimal weights at each level through multi-fidelity Bayesian optimization iterative search, avoiding the subjectivity of traditional manual weight setting. This balances computational cost and optimization accuracy, significantly improving the efficiency and rationality of weight solving. Simultaneously, a comprehensive energy consumption score is obtained by weighted summarization of hierarchical index parameters and optimal weights. Combined with a flexibly adjustable preset range to define energy consumption levels, this provides a clear picture of the operating system's comprehensive energy consumption performance and adapts to the evaluation needs of different application scenarios such as mobile terminals and servers, significantly improving the objectivity, accuracy, and scenario adaptability of energy consumption evaluation.
[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description
[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.
[0021] Figure 1 A flowchart illustrating the OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization in an exemplary embodiment of this disclosure is shown. Figure 2 A detailed flowchart of step S100 of the OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization in an exemplary embodiment of this disclosure is shown. Figure 3 A detailed flowchart of step S200 of the OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization in an exemplary embodiment of this disclosure is shown. Figure 4 A detailed flowchart of step S300 of the OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization in an exemplary embodiment of this disclosure is shown. Figure 5 A detailed flowchart of step S400 of the OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization in an exemplary embodiment of this disclosure is shown. Figure 6 This diagram illustrates the structure of an OS energy consumption adaptive scoring system based on multi-fidelity Bayesian optimization in an exemplary embodiment of this disclosure. Detailed Implementation
[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
[0023] This example implementation first provides an OS power consumption adaptive scoring method based on multi-fidelity Bayesian optimization. This method can be applied to a terminal device, such as a mobile terminal like a mobile phone, desktop computer, personal digital assistant, laptop, tablet, or smartwatch. (Reference) Figure 1 As shown, the method may include the following steps: Step S100: Collect energy consumption data of the tested operating system and the benchmark operating system, and perform natural logarithmic transformation after calculating the energy consumption ratio to obtain standard energy consumption data.
[0024] Step S200: Construct a mixed-effect model based on the standard energy consumption data, and convert the parameters of the mixed-effect model into hierarchical index parameters.
[0025] Step S300: With energy consumption ranking consistency as the optimization objective, construct an optimization problem based on the hierarchical index parameters, and adaptively solve for the optimal weights of each level through multi-fidelity Bayesian optimization iterative search.
[0026] Step S400: Based on the optimal weights of the hierarchical index parameters and the scores of each level, a comprehensive energy consumption score is obtained through weighted summation, and energy consumption levels are defined according to a preset range.
[0027] The above methods can effectively offset non-OS interference factors such as individual device differences and batch fluctuations, ensuring the standardization and accuracy of energy consumption data. A hierarchical indicator system comprehensively covers the OS's global energy efficiency, scenario adaptability, and device / batch adaptability stability, avoiding evaluation bias from single indicators. Multi-fidelity Bayesian optimization enables adaptive weighting without human subjective bias, adapting to the evaluation needs of different application scenarios. Ultimately, it efficiently outputs accurate comprehensive energy consumption scores and classifications, significantly improving the consistency and reliability of energy consumption evaluation while reducing high-fidelity computation costs, and adapting to the OS energy consumption evaluation needs of various terminals such as smartphones, data center servers, and embedded devices.
[0028] Below, we will refer to Figures 2 to 5 The steps of the method described above in this example embodiment will be explained in more detail.
[0029] In step S100, energy consumption data of the tested operating system and the benchmark operating system are collected, and after calculating the energy consumption ratio, a natural logarithmic transformation is performed to obtain standard energy consumption data.
[0030] It should be noted that this step eliminates interference from non-operating system factors such as equipment, scenario, and batch to the energy consumption data, and improves the data distribution characteristics to adapt to the requirements of subsequent statistical modeling; the construction of the subsequent mixed-effect model and the transformation of hierarchical index parameters all use the standard energy consumption data output in this step as the sole input source.
[0031] In one embodiment, such as Figure 2 As shown, step S100 may include the following sub-steps.
[0032] In step S110, energy consumption data of the tested operating system and the benchmark operating system are collected under the same device, in the same scenario, and in the same batch.
[0033] It should be noted that the constraints of the same device, same scenario, and same batch are to ensure that the comparison benchmark for energy consumption data is consistent. Among them, the same batch covers production dimensions that may affect energy consumption, such as device hardware batch and battery power batch. This constraint can eliminate the interference of non-operating system factors such as hardware and environment to the greatest extent, so that the energy consumption differences in subsequent calculations only reflect the characteristics of the operating system itself.
[0034] In step S120, the energy consumption ratio between the energy consumption data of the tested operating system and the energy consumption data of the benchmark operating system is calculated, and background noise is removed.
[0035] It should be noted that the calculation of the energy consumption ratio is essentially that background factors such as device hardware and environmental load that jointly affect energy consumption will cancel each other out in the ratio calculation, thus retaining only the energy consumption difference brought by the operating system itself, thereby removing background noise.
[0036] The formula for calculating the energy consumption ratio is as follows: in, This is the energy consumption ratio. This refers to the power consumption data of the operating system under test. This is the power consumption data for the benchmark operating system.
[0037] It should be noted that s, d, and b in the formula correspond to the identification parameters of the scene, device, and batch, respectively. This means that the ratio is an independent calculation result under "single scene, single device, single batch". In experiments with multiple devices, multiple scenes, and multiple batches, the corresponding energy consumption ratio needs to be calculated for each combination of (s, d, b) and then summarized to form the basic dataset for subsequent modeling.
[0038] In step S130, the energy consumption ratio is transformed by natural logarithm to improve the normality of the data.
[0039] It should be noted that in actual implementation, the data distribution of energy consumption ratios is usually right-skewed, that is, the proportion of samples with extremely high energy consumption is low but the numerical differences are large. This skewed distribution will lead to deviations in the parameter estimation of subsequent linear mixed-effects models. The natural logarithm transformation can equivalently transform the geometric mean of the original space into the arithmetic mean of the logarithmic space, effectively improving the normality of the data and making the data more suitable for the assumptions of the linear statistical model.
[0040] The formula for calculating the standard energy consumption data is as follows: in, This is standard energy consumption data.
[0041] It should be noted that, The numerical meaning of these values is directly related to energy consumption performance: when When <0, the corresponding <1 indicates that the tested operating system consumes less power than the benchmark operating system, meaning it is more energy-efficient; when When >0, the corresponding A value >1 indicates that the power consumption of the tested operating system is higher than that of the benchmark operating system; after this conversion, The distribution is closer to a normal distribution, and subsequent models can be based on this to achieve a more accurate decomposition of energy consumption effects.
[0042] In step S200, a hybrid effect model is constructed based on the standard energy consumption data, and the parameters of the hybrid effect model are converted into hierarchical index parameters.
[0043] It should be noted that this step transforms the standardized energy consumption data output from step S100 from raw values into hierarchical parameters that can characterize the energy consumption characteristics of the OS. By using a mixed-effects model to separate energy consumption influencing factors of different dimensions, it can accurately locate the energy consumption performance of the OS itself and quantify the fluctuations of interference factors such as devices and batches, providing an indicator carrier with clear physical meaning for subsequent weight calculation.
[0044] In one embodiment, such as Figure 3 As shown, step S200 may include the following sub-steps.
[0045] In step S210, a structured data table is constructed based on the standard energy consumption data, and a mixed-effect model is constructed.
[0046] It should be noted that constructing structured data tables, such as This method is used to organize scattered standard energy consumption data into a model-recognizable format according to the correspondence between sample and effect factor. The mixed-effects model includes global average effect, scene fixed effect, and OS-scene interaction term, which are fixed effects reflecting the deterministic energy consumption characteristics of OS and scene. Equipment volume difference and batch difference are random effects reflecting the random effects of hardware and production batch. This hierarchical modeling method can simultaneously take into account the inherent characteristics of OS energy consumption and the random effects of experimental interference, avoiding the result bias caused by the traditional single model ignoring interference factors.
[0047] The mixed-effects model is as follows: in, , , , For standard energy consumption data, For global average effect, For scene fixed effect, For the interaction between the operating system and the scene, Due to equipment differences, This represents the variance due to equipment differences. Due to batch differences, For batch variation variance, For residuals, Other noise during measurement.
[0048] It should be noted that in fixed effects, This represents the overall energy consumption ratio of the tested OS relative to the reference OS. This represents the average difference across different scenarios (e.g., video playback vs. file operations); in random effects, This indicates whether the OS behaves consistently across different scenarios; This indicates the impact of differences in hardware between different devices; This indicates the impact of production / battery batches.
[0049] In step S220, the parameters of the mixed-effects model are estimated by the maximum likelihood method to obtain the mixed-effects model parameters; the mixed-effects model parameters include global average effect, scene fixed effect, interaction term between operating system and scene, device volume difference, batch difference and residual.
[0050] It should be noted that the maximum likelihood method was chosen because it is suitable for the hierarchical structure of mixed-effects models and can simultaneously provide optimal estimates of the variance parameters of both fixed-effects and random-effects. Compared with ordinary least squares, it can more accurately capture the stratified fluctuations of the data. The parameters output in this step are abstract statistics and do not have direct scoring significance. They need to be transformed by S230 to become weighted hierarchical indicators. Therefore, the accuracy of parameter estimation directly affects the reliability of subsequent scoring results.
[0051] In step S230, the parameters of the mixed effect model are converted into hierarchical index parameters; the hierarchical index parameters are equipment difference variance, batch difference variance and other noise during measurement.
[0052] It should be noted that, as shown in Table 1, this transformation is an engineered mapping of statistical parameters. The model parameters need to match the scoring objective. For example, the global average effect μ corresponds to the core parameter of the global score, and the scene fixed effect... Scene interaction items The parameters for the corresponding scenario scoring, and the device difference variance. Batch variation variance These correspond to the parameters of the stability score.
[0053] Table 1 In step S300, with energy consumption ranking consistency as the optimization objective, an optimization problem is constructed based on the hierarchical index parameters, and the optimal weights of each level are adaptively solved through multi-fidelity Bayesian optimization iterative search.
[0054] It should be noted that this step can solve the problem of strong subjectivity in the manual setting of weights in traditional energy consumption scoring. With the consistency of energy consumption ranking as the optimization goal, it means that the selection of weights should be consistent with the actual needs of the comprehensive score and the actual energy consumption performance. The technical path of multi-fidelity Bayes optimization achieves a balance between computational cost and optimization accuracy. The optimal weights output will be directly used for the weighted calculation of the subsequent comprehensive energy consumption score.
[0055] The AUC (Area Under ROC Curve) metric measures ranking consistency. Using AUC as the objective factor to determine weights allows us to determine that "OS with higher overall scores is more likely to be a true low-energy label." "The label is based on the actual power consumption of the operating system under test measured in step S100." Labeling. For example, using Top-K pseudo-labels, sorting by energy consumption from low to high within a single conditional block, and labeling the top K%. The rest are .
[0056] That is, the probability of correctly ranking the scores when comparing each pair of positive samples (true low-energy OS) with negative samples (high-energy OS).
[0057] In one embodiment, such as Figure 4 As shown, step S300 may include the following sub-steps: In step S310, based on the hierarchical index parameters and through labels labeled with actual energy consumption data, an optimization problem is constructed with the goal of energy consumption ranking consistency.
[0058] It should be noted that the labels here are Top-K pseudo-labels based on actual energy consumption data. That is, they are sorted from low to high based on the actual energy consumption within a single condition block, and the top K% are labeled as low-energy-consumption positive samples, while the rest are high-energy-consumption negative samples. Their purpose is to provide an objective reference benchmark for the consistency of energy consumption ranking. The constructed optimization problem has variables that are weight vectors of global score, scene score, and stability score. The constraint is that the weights are non-negative and their sum is 1. This constraint ensures that the weights conform to the physical meaning of the scoring system and avoids mathematical solutions without practical interpretation.
[0059] The optimization problem is represented as follows: in, , , ; in, This is the weight vector used for comprehensive energy consumption scoring; To make the objective function within the feasible region The weight vector that achieves the maximum value; The weighted feasible region is a three-dimensional probabilistic simplex. This serves as a consistency indicator for ranking. For sorting labels; This is the weighted feature vector.
[0060] In step S320, a multi-fidelity Bayesian optimization model is constructed through a joint Gaussian process. The multi-fidelity Bayesian optimization model performs coarse screening of the weight space using a low-fidelity approximation index, and then performs local precise optimization using a high-fidelity energy consumption ranking consistency.
[0061] It should be noted that the multi-fidelity of this model is proposed for the engineering practicality of weight solution. The low-fidelity approximation index has extremely low computational cost, can quickly traverse the weight space and lock the potential optimal solution region, avoiding indiscriminate high-cost computation, such as Fisher's discrimination ratio; the high-fidelity energy consumption ranking consistency is the direct quantification of the optimization objective, and accurately evaluates the region after coarse screening. The combination of the two can achieve an efficiency balance between low-cost range locking and high-precision local optimization.
[0062] The multi-fidelity Bayesian optimization model is as follows: in, For low-fidelity optimization model, To optimize the model for high fidelity, For scaling / offset relationships from low-fidelity to high-fidelity, This is the first independent Gaussian process in the joint Gaussian process. It is the second independent Gaussian process in the joint Gaussian process.
[0063] It should be noted that in the model and The independent characteristics are intended to separate random noise from low-fidelity and high-fidelity evaluation results, and to prevent interference factors from affecting each other in the two evaluation layers; while the scaling / offset relationship It is a statistical association obtained by fitting sample data. Its function is to map the low-fidelity evaluation results to the high-fidelity target space, so that the low-fidelity samples can effectively assist the prediction of the high-fidelity target, thereby reducing the number of high-cost high-fidelity evaluations.
[0064] Scene modeling can be represented as: low fidelity is Fisher's discrimination ratio. High fidelity is AUC .
[0065] In step S330, based on the multi-fidelity Bayesian optimization model, the optimal weights at each level are adaptively solved by iteratively searching the cost-weighted expectation improvement function under budget constraints.
[0066] It should be noted that the budget constraint refers to the computational resource limitation of the high-fidelity assessment, while the core logic of the cost-weighted expected value improvement function is to select the weight sample points with the greatest potential for increasing the fidelity target (AUC) and the lowest computational cost, and to gradually converge the weights to the global optimum through iterative search. This process does not require manual intervention, and the final output of the optimal weights not only conforms to the physical constraints of the scoring system, but also maximizes the consistency between the comprehensive score and the actual energy consumption ranking, ensuring the objectivity and effectiveness of the weights.
[0067] The cost-weighted expected improvement function is as follows: in, To improve the a posteriori expectation of the current joint Gaussian process for this fidelity, To assess costs.
[0068] It should be noted that this function is used to implement multi-fidelity Bayesian optimization iterative search: Quantify the potential improvement of the current candidate weights for the high-fidelity target. The computational cost corresponding to the fidelity assessment is considered. By selecting the next assessment point based on the ratio of expected improvement to assessment cost, we avoid wasting resources by pursuing high improvement at the expense of cost, and also prevent inefficient searches by focusing only on low cost and abandoning high-quality weights, ultimately achieving optimal search efficiency under budget constraints.
[0069] Furthermore, the overall algorithm flow can be represented as follows: 1. Input: Parameter set obtained from the mixed-effects model With tags ,Budget Cost vector (The initial sample size can be determined based on the actual measurement calculation time.) , ; 2. Initialization: Sampling on the simplex Group weights, evaluation Sample a small amount Group weights, evaluation ; 3. Fitting MF-GP (co-kriging): Estimate kernel parameters using existing data and (MLE method); 4. Iteration: exist Spatial data acquisition is performed, the acquisition function EI / cost is calculated, and selection is made. Maximum point Evaluate at the selected fidelity level, and calculate Fisher's or AUC. Add the results to the dataset and update the co-kriging model and posterior distribution. Update remaining budget ; 5. Output: Selected based on the high-fidelity mean of the co-kriging posterior. .
[0070] In step S400, based on the hierarchical index parameters and the optimal weight of each level score, a comprehensive energy consumption score is obtained through weighted summation, and energy consumption levels are defined according to a preset range.
[0071] It should be noted that this step integrates the hierarchical indicator parameters and optimal weights from the preceding steps into a quantitative result that can directly guide decision-making. Based on the logic that the global score reflects overall energy efficiency, the scenario score reflects scenario adaptability, and the stability score reflects hardware / batch compatibility, the comprehensive score can fully characterize the OS's energy consumption performance; and the preset range can be flexibly adjusted according to the energy efficiency requirements of different application scenarios such as mobile terminals and servers.
[0072] In one embodiment, such as Figure 5 As shown, step S400 may include the following sub-steps: In step S410, a hierarchical score is calculated based on the hierarchical index parameters; wherein the hierarchical score is a global score, a scene score, and a stability score.
[0073] It should be noted that the statistical hierarchical index parameters output by S230 are converted into quantitative scores with clear business meanings. The global score corresponds to the overall energy saving level of the OS, the scenario score corresponds to the energy consumption consistency of the OS in different scenarios, and the stability score corresponds to the energy consumption fluctuation of the OS in different devices / batches.
[0074] The expression for the global score is: in, For overall energy efficiency, This is the global average effect estimate obtained through the mixed-effects model; The expression for the scene rating is: in, , As a consistency indicator, To account for the sensitivity of energy consumption in different scenarios, These are the fixed effects estimates for the scenario obtained through a mixed-effects model. These are the estimated interaction terms between the operating system and the scene obtained through a mixed-effects model; The expression for the stability score is: in, , , The proportion of the first variance. This represents the proportion of the second variance. For stability rating.
[0075] In step S420, the comprehensive energy consumption score is calculated using the hierarchical score and the optimal weight of each level.
[0076] It should be noted that the performance of the three dimensions—the overall score reflecting overall energy efficiency, the scenario consistency reflecting scenario energy consumption stability, and the stability score reflecting equipment / batch adaptability—is weighted and integrated according to their actual importance to the energy consumption assessment.
[0077] The overall energy consumption score of the operating system is as follows: in, The operating system is given a comprehensive energy consumption score. As the weight of the global scoring level, The weights for the scene rating levels. The weights for the stability scoring levels.
[0078] It should be noted that the formula uses Instead of using directly Because It is the extreme difference in energy consumption fluctuations in the scenario. It can be converted into a positive scenario energy consumption consistency indicator; Weights in the formula , , The optimal combination obtained through multi-fidelity Bayesian optimization in step S330 strictly satisfies + + Physical constraints with a value of 1 and all weights being non-negative avoid mathematical solutions that are meaningless in practice; while the overall score... The numerical meaning of is intuitive. The smaller the value, the lower the overall energy consumption of the tested OS relative to the benchmark OS, and vice versa.
[0079] In step S430, the energy consumption level is determined based on the preset energy consumption level range and the comprehensive energy consumption score.
[0080] It should be noted that the preset energy consumption level ranges are usually determined by combining industry energy efficiency standards and the actual needs of different application scenarios, and are not fixed thresholds. For example, preset energy consumption level ranges are shown in Table 2.
[0081] Table 2 Furthermore, this example implementation also provides an OS power consumption adaptive scoring system based on multi-fidelity Bayesian optimization. (Reference) Figure 6 As shown, the system may include: The data acquisition and processing module is used to collect energy consumption data of the tested operating system and the benchmark operating system, and after calculating the energy consumption ratio, it performs a natural logarithmic transformation to obtain standard energy consumption data. The hierarchical indicator conversion module is used to construct a hybrid effect model based on the standard energy consumption data and convert the parameters of the hybrid effect model into hierarchical indicator parameters. The scoring weight solution module is used to construct a multi-fidelity Bayesian optimization problem based on the hierarchical index parameters with energy consumption ranking consistency as the optimization objective, and adaptively solve the optimal weight of each level of scoring through iterative search. The energy consumption level classification module is used to obtain a comprehensive energy consumption score by weighted summation based on the hierarchical indicator parameters and the optimal weight of the scores of each level, and to classify the energy consumption level through a preset range.
[0082] The aforementioned system enables standardized preprocessing of energy consumption data by the data acquisition and processing module, completely eliminating non-OS interference factors such as equipment and batches. Then, a hierarchical indicator conversion module constructs a hierarchical indicator system covering global energy efficiency, scenario adaptability, and device adaptability stability. A scoring weight solution module uses multi-fidelity Bayesian optimization to achieve subjectively bias-free adaptive weight solution. Finally, the energy consumption level classification module outputs a quantitative comprehensive energy consumption score and level, realizing full-process automation, indicator systematization, and result accuracy in OS energy consumption evaluation. This reduces high-fidelity computing costs and is adaptable to OS energy consumption evaluation scenarios for various terminals such as smartphones, data center servers, and embedded devices, providing efficient and reliable tool support for OS energy efficiency assessment, optimization iteration, and selection deployment.
[0083] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.
Claims
1. An OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization, characterized in that, include: Collect energy consumption data of the tested operating system and the benchmark operating system, and perform a natural logarithmic transformation after calculating the energy consumption ratio to obtain standard energy consumption data; A mixed-effect model is constructed based on the standard energy consumption data, and the parameters of the mixed-effect model are converted into hierarchical index parameters. With energy consumption ranking consistency as the optimization objective, an optimization problem is constructed based on the hierarchical index parameters, and the optimal weights of each level are adaptively solved through multi-fidelity Bayesian optimization iterative search. Based on the optimal weights of the hierarchical index parameters and the scores of each level, a comprehensive energy consumption score is obtained through weighted summation, and energy consumption levels are defined according to a preset range.
2. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 1, characterized in that, The steps of collecting energy consumption data of the tested operating system and the benchmark operating system, and performing a natural logarithmic transformation after calculating the energy consumption ratio to obtain standard energy consumption data include: Collect energy consumption data of the tested operating system and the benchmark operating system under the same device, in the same scenario, and in the same batch; Calculate the energy consumption ratio between the tested operating system's energy consumption data and the benchmark operating system's energy consumption data, and remove background noise; The energy consumption ratio is transformed by natural logarithm to improve the normality of the data.
3. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 2, characterized in that, The formula for calculating the energy consumption ratio is: in, This is the energy consumption ratio. This refers to the power consumption data of the operating system under test. This is the power consumption data for the benchmark operating system; The formula for calculating the standard energy consumption data is: in, This is standard energy consumption data.
4. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 1, characterized in that, The step of constructing a mixed-effect model based on the standard energy consumption data and converting the parameters of the mixed-effect model into hierarchical index parameters includes: A structured data table is constructed based on the standard energy consumption data, and a mixed-effect model is built. The parameters of the mixed-effects model are estimated by means of the maximum likelihood method. The mixed-effects model parameters include global average effect, scene fixed effect, interaction term between operating system and scene, device volume difference, batch difference and residual. The parameters of the mixed-effects model are converted into hierarchical index parameters; the hierarchical index parameters are equipment variance variance, batch variance variance, and other noise during measurement.
5. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 4, characterized in that, The mixed-effects model is as follows: in, , , , For standard energy consumption data, For global average effect, For scene fixed effect, For the interaction between the operating system and the scene, Due to equipment differences, For equipment variation variance. Due to batch differences, For batch variation variance, For residuals, Other noise during measurement.
6. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 1, characterized in that, The steps of constructing an optimization problem based on the hierarchical index parameters, with energy consumption ranking consistency as the optimization objective, and adaptively solving for the optimal weights at each level through multi-fidelity Bayesian optimization iterative search, include: Based on the hierarchical index parameters and using labels based on actual energy consumption data, an optimization problem is constructed with the goal of energy consumption ranking consistency. A multi-fidelity Bayesian optimization model is constructed by a joint Gaussian process; the multi-fidelity Bayesian optimization model performs coarse screening of the weight space through a low-fidelity approximation index, and then performs local precise optimization through a high-fidelity energy consumption ranking consistency. Based on the aforementioned multi-fidelity Bayesian optimization model, the optimal weights for each level are adaptively solved through iterative search using a cost-weighted expectation improvement function under budget constraints.
7. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 6, characterized in that, The multi-fidelity Bayesian optimization model is as follows: in, For low-fidelity optimization model, To optimize the model for high fidelity, For scaling / offset relationships from low-fidelity to high-fidelity, This is the first independent Gaussian process in the joint Gaussian process. It is the second independent Gaussian process in the joint Gaussian process; The cost-weighted expected improvement function is: in, To improve the a posteriori expectation of the current joint Gaussian process for this fidelity, To assess costs.
8. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 1, characterized in that, The step of obtaining a comprehensive energy consumption score by weighted summation based on the optimal weights of the hierarchical index parameters and the scores of each level, and defining the energy consumption level according to a preset range, includes: The hierarchical score is calculated based on the hierarchical index parameters; wherein, the hierarchical score is a global score, a scene score, and a stability score; The comprehensive energy consumption score is calculated using the hierarchical scores and the optimal weights for each level. Based on a preset energy consumption level range, the energy consumption level is determined according to the comprehensive energy consumption score.
9. The OS energy consumption adaptive scoring method based on multi-fidelity Bayesian optimization according to claim 8, characterized in that, The expression for the global score is: in, For overall energy efficiency, This is the global average effect estimate obtained through the mixed-effects model; The expression for the scene rating is: in, , As a consistency indicator, To account for the sensitivity of energy consumption in different scenarios, These are the fixed effects estimates for the scenario obtained through a mixed-effects model. These are the estimated interaction terms between the operating system and the scene obtained through a mixed-effects model; The expression for the stability score is: in, , , The proportion of the first variance. This represents the proportion of the second variance. Assess stability score; The overall energy consumption score of the operating system is: in, The operating system is given a comprehensive energy consumption score. As the weight of the global scoring level, The weights for the scene rating levels. The weights for the stability scoring levels.
10. An OS energy consumption adaptive scoring system based on multi-fidelity Bayesian optimization, characterized in that, The system is used to perform the method as described in any one of claims 1 to 9, the system comprising: The data acquisition and processing module is used to collect energy consumption data of the tested operating system and the benchmark operating system, and after calculating the energy consumption ratio, it performs a natural logarithmic transformation to obtain standard energy consumption data. The hierarchical indicator conversion module is used to construct a hybrid effect model based on the standard energy consumption data and convert the parameters of the hybrid effect model into hierarchical indicator parameters. The scoring weight solution module is used to construct a multi-fidelity Bayesian optimization problem based on the hierarchical index parameters with energy consumption ranking consistency as the optimization objective, and adaptively solve the optimal weight of each level of scoring through iterative search. The energy consumption level classification module is used to obtain a comprehensive energy consumption score by weighted summation based on the hierarchical indicator parameters and the optimal weight of the scores of each level, and to classify the energy consumption level through a preset range.