Automated resource configuration policy optimization method, apparatus, and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and provides an automatic resource configuration strategy optimization method, device and storage medium. The method comprises the following steps: receiving an optimization target, determining a strategy parameter set associated with the optimization target, and defining the value of each parameter in the strategy parameter set; based on each parameter in the strategy parameter set and the defined value of the parameter, applying orthogonal experiment design to generate an orthogonal experiment table containing M strategy combinations; inputting the M strategy combinations into a prediction model based on a pre-trained machine learning model, and outputting a virtual effect prediction value corresponding to each combination; based on the virtual effect prediction value, filtering out the first N strategies with the optimal prediction effect as high-potential candidate strategies through a decision tree, an integrated learning model, a support vector machine, a deep learning model, a reinforcement learning model or a privacy computing technology; and creating the high-potential candidate strategies into real execution plans and performing monitoring.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an automated resource allocation strategy optimization method, apparatus, and storage medium. Background Technology

[0002] In fields such as digital marketing and computing resource scheduling that require automated strategy optimization, the effectiveness of a strategy is typically determined by multiple configurable parameters (e.g., bidding, targeting conditions, display time, creative content, etc.). Different combinations of these parameters constitute a vast strategy space. Traditionally, finding the optimal or near-optimal strategy combination relies on expert experience or simple rules, which have limited optimization effectiveness and are difficult to adapt to dynamically changing environments.

[0003] Currently, A / B testing, multivariate testing (MVT), or grid search are commonly used in practice for strategy optimization. A / B testing typically only tests a few variables at a time, resulting in low efficiency. While MVT and grid search can test multiple parameters simultaneously, the number of complete experimental combinations required increases exponentially when there are many parameters and values. This leads to lengthy testing cycles and enormous computational and real-world deployment resource consumption, making it difficult to conduct frequently and on a large scale in actual business operations. Furthermore, these methods are essentially based on "trial and error" using real traffic, which may consume significant resources on strategies that show significantly poor results, leading to resource waste.

[0004] Therefore, the shortcomings of existing technologies lie in the lack of an automated optimization method that can efficiently explore a broad policy space within a small number of experiments and pre-evaluate the potential effects of policies, thereby significantly reducing the consumption of real resources, when faced with complex policy optimization problems involving multiple parameters and levels. This results in inefficient and costly optimization processes, and makes it difficult to achieve real-time or near-real-time policy tuning. Summary of the Invention

[0005] This application provides an automated resource allocation strategy optimization method, apparatus, and storage medium. By integrating multi-time granular data with heterogeneous models, it can more comprehensively capture the temporal patterns and complex relationships of promotion effects, thereby improving the accuracy and robustness of predictions and providing multi-dimensional information for deployment decisions.

[0006] On the one hand, this application provides an automated resource allocation strategy optimization method, the method comprising:

[0007] Step S1: Receive the optimization target, determine the set of strategy parameters associated with the optimization target, and define the value of each parameter in the set of strategy parameters;

[0008] Step S2: Based on each parameter and its defined value, orthogonal experimental design is applied to generate an orthogonal experimental table containing M strategy combinations, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically cover the effects of each parameter and its interaction in a balanced manner.

[0009] Step S3: Input the M strategy combinations into the pre-trained prediction model based on the machine learning model, and output the virtual effect prediction value corresponding to each combination;

[0010] Step S4: Based on the virtual effect prediction value, select the top N strategies with the best prediction effect as high-potential candidate strategies through decision tree, ensemble learning model, support vector machine, deep learning model, reinforcement learning model or privacy computing technology, where N is a positive integer less than M.

[0011] Step S5: Create the high-potential candidate strategies as actual execution plans and monitor their deployment.

[0012] On the other hand, this application provides an automated resource allocation strategy optimization device, the device comprising:

[0013] The determination module is used to receive the optimization target, determine the set of strategy parameters associated with the optimization target, and define the value of each parameter in the set of strategy parameters;

[0014] The generation module is used to generate an orthogonal experimental table containing M strategy combinations based on each parameter and its defined values, using orthogonal experimental design, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically balance the influence of each parameter and its interaction.

[0015] The prediction module is used to input the M strategy combinations into a pre-trained prediction model based on a machine learning model, and output the virtual effect prediction value corresponding to each combination.

[0016] The filtering module is used to select the top N strategies with the best prediction results as high-potential candidate strategies based on the virtual effect prediction value, through decision tree, ensemble learning model, support vector machine, deep learning model, reinforcement learning model or privacy computing technology, where N is a positive integer less than M.

[0017] The monitoring module is used to create the high-potential candidate strategies as actual execution plans and monitor their deployment.

[0018] Thirdly, this application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the technical solution of the automated resource allocation strategy optimization method described above.

[0019] Fourthly, this application provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described automated resource allocation strategy optimization method.

[0020] As can be seen from the technical solution provided in this application, on the one hand, by applying orthogonal experimental design to generate strategy combinations, the technical solution of this application can statistically cover the influence of each parameter and its interaction with a much smaller number of permutations and combinations. This means that only a small number of carefully designed representative strategy combinations need to be evaluated to effectively infer the influence trend of each factor in the entire strategy space, thereby greatly reducing the number of strategies that need to be examined, solving the combination explosion problem caused by massive combinations, and laying the foundation for rapid optimization. On the other hand, by introducing a prediction model based on machine learning models to predict the virtual effects of the generated strategy combinations, the technical solution of this application can estimate the potential effects of a large number of strategy combinations in advance and in parallel before the actual deployment. This replaces... This approach replaces the traditional method of obtaining feedback through real A / B testing or online canary releases, avoiding the waste of real traffic and budget on potentially ineffective strategies. It achieves low-cost, near-zero-risk strategy pre-screening. Thirdly, the technical solution combines efficient spatial sampling (i.e., orthogonal experimental design), intelligent virtual evaluation (i.e., machine learning prediction), and automated strategy selection and execution, forming a coherent automated process. This process automatically generates candidate strategies based on the optimization objective, intelligently predicts their effects, selects the optimal subset, and executes them. Human intervention is minimal throughout the process. This not only reduces reliance on professional experience but also enables strategy optimization to quickly and automatically adapt to constantly changing environments and objectives, improving the overall intelligence and responsiveness of the optimization process. In summary, the technical solution of this application significantly reduces the number of strategy combinations to be evaluated through orthogonal experimental design and utilizes machine learning models for virtual effect prediction, thereby rapidly selecting high-potential strategies at extremely low cost. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the automated resource allocation strategy optimization method provided in the embodiments of this application;

[0023] Figure 2 This is a schematic diagram of the structure of the automated resource allocation strategy optimization device provided in the embodiments of this application;

[0024] Figure 3 This is a schematic diagram of the device provided in the embodiments of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] In this specification, adjectives such as "first" and "second" are used only to distinguish one element or action from another, without necessarily requiring or implying any actual such relationship or order. Where circumstances permit, reference to an element or component or step (etc.) should not be construed as being limited to only one of the elements, components, or steps, but may be one or more of the elements, components, or steps, etc.

[0027] For ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale.

[0028] Currently, A / B testing, multivariate testing (MVT), or grid search are commonly used for strategy optimization in practice. A / B testing typically tests only a few variables at a time, resulting in low efficiency. While MVT and grid search can test multiple parameters simultaneously, the number of complete experimental combinations required increases exponentially when there are many parameters and values. This leads to lengthy testing cycles and enormous computational and real-world resource consumption, making it difficult to conduct frequently and on a large scale in actual business applications. Furthermore, these methods are essentially based on "trial and error" with real traffic, potentially consuming significant resources on strategies with clearly poor results, leading to resource waste. Therefore, the current technology suffers from the following deficiency: when facing complex strategy optimization problems with multiple parameters and levels, there is a lack of an automated optimization method that can efficiently explore a broad strategy space within a small number of experiments, pre-evaluate the potential effects of strategies, and thus significantly reduce the consumption of real-world resources. This results in inefficient and costly optimization processes, and makes it difficult to achieve real-time or near-real-time strategy tuning.

[0029] To address the aforementioned problems in the prior art, this application proposes an automated resource allocation strategy optimization method, the flowchart of which is attached. Figure 1 As shown, it mainly includes steps S1 to S5, which are detailed below:

[0030] Step S1: Receive the optimization target, determine the set of strategy parameters associated with the optimization target, and define the value of each parameter in the set of strategy parameters.

[0031] In this embodiment, the optimization objective is specified by the user or the upstream system. For example, in a digital marketing scenario, it could be maximizing return on investment (ROI), maximizing conversions, or controlling the cost per conversion (CPA) below a certain threshold. After receiving the optimization objective, the system needs to determine which strategy parameters (i.e., adjustable variables) adjustments will affect that objective.

[0032] In one specific implementation, determining the set of strategy parameters does not involve simply listing all adjustable parameters, but rather conducting intelligent filtering to focus on key influencing factors, thereby improving the efficiency and targeting of subsequent optimizations. This process can be further implemented through the following sub-steps: determining the set of strategy parameters associated with the optimization objective can be achieved through steps S1.1 and S1.2, detailed below:

[0033] Step S1.1: Extract features from historical deployment data and use a feature importance assessment algorithm to identify the K core parameters that rank highest in terms of their impact on the optimization objective.

[0034] Specifically, the system collects past advertising campaign logs. Each log entry contains the strategy parameter combinations used in a campaign (e.g., bid, audience targeting tags, creative ID, campaign time, etc.) and the final performance metrics achieved (e.g., conversions, spend). By processing this data, a supervised learning dataset can be constructed, where features are the strategy parameters (which can be processed using one-hot encoding, etc.) and labels are the performance metrics corresponding to the optimization target. Then, using a feature importance evaluation algorithm, such as training a Gradient Boosting Decision Tree (GBDT) model, after model training, the importance score of each feature to the prediction target is quantified by calculating the sum of the impurity reduction resulting from each feature branching across all decision trees in the model. Based on the scores, the top K core parameters with the highest impact on the optimization target can be identified. This data-driven approach avoids biases that may arise from subjective experience, ensuring the scientific rigor of parameter selection.

[0035] Step S1.2: Select the parameters that can be adjusted within a preset time period from the K core parameters to form a set of strategy parameters.

[0036] After identifying the core parameters, their feasibility in actual operation must also be considered. Some parameters may not be modifiable within a specific time period due to business rules, technical limitations, or contractual agreements. Traditional static whitelist methods cannot adapt to dynamic environments. Therefore, this application adopts a dynamic adaptive filtering mechanism. That is, the parameters that can be adjusted within a preset time period from the K core parameters in the above embodiment can be achieved through steps S1.2a to S1.2c, as detailed below:

[0037] Step S1.2a: Monitor the current market environment status characteristics in real time.

[0038] The current market environment can be characterized by factors such as the intensity of market competition, major holidays, updates to relevant policies and regulations, and peak traffic periods on the platform.

[0039] Step S1.2b: Match the corresponding parameter adjustability rules from the pre-configured rule base according to the characteristics of the market environment.

[0040] The rule base predefines the logical relationships between various environmental states and parameter adjustability. For example, one rule might be: "If it is currently a major holiday, the 'deployment period' parameter is locked and cannot be adjusted"; another rule might be: "If the market competition index exceeds the threshold X, the 'bid' parameter is marked as a high-priority adjustable parameter".

[0041] Step S1.2c: Based on the matched rules, dynamically select the parameters that can be adjusted from the K core parameters.

[0042] The system applies all matched rules to filter out adjustable parameters from K core parameters within a preset time period. This set of parameters ultimately constitutes the strategy parameter set. This mechanism ensures the optimization system can perceive environmental changes and always optimizes within the feasible parameter space, improving the system's practicality and robustness. After determining the parameter set, the values ​​or ranges to be examined for each parameter need to be defined. For example, the "bid" parameter can be defined as [1.0, 1.5, 2.0, 2.5] yuan; the "age group orientation" parameter can be defined as [18-25, 26-35, 36-45]. These values ​​form the basis for the next step of experimental design.

[0043] The technical solutions described in steps S1.2a to S1.2c above effectively solve the problem of how to intelligently and in real-time adjust the set of optimizable parameters according to the dynamic environment. By introducing a rule engine that is linked to the environment, the system achieves dynamic self-adaptation of the set of strategy parameters. This enables the optimization system to perceive changes in the environment (e.g., intensified market competition or policy adjustments) and automatically focus on the parameters that are most likely to have an impact and are operable, thereby improving the real-time performance, adaptability, and intelligence of the optimization system.

[0044] As another embodiment of this application, selecting adjustable parameters from K core parameters within a preset time period can also involve: constructing a causal graph model representing the structured relationship between strategy parameters, environmental variables, and effect indicators using historical data; when receiving a new task, based on the current environmental variables, performing intervention reasoning through the causal graph model to quantitatively estimate the average causal effect of each core parameter in the current environment; and selecting parameters whose average causal effect exceeds a preset threshold to form a set of strategy parameters. Compared to the technical solutions exemplified in steps S1.2a to S1.2c, this solution addresses the essential questions of "why adjust this parameter" and "how much expected benefit it can bring," identifying parameters that truly have a driving effect in a specific environment, eliminating spurious associations, and making parameter selection decisions more scientific and interpretable.

[0045] Step S2: Based on each parameter and its defined value, apply orthogonal experimental design to generate an orthogonal experimental table containing M strategy combinations, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically cover the effects of each parameter and its interaction in a balanced manner.

[0046] When there are many policy parameters and a large number of possible values, the sheer number of permutations and combinations (i.e., combinatorial explosion) leads to extremely low efficiency in policy space exploration. Traditional methods cannot effectively cover and evaluate such a vast space in a limited number of experiments. For example, grid search, an existing approach, tests all possible combinations for each parameter. However, the number of experiments required for this approach increases exponentially with the number of parameters, making it completely infeasible in complex scenarios due to unacceptable costs and time. Another existing approach is random search, which randomly selects several combinations in the policy space for testing. Its drawback is that the selected combinations may lack representativeness, failing to systematically and evenly examine the influence of each parameter and its interactions, easily overlooking optimal solution regions, and resulting in an unstable and unscientific optimization process.

[0047] To address this, this application introduces the principle of orthogonal experimental design from the field of mathematical statistics. The technical solution involves generating an orthogonal experimental table containing M strategy combinations based on each parameter and its defined values. Here, M is a positive integer much smaller than the total number of permutations of the parameter values, and the M combinations statistically evenly cover the influence of each parameter and its interactions. Orthogonal experimental design is a scientific method for studying and handling multi-factor, multi-level experimental problems. It utilizes a pre-compiled, standardized table—an orthogonal table—to scientifically arrange experiments. By introducing this mathematical statistical method, this approach can evenly and efficiently cover all parameters and their interactions between different levels with a very small number (M) of statistically representative experimental combinations. This allows the system to use a carefully designed small-scale experimental scheme to scientifically and systematically explore a vast strategy space, laying an efficient and reliable data foundation for subsequent intelligent evaluation and fundamentally solving the bottleneck of exploration efficiency.

[0048] As an embodiment of this application, based on each parameter and its defined value, generating an orthogonal experimental table containing M strategy combinations using orthogonal experimental design can be achieved through the following steps S2.1 to S2.3:

[0049] Step S2.1: Select an orthogonal array with matching order from the standard orthogonal array library based on the number of values ​​for each parameter.

[0050] Orthogonal arrays are usually represented as The orthogonal array is in the form of L, where L represents an orthogonal array, n is the number of experiments (i.e., the number of rows), t is the number of levels, and c is the number of factors (i.e., the number of columns). For example, for the scenario of 5 factors and 4 levels mentioned above, an orthogonal array can be chosen. This means that only 16 experiments (n=16, i.e., M=16) are needed to evenly arrange five experiments with four levels of factors. Since 16 is much smaller than the full permutation of 1024, this greatly reduces the number of strategy combinations that need to be evaluated.

[0051] Step S2.2: Map the values ​​of each parameter to the factors and levels of the selected orthogonal array to generate a preliminary strategy combination.

[0052] The five strategy parameters are mapped to the five columns (factors) of an orthogonal array, and the four values ​​of each parameter are mapped to the four levels of that column (usually represented by 1, 2, 3, 4). Then, according to the factor levels specified in each row of the orthogonal array, specific values ​​for the corresponding parameters are selected to generate preliminary strategy combinations. This results in an experimental design containing 16 strategy combinations.

[0053] Step S2.3: Calculate the distribution uniformity index of the initial strategy combination in the parameter space. If the index is lower than the preset standard, iteratively optimize the mapping relationship until an orthogonal experimental table that meets the requirements of balanced coverage is generated.

[0054] To ensure the quality of the generated experimental combinations, their distribution uniformity can be further evaluated. For example, the frequency of any two parameter combinations appearing in all experimental combinations can be calculated to check if they are close to equilibrium. Distribution uniformity indicators can be based on variance or dispersion. If the indicator is below a preset standard, it indicates that the current parameter-factor mapping may lead to insufficient consideration of some parameter combinations. In this case, the system can automatically adjust the mapping (e.g., swapping the orthogonal table columns corresponding to the two parameters), regenerate the combinations, calculate the indicator, and iteratively optimize the mapping until an orthogonal experimental table that meets the balanced coverage requirement is generated. This process ensures that the selected M combinations are not only few in number but also statistically balanced in covering the influence of each parameter and its interactions. This allows subsequent evaluation results based on this small sample to effectively infer the characteristics of the entire parameter space, laying a solid foundation for scientific optimization.

[0055] After generating the initial orthogonal experimental table, in order to further improve the quality of the experimental combinations and avoid wasting computational resources on strategies with almost certain poor results, this application can also perform a round of optimization and screening on the orthogonal experimental table. That is, the above step S2 can also include the following technical solutions exemplified by steps S2.4 and S2.5:

[0056] Step S2.4: Based on the pre-trained agent model, perform a fast pre-evaluation on the generated M policy combinations to obtain the pre-evaluation results.

[0057] The surrogate model is a lighter and faster simplified model than the main prediction model in step S3, such as a shallow neural network or linear model. It is pre-trained using historical data, and although its prediction accuracy may be slightly lower than the main model, it is sufficient to quickly rank the "good" or "bad" performance of the policies. The system inputs the M policy combinations generated in step S2.3 into this surrogate model to obtain a pre-evaluation result, which is typically the prediction score for each combination.

[0058] Step S2.5: Eliminate combinations whose pre-evaluation results are significantly lower than the historical benchmark, and supplement the remaining combinations to keep the total number M unchanged, forming the final orthogonal experimental table for evaluation in step S3.

[0059] The system establishes a historical performance benchmark (e.g., the average performance of all historical strategies). Combinations whose pre-evaluation results are significantly lower than the historical benchmark are removed; for example, combinations scoring 20% ​​below the benchmark are considered low-potential and removed from the experimental table. After removal, the number of experimental combinations decreases. To maintain the balance and representativeness of the orthogonal experiments, the system needs to supplement the remaining combinations. This supplementation can be done by reselecting underrepresented combinations from the parameter levels corresponding to the removed combinations, or by selecting new replacement rows from the standard orthogonal table. Ultimately, a pre-screened orthogonal experimental table with a constant total number M is formed for evaluation in step S3. This step acts as an "initial screening," filtering out obviously poor strategies before entering the main evaluation process, further improving the efficiency of the overall optimization process.

[0060] Step S3: Input the M strategy combinations into the pre-trained prediction model based on the machine learning model, and output the virtual effect prediction value corresponding to each combination.

[0061] For strategy optimization, relying on real-world A / B testing suffers from drawbacks such as high resource consumption, long evaluation cycles, and high trial-and-error risks. For example, real-world A / B testing, which compares different strategy combinations in a real online environment, consumes real budget and traffic resources for each strategy test. Furthermore, obtaining statistically significant results requires a long deployment period, hindering rapid feedback and resulting in high costs, slow speed, and high risk. Existing solutions based on simple rule-based estimations, using fixed formulas or historical averages for rough estimations, still suffer from low prediction accuracy, inability to characterize complex parameters and nonlinear effects, poor reliability, and susceptibility to decision-making errors. Therefore, this application replaces most real-world experiments by constructing a high-fidelity virtual simulation environment—a prediction model. Specifically, M strategy combinations are input into a pre-trained prediction model based on a machine learning model, which outputs a virtual effect prediction value for each combination. This solution, by introducing a pre-trained machine learning model as a virtual evaluation environment, achieves rapid, parallel, and zero-cost prediction of strategy effects. This model can learn the complex mapping relationship between policy parameters and final effects in historical data, thereby providing effect predictions for any new policy combination with input within milliseconds. This completely transforms policy evaluation from expensive real-world trial and error to efficient digital simulation, making large-scale, high-frequency policy iteration and selection possible. It is the core of reducing optimization costs and improving decision-making speed.

[0062] The pre-trained predictive model based on a machine learning model is a key component of this method, aiming to learn the complex mapping relationship between policy parameter combinations and contextual environment to the final performance metric. The training method of this model itself constitutes an important aspect of this application and can be implemented through the training flow exemplified in steps S3.T1 to S3.T3:

[0063] Step S3.T1: Collect historical delivery data, which includes strategy parameter combinations, corresponding contextual features, and final performance metrics.

[0064] The training data comes from past real-world deployment records. Each data point is a sample and contains:

[0065] Strategy parameter combination: This refers to the specific values ​​of all adjustable parameters used at that time, such as bid, targeting, creative, and time period.

[0066] Contextual characteristics: Uncontrollable external conditions at the time of campaign execution, such as day of the week, whether it is a holiday, the industry, and the competition intensity index. These characteristics are crucial for the model to understand that "the same strategy may have different effects in different environments."

[0067] Final performance metrics: The actual business metrics achieved in this campaign, such as conversion rate, spending amount, ROI, etc., which are aligned with the optimization goals.

[0068] Step S3.T2: Clean and enhance the historical delivery data to construct feature-label pairs for model training.

[0069] Considering that raw historical deployment data cannot usually be used directly for training, preprocessing is required. Specifically, cleaning and augmenting the historical deployment data can be achieved through steps S3.T2a and S3.T2b:

[0070] Step S3.T2a: Clean the historical delivery data, including handling missing values, outliers, and normalization.

[0071] For example, missing contextual features can be filled with the mean or median; extreme outliers in performance metrics (which may be due to data entry errors) can be truncated or removed; and continuous features and labels can be normalized or standardized to accelerate model convergence.

[0072] Step S3.T2b: Augment the cleaned historical deployment data, including generating simulated samples for strategy categories with small sample sizes using synthetic minority class oversampling techniques to balance the distribution of training data.

[0073] In advertising, certain special strategies (e.g., high bids targeting a very small group of people) may have very few samples, leading to inaccurate model predictions for these areas. Algorithms such as the Synthetic Minority Over-sampling Technique (SMOTE) can interpolate new, reasonable virtual samples within the feature space of existing minority class samples, thereby balancing the distribution of training data and improving the model's generalization ability.

[0074] Step S3.T3: Input the cleaned and enhanced historical delivery data into the ensemble learning framework for training to obtain a prediction model that can predict performance indicators based on input policy parameters and contextual features.

[0075] Ensemble learning frameworks (e.g., Random Forest, Gradient Boosting Decision Tree (GBDT), XGBoost, LightGBM, etc.) typically achieve more stable and accurate prediction performance than single models by combining the prediction results of multiple base learners. The system inputs cleaned and augmented historical delivery data into the ensemble learning framework for training. During this process, policy parameters and contextual features are used as input features, and performance metrics are used as prediction labels. After training, a prediction model is obtained that can predict performance metrics based on input policy parameters and contextual features. This model captures the potential nonlinear relationships and interaction effects between policy, environment, and performance.

[0076] When it is necessary to evaluate the orthogonal experimental table generated in step S2, the system inputs M strategy combinations into this pre-trained prediction model based on a machine learning model. The model, in conjunction with the current task context (e.g., the planned launch date, industry), calculates a virtual effect prediction value for each strategy combination in parallel. This process is completed within milliseconds or seconds, outputting the virtual effect prediction value corresponding to each combination, thereby realizing instantaneous, parallel, and zero-cost virtual experiments on a large number of strategies.

[0077] To provide richer decision-making information and enhance the robustness of the screening, the prediction model of this application can also output a measure of prediction uncertainty. That is, step S3 can also include the confidence assessment process exemplified in steps S3.1 and S3.2:

[0078] Step S3.1: While outputting the predicted value of the virtual effect, the prediction model also outputs the confidence interval or uncertainty score of the predicted value.

[0079] For certain models capable of estimating uncertainty—such as Bayesian neural networks, neural networks using Dropout as an approximate Bayesian inference, or the variance predicted by different sub-models in ensemble learning—it's possible to output a confidence interval or uncertainty score along with the predicted virtual effect value. The confidence interval gives the range within which the predicted value may fluctuate, while the uncertainty score quantifies how confident the model is in the prediction.

[0080] Step S3.2: During the screening in step S4, the confidence interval or uncertainty score is used as an auxiliary decision factor to weigh the combination of high predicted value but low confidence and the combination of medium predicted value but high confidence.

[0081] In subsequent screening, the system not only considers the predicted value but also uses confidence intervals or uncertainty scores as auxiliary decision-making factors. For example, one strategy might predict a high ROI but a wide confidence interval (i.e., high uncertainty) due to a lack of similar historical data; another strategy might predict a moderate ROI but a narrow confidence interval (i.e., high certainty). Decision-makers or automated screening logic can weigh the risks of high-predicted-value but low-confidence combinations against moderate-predicted-value but high-confidence combinations, choosing the latter to ensure stable returns or the former to explore high-potential areas. This upgrades decision-making from simply pursuing the "optimal expected value" to balancing "return" and "risk," making it more intelligent.

[0082] After obtaining the virtual effect predictions for M strategy combinations and optional uncertainty measures through the prediction model, the process enters the intelligent screening stage.

[0083] Step S4: Based on the virtual effect prediction values, select the top N strategies with the best prediction effects as high-potential candidate strategies through decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models, or privacy computing techniques, where N is a positive integer less than M.

[0084] The purpose of step S4 is to intelligently select the top N most promising strategies (e.g., N=3 or 5) from the M representative combinations that have undergone virtual evaluation, as candidates for final deployment in the real environment. Here, screening using decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models, or privacy-preserving computation techniques means that these advanced artificial intelligence technologies can be applied to assist or directly execute the screening decision, rather than simply ranking. Several typical screening paradigms are illustrated below. As an advanced screening paradigm, step S4 can be implemented through steps S4.1 to S4.3, detailed below:

[0085] Step S4.1: Based on the virtual effect prediction value and its uncertainty, construct a multi-armed slot machine model, treating each strategy combination as an "arm", whose expected return and uncertainty correspond to the virtual effect prediction value and its uncertainty, respectively.

[0086] In the field of reinforcement learning, the multi-armed slot machine is a classic model used to study how to balance exploiting known high-reward options with exploring potentially higher-reward options in a limited number of trials. Here, the selection problem is formalized into this model: each policy combination is an "arm," and pulling an arm selects that policy. The expected reward of an arm can be estimated using the virtual effect predictions of a predictive model, while the uncertainty of the arm's reward is characterized by the uncertainty score or prediction variance of the model output.

[0087] Step S4.2: In the multi-armed slot machine model, the upper confidence bound algorithm in reinforcement learning is applied to simulate pulling each arm through multiple rounds of iteration, so as to balance the use of high expected reward strategies and the exploration of high uncertainty strategies.

[0088] The upper confidence bound algorithm is a classic algorithm for solving the multi-armed slot machine problem. Its core idea is to calculate an "upper confidence bound" score for each arm: score = current estimated expected reward + exploration coefficient * uncertainty. The system does not directly select the N arms with the highest predicted values. Instead, it applies the upper confidence bound algorithm from reinforcement learning to perform multiple rounds of iterative simulations in a purely virtual environment to "pull" each arm. In each round of simulation, the algorithm selects the arm to "pull" (i.e., virtually evaluate) based on the upper confidence bound scores of all arms. Selecting high-scoring arms tends to take advantage of the current high reward, while the presence of uncertainty also gives opportunities to arms with lower predicted values ​​but high uncertainty (i.e., arms with unknown potential). This process balances the utilization of high-expected-reward strategies with the exploration of high-uncertainty strategies.

[0089] Step S4.3: Based on the cumulative virtual returns of each arm obtained from multiple rounds of iterative simulation, select the top N strategies with the highest cumulative virtual returns as high-potential candidate strategies.

[0090] After multiple rounds (e.g., hundreds of rounds) of simulation, each arm is "pulled" several times, accumulating simulated virtual rewards (i.e., the cumulative predicted reward for each pull). The system sorts the arms based on the cumulative virtual rewards obtained from multiple rounds of iterative simulation, and finally selects the top N strategies with the highest cumulative virtual rewards as high-potential candidate strategies. This method selects strategies that consider not only their point estimates but also their potential space, making it more likely to discover globally optimal or near-optimal strategies.

[0091] As can be seen from the technical solutions exemplified in steps S4.1 to S4.3 above, by constructing the policy selection problem as a multi-armed slot machine model and applying reinforcement learning algorithms (e.g., the upper confidence bound algorithm) to conduct strategic simulation exploration in a virtual environment, not only are policies with high prediction values ​​selected, but the potential of policies with high uncertainty is also systematically evaluated, thereby discovering a globally better solution with a higher probability, enhancing the robustness of optimization and global search capability.

[0092] In another embodiment of this application, step S4 can also be implemented as follows: Initialize a Bayesian optimization surrogate model, taking the policy combination space as input and the virtual effect prediction value as the observation target; utilize a meta-learned acquisition function, which dynamically adjusts its exploration preferences according to the characteristics of the current optimization task in each iteration; perform multiple iterations through the Bayesian optimization surrogate model and the dynamic acquisition function, select the next policy combination to be evaluated, and update the surrogate model until the evaluation limit is reached, selecting the top N policies with the highest evaluation values ​​as high-potential candidate policies. Since the acquisition function itself can be adjusted online according to the task, this means that the exploration strategy in the above scheme is no longer fixed, but can learn from historical exploration experience and be personalized for the characteristics of the current policy space (e.g., smoothness, multimodality), thus realizing the exploration of the exploration strategy.

[0093] When the optimization objective consists of multiple competing goals, such as simultaneously pursuing "high click-through rate" and "low cost per click," the selection logic becomes more complex. Step S4, which selects the top N strategies with the best predictive performance as high-potential candidate strategies, can be achieved through steps S4a to S4c as follows:

[0094] Step S4a: Use the prediction model to output the virtual effect prediction values ​​of each strategy combination under multiple optimization objectives, and form the objective vector.

[0095] For each of the M strategy combinations, the prediction model needs to output a vector, for example, [predicted click-through rate, predicted cost per click]. This vector constitutes the target vector, representing the performance of the strategy across multiple target dimensions.

[0096] Step S4b: Based on the multi-objective vector, apply the Pareto front algorithm to select the non-dominated solution set from the M policy combinations.

[0097] In multi-objective optimization, a Pareto optimal solution (or non-dominated solution) is one that cannot improve any other objective without worsening at least one other objective. Pareto front algorithms (e.g., the core sorting step of the non-dominated sorting genetic algorithm NSGA-II) are applied to compare the objective vectors of all strategies. A set of non-dominated solutions is selected from M strategy combinations. The strategies in this set are mutually non-dominated; that is, a strategy better on one objective must be worse on another. Together, they constitute the optimal trade-off boundary, called the Pareto front.

[0098] Step S4c: From the non-dominated solution set, select the top N strategies as high-potential candidate strategies according to the preset preference weights or reference points.

[0099] Decision-makers can set preference weights for each objective based on business needs (e.g., click-through rate weight 0.7, cost weight 0.3), then calculate the weighted score of each strategy in the non-dominated solution set, and select the top N strategies based on the preset preference weights or reference points. Alternatively, an ideal reference point can be specified, and then the N strategies closest to the reference point (e.g., those with the smallest Euclidean distance) can be selected. This completes the screening under multiple objectives as high-potential candidate strategies.

[0100] The screening schemes in the examples of steps S4a to S4c above can identify all "non-dominated solutions" (i.e. strategies that cannot improve one objective without compromising another) at once by constructing multi-objective vectors and solving the Pareto front. This provides decision-makers with a clear optimal trade-off surface, supporting them to make scientific choices based on actual business preferences, achieving multi-objective collaborative optimization, and meeting complex actual business needs.

[0101] When there are multiple optimization objectives, the above embodiment can further select the top N strategies with the best prediction performance as high-potential candidate strategies by using a multi-task learning architecture with a shared backbone network and multiple specific task output heads as the prediction model to simultaneously output the predicted values ​​of multiple objectives; training a conditional generative adversarial network to generate corresponding policy parameter combinations based on the preference weight vector of the optimization objective; during the selection process, inputting the preset preference weight vector into the conditional generative adversarial network to generate a set of candidate strategies that meet the preference, and then performing a fine evaluation through the multi-task prediction model to select the top N as high-potential candidate strategies. In this scheme, because multi-task learning (for accurate evaluation) and the conditional generative model (for targeted generation) are deeply integrated, the conditional generative model can learn the complex mapping from objective preferences to the policy space, realizing the "inverse design" of multi-objective optimization.

[0102] After selecting high-potential candidate strategies through any of the above methods, in order to further improve the accuracy of the strategies, they can be finely adjusted locally. That is, after step S4, the strategy fine-tuning process exemplified in steps S41 and S42 can also be included:

[0103] Step S41: Centered on the parameter values ​​of each high-potential candidate strategy, perform local sampling in its neighborhood space to generate a subset of fine-tuning strategies.

[0104] Since the policy combinations provided by the orthogonal experimental table are discrete, the candidate policy may not be the optimal point in its local region. The system centers on the parameter values ​​of each high-potential candidate policy, sampling uniformly or randomly within a small neighborhood for continuous parameters; for discrete parameters, it tries neighboring values. This local sampling within the neighborhood space generates dozens of fine-tuned policy subsets for each candidate policy.

[0105] Step S42: Call the prediction model to quickly evaluate the subset of fine-tuning strategies, and replace the original candidate strategies with the best-performing fine-tuning strategy.

[0106] The prediction model is invoked to quickly evaluate the subset of fine-tuning policies. For each subset of candidate policies, the fine-tuning policy with the best prediction performance is identified. Then, the original candidate policy is replaced with the best-performing fine-tuning policy. This process effectively improves the overall quality of the final output policy set, making it closer to the true local optimum.

[0107] The policy fine-tuning process exemplified in steps S41 and S42 above achieves policy refinement by performing local sampling and rapid evaluation within the neighborhood of candidate policies. This can further improve a good policy into an even better one, significantly enhancing the accuracy and quality of the final output policy and overcoming the problem of missing optimal solutions that may be caused by the discretization of experimental design.

[0108] As another embodiment of this application, the policy fine-tuning process can also be as follows: the prediction model is regarded as a differentiable function, and the policy parameters are regarded as optimizable variables; taking each high-potential candidate policy as the initial point, the gradient of the prediction effect with respect to the policy parameters is calculated by automatic differentiation technology; multi-step projection gradient ascent is performed along the gradient direction to ensure that the parameters are updated within the feasible region to obtain the fine-tuned policy; and the original candidate policy is replaced with the fine-tuned policy.

[0109] Step S5: Create high-potential candidate strategies as actual execution plans and monitor their deployment.

[0110] After the above steps, N high-quality strategies were obtained, which underwent virtual evaluation, intelligent screening, and possible fine-tuning. Creating high-potential candidate strategies as actual execution plans involves creating corresponding advertising plans or tasks within the advertising platform or resource management system, according to the specific parameter configurations of each strategy, and allocating a real budget to launch the campaign.

[0111] Monitoring the campaign is crucial to ensuring that the theoretical potential is realized in reality. Campaign monitoring includes the following steps S5.1 to S5.3, which illustrate a bid calibration process based on real-time feedback:

[0112] Step S5.1: After the actual deployment is started, collect real-time effect data at the set frequency.

[0113] The system collects real-time performance data, such as spending amount, clicks received, and conversions, at a frequency of minutes or hours.

[0114] Step S5.2: Compare the real-time performance data with the corresponding predicted values ​​output by the prediction model, and calculate the prediction deviation.

[0115] The cumulative real-time results (e.g., current ROI) are compared with the corresponding predicted values ​​(e.g., predicted ROI) given by the prediction model before deployment, and the prediction bias is calculated. Bias = (actual value - predicted value) / predicted value.

[0116] Step S5.3: Based on the prediction deviation, dynamically adjust the value using a proportional-integral-derivative control algorithm to make the cumulative real-time effect approach the predicted trajectory.

[0117] PID control is a classic anti-locking regulation method in industrial control. This system applies this algorithm to bid calibration: the proportional term (P) reacts quickly to the current deviation; the integral term (I) accumulates historical deviations to eliminate steady-state error; and the derivative term (D) responds in advance based on the rate of change of deviation. Based on the predicted deviation, the bid is dynamically adjusted using the proportional-integral-derivative control algorithm. For example, if the actual ROI is consistently lower than the prediction, the bid is finely adjusted to decrease, and vice versa. The goal is to make the cumulative real-time effect approach the predicted trajectory, thereby ensuring that the virtual optimization results can be realized in reality.

[0118] In the real-time monitoring and calibration process of step S5, the accuracy of the prediction model is the cornerstone of the effective operation of the entire system. However, the online environment is complex and ever-changing, and the predictive ability of the model may decline over time. Therefore, this application designs an online evolution mechanism for the prediction model, namely, the online update process of the prediction model is achieved through the following steps S5.4 and S5.5:

[0119] Step S5.4: When the prediction deviation continues to exceed the preset threshold, or when the deployment environment is detected to have experienced significant drift, the model update process is triggered.

[0120] The system continuously monitors the prediction bias calculated in step S5.2. When the prediction bias continues to exceed a preset threshold (e.g., the absolute value of the bias is greater than 10% for multiple consecutive time windows), or when statistical detection methods (e.g., KS test, PSI index) reveal a significant drift between the distribution of newly collected real-time data and the distribution of historical data during model training, the system automatically triggers the model update process.

[0121] Step S5.5: Mix the newly collected real deployment data with the historical training data, and update the parameters of the prediction model by online learning or periodic batch retraining.

[0122] One direct approach is to mix newly collected real-world deployment data with historical training data to create a new training dataset. Then, a new predictive model is trained from scratch using periodic batch retraining to replace the old model. This method is simple but computationally expensive, and the new model may completely forget the knowledge from the old data—a catastrophic forgetting. A better implementation is online learning. Online learning allows the model to perform incremental updates in small batches after receiving new data samples. For example, for neural network-based predictive models, gradients can be calculated using new data, and the model parameters can be fine-tuned in one step. This approach updates the predictive model's parameters, enabling it to quickly adapt to new trends. To prevent catastrophic forgetting, regularization constraints on changes in important parameters (e.g., Elastic Weight Consolidation) can be added to the loss function of online learning. This is a higher-order implementation that ensures the model retains core knowledge learned in the past while adapting to new knowledge, achieving stable and continuous evolution.

[0123] The online updates of the prediction model exemplified in steps S5.4 and S5.5 above effectively solve the problem of how to enable the prediction model to continuously adapt to the dynamically changing real environment and maintain long-term prediction accuracy. By monitoring prediction bias or environmental drift and triggering the online update process, the prediction model can continuously learn and evolve using the latest real feedback data. This ensures the stability and effectiveness of the entire optimization system in long-term operation and avoids the degradation of optimization performance caused by model aging.

[0124] The technical solution of this application not only focuses on a single optimization task, but also aims to build a knowledge system that can accumulate and reuse optimization experience, thereby making the system increasingly intelligent. That is, furthermore, Figure 1The example method may also include the process of building and reusing the strategy knowledge base in steps S6 and S7:

[0125] Step S6: After each complete optimization cycle from Step S1 to Step S5, the optimization objective, the set of strategy parameters used, the generated orthogonal experimental table, the high-potential candidate strategies and their actual deployment effects are associated and stored in the strategy knowledge base.

[0126] After each optimization task is completed, regardless of success or failure, the entire process and result constitute a valuable experience. The system records in a structured manner: the optimization objective (e.g., "maximizing day-one ROI"), the set of strategy parameters used (e.g., [bid, age group, creative type]), the generated orthogonal experimental table (M specific combinations), the high-potential candidate strategies (the final N strategies and their parameters), and the actual delivery effect observed in a real environment (e.g., actual ROI = 1.5). This information is stored in a strategy knowledge base, forming a series of queryable historical optimization cases.

[0127] Step S7: When a new optimization task is received, retrieve historical similar tasks and effective strategies from the strategy knowledge base, and use the retrieval results as prior knowledge to guide the initial setting of parameters in step S1 or the key areas of orthogonal experimental design in step S2.

[0128] When a new optimization task arrives, the system does not start from scratch. It first retrieves historical similar tasks and effective strategies from the policy knowledge base. The core of the retrieval lies in how to define similarity; that is, as an embodiment of this application, retrieving historical similar tasks from the policy knowledge base can be achieved through steps S7.1 to S7.3:

[0129] Step S7.1: Transform the optimization objective and context features of the new task into a feature vector of the new task.

[0130] The textual optimization objective of the new task (such as "increase the paid conversion rate") is transformed into a vector through a word embedding model, and then concatenated with numerical contextual features (such as industry code and budget size) to form a new task feature vector.

[0131] Step S7.2: Use a deep learning-based semantic matching model to calculate the similarity between the feature vector of the new task and the feature vector of the historical tasks in the knowledge base.

[0132] Each historical task record in the strategy knowledge base is also transformed into a feature vector in the same way. The system uses a deep learning-based semantic matching model (e.g., a sentence matching model based on Siamese neural networks or BERT structures) to calculate the similarity between the feature vector of the new task and the feature vectors of historical tasks in the knowledge base. The model is able to understand the similarity of the semantics of the optimization target (e.g., "increase the paid conversion rate" and "improve the conversion rate" are similar) and perform a comprehensive matching based on the context.

[0133] Step S7.3: Output the top K historical tasks with the highest similarity and their associated policy information as prior knowledge.

[0134] The system outputs the top K most similar historical tasks and their associated strategy information as prior knowledge. This prior knowledge can be used for: 1) setting initial values ​​for parameters in step S1: For example, if "user interest tags" are proven to be core parameters in similar historical tasks, they can be prioritized for inclusion in the parameter set in new tasks. 2) designing key areas for orthogonal experiments in step S2: For example, if strategies near "bid = 2.0 yuan" generally perform well in similar historical tasks, denser value points can be set near this parameter level when generating the orthogonal array to achieve biased exploration and accelerate the optimization process.

[0135] In today's world, where data privacy protection is increasingly important, training models by pooling data from various sources may face compliance challenges. To address this, this application provides a distributed training scheme based on federated learning. Specifically, the federated learning-based distributed training architecture exemplified in steps S3.T3a and S3.T3b, where historical delivery data after cleaning and augmentation is input into the ensemble learning framework for training, can be implemented using the following steps:

[0136] Step S3.T3a: Train a local prediction sub-model locally on the local premises of multiple data holders, using their respective historical delivery data.

[0137] Suppose multiple advertisers (data holders) want to collaboratively improve the predictive capabilities of a model, but cannot share raw data for privacy reasons. In a federated learning architecture, a coordinating server distributes the initial global model parameters to each participant. Each participant, locally on multiple data holders' datasets, uses their respective historical campaign data to train the model for several rounds, creating a local predictive sub-model. The data never leaves the local machine.

[0138] Step S3.T3b: The coordination server uses a secure aggregation algorithm to periodically aggregate the parameter updates of each local prediction sub-model without exposing the original local data, in order to form a global prediction model.

[0139] After a certain number of training rounds, each participant sends encrypted parameter updates (usually gradients or parameter differences) of its local sub-model compared to the previous round's global model to the coordination server. The server then periodically aggregates the parameter updates of each local prediction sub-model using a secure aggregation algorithm (e.g., SecureAggregation) without decrypting the updates from individual participants. The average update after aggregation is used to update the global model. This process is repeated until a global prediction model is formed. This method ensures that the original local data is not exposed during the aggregation process. While protecting data privacy, it leverages multi-party data to jointly train a more powerful shared prediction model, achieving collaborative intelligence where "data is available but not visible."

[0140] The federated learning-based distributed training architecture exemplified in steps S3.T3a and S3.T3b above is based on the assumption of identical data distribution. However, one of the core challenges in federated learning is data heterogeneity. Therefore, in order to generate customized models for each participant, personalized modules or prototype networks can be introduced to explicitly acknowledge and utilize the differences in data distribution. That is, as another embodiment of this application, the ensemble learning framework of the above embodiment can adopt a personalized federated learning architecture, specifically including: the server maintains a globally shared model; each data holder locally not only trains the model but also learns a lightweight personalized adapter module, or uses a prototype network to learn a prototype representation for local data within a shared feature space; during aggregation, the parameters of the globally shared model are mainly aggregated, while the personalized adapter or prototype remains local; during inference, the global model and the local personalized module or prototype are combined to generate a personalized prediction model suitable for the data distribution of that participant. This approach proposes a hybrid model architecture that combines global sharing with local personalization. Instead of pursuing a consensus model, it seeks a flexible modeling framework that can accommodate both commonalities and individual characteristics. The design is more creative, and the generated personalized models theoretically have better performance on their respective local data. At the same time, they still benefit from data expansion through the global sharing component, achieving a better balance between personalized performance while protecting privacy.

[0141] From the above appendix Figure 1As can be seen from the example of the automated resource allocation strategy optimization method, on the one hand, by applying orthogonal experimental design to generate strategy combinations, the technical solution of this application can statistically and evenly cover the influence of each parameter and its interaction with far fewer permutations and combinations. This means that only a small number of carefully designed representative strategy combinations need to be evaluated to effectively infer the influence trend of each factor in the entire strategy space, thereby greatly reducing the number of strategies that need to be examined, solving the combination explosion problem caused by massive combinations, and laying the foundation for rapid optimization. On the other hand, by introducing a prediction model based on machine learning models to predict the virtual effects of the generated strategy combinations, the technical solution of this application can estimate the potential effects of a large number of strategy combinations in advance and in parallel before the actual deployment. This approach replaces the traditional method of obtaining feedback through real A / B testing or online canary releases, avoiding the waste of real traffic and budget on potentially ineffective strategies. It achieves low-cost, near-zero-risk strategy pre-screening. Thirdly, the technical solution combines efficient spatial sampling (i.e., orthogonal experimental design), intelligent virtual evaluation (i.e., machine learning prediction), and automated strategy selection and execution, forming a coherent automated process. This process automatically generates candidate strategies based on the optimization objective, intelligently predicts their effects, selects the optimal subset, and executes them. Human intervention is minimal throughout the process. This not only reduces reliance on professional experience but also enables strategy optimization to quickly and automatically adapt to constantly changing environments and objectives, improving the overall intelligence and responsiveness of the optimization process. In summary, the technical solution of this application significantly reduces the number of strategy combinations to be evaluated through orthogonal experimental design and utilizes machine learning models for virtual effect prediction, thereby rapidly selecting high-potential strategies at extremely low cost.

[0142] Please see the appendix Figure 2 This application provides an automated resource allocation strategy optimization device, which includes a determination module 201, a generation module 202, a prediction module 203, a filtering module 204, and a monitoring module 205, as detailed below:

[0143] The determination module 201 is used to receive the optimization target, determine the set of strategy parameters associated with the optimization target, and define the value of each parameter in the set of strategy parameters;

[0144] The generation module 202 is used to generate an orthogonal experimental table containing M policy combinations based on each parameter in the policy parameter set and its defined values, using orthogonal experimental design. Here, M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically cover the effects of each parameter and its interaction in a balanced manner.

[0145] The prediction module 203 is used to input the combination of M strategies into a pre-trained prediction model based on a machine learning model, and output the virtual effect prediction value corresponding to each combination.

[0146] The filtering module 204 is used to select the top N strategies with the best prediction results as high-potential candidate strategies based on the virtual effect prediction values, through decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models or privacy computing techniques, where N is a positive integer less than M.

[0147] The monitoring module 205 is used to create high-potential candidate strategies as actual execution plans and monitor their deployment.

[0148] From the above appendix Figure 2 As can be seen from the example of the automated resource allocation strategy optimization device, on the one hand, by applying orthogonal experimental design to generate strategy combinations, the technical solution of this application can statistically and evenly cover the influence of each parameter and its interaction with far fewer permutations and combinations. This means that only a small number of carefully designed representative strategy combinations need to be evaluated to effectively infer the influence trend of each factor in the entire strategy space, thereby greatly reducing the number of strategies that need to be examined, solving the combination explosion problem caused by massive combinations, and laying the foundation for rapid optimization. On the other hand, by introducing a prediction model based on machine learning models to predict the virtual effects of the generated strategy combinations, the technical solution of this application can estimate the potential effects of a large number of strategy combinations in advance and in parallel before the actual deployment. This approach replaces the traditional method of obtaining feedback through real A / B testing or online canary releases, avoiding the waste of real traffic and budget on potentially ineffective strategies. It achieves low-cost, near-zero-risk strategy pre-screening. Thirdly, the technical solution combines efficient spatial sampling (i.e., orthogonal experimental design), intelligent virtual evaluation (i.e., machine learning prediction), and automated strategy selection and execution, forming a coherent automated process. This process automatically generates candidate strategies based on the optimization objective, intelligently predicts their effects, selects the optimal subset, and executes them. Human intervention is minimal throughout the process. This not only reduces reliance on professional experience but also enables strategy optimization to quickly and automatically adapt to constantly changing environments and objectives, improving the overall intelligence and responsiveness of the optimization process. In summary, the technical solution of this application significantly reduces the number of strategy combinations to be evaluated through orthogonal experimental design and utilizes machine learning models for virtual effect prediction, thereby rapidly selecting high-potential strategies at extremely low cost.

[0149] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. For example... Figure 3As shown, the electronic device 3 in this embodiment mainly includes: a processor 30, a memory 31, and a computer program 32 stored in the memory 31 and executable on the processor 30, such as a program for an automated resource allocation strategy optimization method. When the processor 30 executes the computer program 32, it implements the steps in the above-described automated resource allocation strategy optimization method embodiment, for example... Figure 1 Steps S1 to S5 are shown. Alternatively, when processor 30 executes computer program 32, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 2 The functions of the determination module 201, generation module 202, prediction module 203, filtering module 204, and monitoring module 205 are shown.

[0150] For example, the computer program 32 of the automated resource allocation strategy optimization method mainly includes: receiving an optimization objective, determining a set of strategy parameters associated with the optimization objective, and defining the value of each parameter in the set of strategy parameters; based on each parameter in the set of strategy parameters and its defined value, applying orthogonal experimental design to generate an orthogonal experimental table containing M strategy combinations, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically balance the influence of each parameter and its interaction; inputting the M strategy combinations into a pre-trained prediction model based on a machine learning model, and outputting the virtual effect prediction value corresponding to each combination; based on the virtual effect prediction value, using decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models, or privacy computing techniques, selecting the top N strategies with the best prediction effect as high-potential candidate strategies, where N is a positive integer less than M; creating real execution plans for the high-potential candidate strategies and monitoring their deployment. The computer program 32 can be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete this application. One or more modules / units can be a series of computer program instruction segments capable of performing specific functions. These instruction segments describe the execution process of the computer program 32 in the electronic device 3. For example, the computer program 32 can be divided into the functions of a determination module 201, a generation module 202, a prediction module 203, a screening module 204, and a monitoring module 205 (a module in a virtual device). The specific functions of each module are as follows: The determination module 201 is used to receive the optimization target, determine the set of policy parameters associated with the optimization target, and define the value of each parameter in the set of policy parameters; The generation module 202 is used to generate an orthogonal experimental table containing M policy combinations based on each parameter in the set of policy parameters and its defined value, using orthogonal experimental design, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M groups The combination of strategies can statistically balance the impact of each parameter and its interaction; the prediction module 203 is used to input the M strategy combinations into a pre-trained prediction model based on a machine learning model, and output the virtual effect prediction value corresponding to each combination; the screening module 204 is used to select the top N strategies with the best prediction effect as high-potential candidate strategies based on the virtual effect prediction values, through decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models or privacy computing techniques, where N is a positive integer less than M; the monitoring module 205 is used to create real execution plans for the high-potential candidate strategies and monitor their deployment.

[0151] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0152] The processor 30 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0153] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or RAM. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store computer programs and other programs and data required by the electronic device. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed. That is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above-described device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0155] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0156] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0157] In the embodiments provided in this application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0158] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0159] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0160] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program for the automated resource allocation strategy optimization method can be stored in a storage medium. When executed by a processor, this computer program can implement the steps of the various method embodiments described above, namely, receiving an optimization target, determining a set of strategy parameters associated with the optimization target, and defining a value for each parameter in the strategy parameter set; based on each parameter in the strategy parameter set and its defined value, applying orthogonal experimental design to generate an orthogonal experimental table containing M strategy combinations, wherein... M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations statistically balance the influence of each parameter and its interactions. The M policy combinations are input into a pre-trained prediction model based on a machine learning model, which outputs a virtual effect prediction value for each combination. Based on the virtual effect prediction values, the top N policies with the best prediction performance are selected as high-potential candidate policies using decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models, or privacy computing techniques, where N is a positive integer less than M. These high-potential candidate policies are then used to create real execution plans and monitored for deployment. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. The storage medium can include any entity or device capable of carrying computer program code, recording media, USB flash drives, external hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of the storage medium may be appropriately added to or subtracted from the contents according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium may not include electrical carrier signals and telecommunication signals.

[0161] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application. The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the protection scope of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An automated resource allocation strategy optimization method, characterized in that, The method includes: Step S1: Receive the optimization target, determine the set of strategy parameters associated with the optimization target, and define the value of each parameter in the set of strategy parameters; Step S2: Based on each parameter and its defined value, orthogonal experimental design is applied to generate an orthogonal experimental table containing M strategy combinations, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically cover the effects of each parameter and its interaction in a balanced manner. Step S3: Input the M strategy combinations into the pre-trained prediction model based on the machine learning model, and output the virtual effect prediction value corresponding to each combination; Step S4: Based on the virtual effect prediction value, select the top N strategies with the best prediction effect as high-potential candidate strategies through decision tree, ensemble learning model, support vector machine, deep learning model, reinforcement learning model or privacy computing technology, where N is a positive integer less than M. Step S5: Create the high-potential candidate strategies as actual execution plans and monitor their deployment.

2. The automated resource allocation strategy optimization method according to claim 1, characterized in that, The determination of the set of strategy parameters associated with the optimization objective includes: Step S1.1: Extract features from historical deployment data and use a feature importance evaluation algorithm to identify the K core parameters that rank highest in terms of their impact on the optimization objective; Step S1.2: Select the parameters that can be adjusted within a preset time period from the K core parameters to form the strategy parameter set.

3. The automated resource allocation strategy optimization method according to claim 2, characterized in that, The step of selecting adjustable parameters from the K core parameters within a preset time period includes: Step S1.2a: Real-time monitoring of the current market environment characteristics; Step S1.2b: Based on the market environment state characteristics, match the corresponding parameter adjustability rules from the pre-configured rule base; Step S1.2c: Based on the matched rules, dynamically select the currently adjustable parameters from the K core parameters.

4. The automated resource allocation strategy optimization method according to claim 1, characterized in that, Based on the virtual effect prediction values, the top N strategies with the best prediction performance are selected as high-potential candidate strategies using decision trees, ensemble learning models, support vector machines, deep learning models, reinforcement learning models, or privacy computing techniques, including: Step S4.1: Based on the predicted virtual effect value and its uncertainty, construct a multi-armed slot machine model, treating each strategy combination as an "arm", whose expected return and uncertainty correspond to the predicted virtual effect value and its uncertainty, respectively; Step S4.2: In the multi-armed slot machine model, the upper confidence bound algorithm in reinforcement learning is applied to simulate pulling each arm through multiple rounds of iteration, so as to balance the use of high expected reward strategies and the exploration of high uncertainty strategies. Step S4.3: Based on the cumulative virtual returns of each arm obtained from the multi-round iterative simulation, select the top N strategies with the highest cumulative virtual returns as the high-potential candidate strategies.

5. The automated resource allocation strategy optimization method according to claim 1, characterized in that, When there are multiple optimization objectives, the process of selecting the top N strategies with the best prediction performance as high-potential candidate strategies includes: Step S4a: Use the prediction model to output the predicted virtual effects of each strategy combination under multiple optimization objectives, forming an objective vector; Step S4b: Based on the multi-objective vector, apply the Pareto front algorithm to select the non-dominated solution set from the M policy combinations; Step S4c: From the non-dominated solution set, select the top N strategies as the high-potential candidate strategies according to the preset preference weights or reference points.

6. The automated resource allocation strategy optimization method according to claim 1, 4, or 5, characterized in that, Following step S4, the following strategy fine-tuning process is also included: Step S41: Centered on the parameter values ​​of each high-potential candidate strategy, perform local sampling in its neighborhood space to generate a subset of fine-tuning strategies; Step S42: Call the prediction model to quickly evaluate the subset of fine-tuning strategies, and replace the original candidate strategies with the optimal fine-tuning strategy.

7. The automated resource allocation strategy optimization method according to claim 1, characterized in that, The aforementioned campaign monitoring includes the following bid calibration process based on real-time feedback: Step S5.1: After the actual deployment is started, collect real-time effect data at the set frequency; Step S5.2: Compare the real-time effect data with the corresponding predicted value output by the prediction model, and calculate the prediction deviation; Step S5.3: Based on the prediction deviation, dynamically adjust the value using a proportional-integral-derivative control algorithm to make the cumulative real-time effect approach the predicted trajectory.

8. An automated resource allocation strategy optimization device, characterized in that, The device includes: The determination module is used to receive the optimization target, determine the set of strategy parameters associated with the optimization target, and define the value of each parameter in the set of strategy parameters; The generation module is used to generate an orthogonal experimental table containing M strategy combinations based on each parameter and its defined values, using orthogonal experimental design, where M is a positive integer much smaller than the total number of permutations of parameter values, and the M combinations can statistically balance the influence of each parameter and its interaction. The prediction module is used to input the M strategy combinations into a pre-trained prediction model based on a machine learning model, and output the virtual effect prediction value corresponding to each combination. The filtering module is used to select the top N strategies with the best prediction results as high-potential candidate strategies based on the virtual effect prediction value, through decision tree, ensemble learning model, support vector machine, deep learning model, reinforcement learning model or privacy computing technology, where N is a positive integer less than M. The monitoring module is used to create the high-potential candidate strategies as actual execution plans and monitor their deployment.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images