Intelligent decision engine adaptive scheduling method and system based on multi-objective optimization

Abstract

The application discloses a multi-target optimization-based intelligent decision engine adaptive scheduling method and system, relates to the technical field of artificial intelligence and intelligent scheduling, and discloses a multi-target optimization-based intelligent decision engine adaptive scheduling method and system, which realizes adaptive decision scheduling by receiving a decision request, constructing a dynamic capability portrait, inputting a pre-training model to generate a candidate scheme and selecting an optimal scheduling, can dynamically match an optimal decision branch according to task characteristics and system states, adapts to a dynamic business environment, improves resource utilization efficiency, and ensures stable decision quality.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and intelligent scheduling technology, and in particular to an adaptive scheduling method and system based on multi-objective optimization intelligent decision engine. Background Technology

[0002] In the actual deployment of intelligent decision engines, traditional architectures generally rely on a single technology stack to handle all decision-making tasks, such as using only a rule engine, statistical model, or large language model to run independently. While this design may maintain basic functionality when business scenarios are simple, its structural flaws become increasingly apparent as task complexity and diversity continue to grow. The system cannot dynamically identify and match the optimal decision branch based on the specific characteristics of the task (such as the semantic depth and logical complexity of the input data) and its real-time running status (such as system load level and resource consumption). For example, for low-complexity routine queries, the system may incorrectly allocate resources to a high-cost large language model, resulting in redundant consumption of computing resources; while for high-complexity deep inference tasks, it may only be handled by the rule engine, leading to inaccurate decision results or omission of key information. A deeper problem is that existing technologies lack continuous tracking and quantitative modeling of the historical performance of decision branches, failing to effectively capture the confidence fluctuation patterns of branches under different task types, the trend of processing cost changes, and the response latency characteristics under the influence of system load. This makes scheduling decisions based only on static configurations or simplified empirical rules, making it difficult to adapt to dynamically changing business environments. Meanwhile, when a technical branch experiences performance degradation or failure due to abnormal conditions, the system lacks real-time awareness and a collaborative scheduling mechanism, failing to trigger adaptive degradation or switch to a backup branch. This easily leads to service interruptions or a decline in decision quality. These shortcomings collectively result in low resource utilization efficiency, continuously rising operating costs, and unstable overall decision quality, restricting the reliability and scalability of the intelligent decision engine in complex business scenarios.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of this application is to provide an adaptive scheduling method and system for an intelligent decision engine based on multi-objective optimization, which aims to adapt to dynamic business environments, improve resource utilization efficiency, and ensure stable decision quality.

[0005] To achieve the above objectives, this application proposes an adaptive scheduling method for an intelligent decision engine based on multi-objective optimization, the method comprising:

[0006] Receive decision requests from business systems, and obtain task characteristics and current system status from the decision requests;

[0007] Based on historical decision-making tasks, construct dynamic capability profiles for decision-making branches;

[0008] The task characteristics, the current system state, and the dynamic capability profile are input into a pre-trained Pareto front prediction model based on deep learning.

[0009] The task features, the current system state, and the dynamic capability profile are processed by the deep learning-based pre-trained Pareto front prediction model to generate a set of candidate scheduling schemes that have been calibrated by capability.

[0010] Based on current business preferences, a scheduling scheme is selected from the set of candidate scheduling schemes;

[0011] According to the scheduling scheme, the decision request is distributed to the corresponding decision branch for processing.

[0012] In one embodiment, the step of constructing a dynamic capability profile for a decision branch based on historical decision-making tasks includes:

[0013] Collect the actual processing results of the decision branches to obtain the actual result set;

[0014] Based on the actual result set, the confidence change pattern of the decision branch under multiple task complexities is analyzed to generate relation curve data, and the cost change pattern of the decision branch in processing multiple feature tasks is analyzed to generate correlation model data. In addition, the latency change pattern of the decision branch under multiple system loads is analyzed to generate response model data.

[0015] Generate a branch robustness score based on historical anomaly handling records of decision branches;

[0016] The relationship curve data, the association model data, the response model data, and the branch robustness score are integrated and processed to generate a dynamic capability profile.

[0017] In one embodiment, the step of inputting the task features, the current system state, and the dynamic capability profile into a pre-trained Pareto front prediction model based on deep learning includes:

[0018] The task features are subjected to multi-scale feature encoding processing to generate task feature codes; the current system state is subjected to temporal state encoding processing to generate system state codes; and the dynamic capability profile is subjected to structured profile encoding processing to generate capability profile codes.

[0019] Based on the cross-attention mechanism, the interaction weights among the task feature encoding, the system state encoding, and the capability profile encoding are calculated.

[0020] Based on the interaction weights, the task feature encoding, the system state encoding, and the capability profile encoding are weighted and fused to generate a fused encoding.

[0021] The fused code is input into a multi-target decoder, and the fused code is processed through multi-task parallel learning to generate predictions of candidate scheduling schemes.

[0022] In one embodiment, the step of calculating the interaction weights among the task feature encoding, the system state encoding, and the capability profile encoding based on the cross-attention mechanism includes:

[0023] Based on the task feature encoding, the system state encoding, and the capability profile encoding, attention weight data between task features and system state, attention weight data between task features and capability profile, and attention weight data between system state and capability profile are calculated simultaneously through a cross-attention mechanism.

[0024] A three-dimensional interactive attention matrix is ​​constructed based on the attention weight data between task features and system state, the attention weight data between task features and capability profile, and the attention weight data between system state and capability profile.

[0025] The three-dimensional interactive attention matrix is ​​subjected to attention normalization processing to generate an interactive weight distribution.

[0026] In one embodiment, the step of processing the task features, the current system state, and the dynamic capability profile using the deep learning-based pre-trained Pareto front prediction model to generate a set of capability-calibrated candidate scheduling schemes includes:

[0027] Based on the relationship curve data, association model data and response model data in the dynamic capability profile, benchmark confidence data, benchmark cost data and benchmark latency data are generated.

[0028] The benchmark confidence data, the benchmark cost data, and the benchmark delay data are used as prior constraints.

[0029] The task feature encoding, the system state encoding, the capability profile encoding, and the prior constraints are jointly constrained and encoded to generate constraint-aware encoding.

[0030] Based on the constraint-aware coding, prediction results for multiple candidate scheduling schemes are generated;

[0031] The prediction results of the multiple candidate scheduling schemes are sorted by non-dominated ordering to select candidate scheduling schemes on the Pareto front and generate a set of candidate scheduling schemes.

[0032] In one embodiment, the step of performing joint constraint encoding processing on the task feature encoding, the system state encoding, the capability profile encoding, and the prior constraints to generate constraint-aware encoding includes:

[0033] The baseline confidence data, the baseline cost data, and the baseline delay data are converted into a priori constraint vector;

[0034] The task feature encoding, the system state encoding, and the capability profile encoding are used as observation feature vectors;

[0035] Based on the gated attention mechanism, the adjustment weight of the prior constraint vector on the observed feature vector is calculated;

[0036] Based on the adjustment weights, the observed feature vector is subjected to feature adjustment processing to generate a constraint-aware feature representation, which is then used as a constraint-aware encoding.

[0037] In one embodiment, the step of generating prediction results for multiple candidate scheduling schemes based on the constraint-aware coding includes:

[0038] Based on the constraint-aware coding, branch selection feature extraction processing is performed to generate the branch selection probability distribution of candidate scheduling schemes;

[0039] Based on the constraint-aware coding, confidence prediction feature extraction is performed to generate the predicted confidence of candidate scheduling schemes;

[0040] Based on the constraint-aware coding, cost evaluation feature extraction processing is performed to generate the predicted cost of candidate scheduling schemes;

[0041] Based on the constraint-aware coding, delay estimation feature extraction is performed to generate candidate scheduling schemes for delay prediction.

[0042] The branch selection probability distribution, the prediction confidence, the prediction cost, and the prediction delay are integrated and processed to generate the prediction results of the candidate scheduling scheme.

[0043] In one embodiment, the step of selecting a scheduling scheme from the set of candidate scheduling schemes based on current business preferences includes:

[0044] Business preference weights are generated through dynamic weight adjustment processing using reinforcement learning.

[0045] Obtain the constraint data for the current business scenario;

[0046] Based on the business preference weights and the constraint data, each candidate scheduling scheme in the candidate scheduling scheme set is calculated and processed to generate a constraint satisfaction score.

[0047] Based on the aforementioned business preference weights, candidate scheduling schemes that meet the constraints are calculated and processed to generate a comprehensive utility score.

[0048] The constraint satisfaction score and the comprehensive utility score are weighted and combined, and the candidate scheduling scheme with the highest score is selected based on the weighted combination result to generate a scheduling scheme.

[0049] In one embodiment, the step of generating business preference weights through reinforcement learning dynamic weight adjustment processing includes:

[0050] Based on system operating parameters and the effects of historical scheduling decisions, system state characterization data is generated.

[0051] Based on the system state characterization data, dynamic weight configuration strategy data is calculated and generated;

[0052] Based on the current task characteristics and the current system state, weight allocation data is generated according to the dynamic weight configuration strategy data.

[0053] Based on the decision execution feedback, the weight allocation data is adjusted to generate business preference weights.

[0054] Furthermore, to achieve the above objectives, this application also proposes an adaptive scheduling system for an intelligent decision engine based on multi-objective optimization. The adaptive scheduling system for an intelligent decision engine based on multi-objective optimization includes: a memory, a processor, and an adaptive scheduler for an intelligent decision engine based on multi-objective optimization stored in the memory and executable on the processor. The adaptive scheduler for an intelligent decision engine based on multi-objective optimization is configured to implement the steps of the adaptive scheduling method for an intelligent decision engine based on multi-objective optimization.

[0055] The adaptive scheduling method and system for intelligent decision engines based on multi-objective optimization proposed in this application achieve adaptive decision scheduling by receiving decision requests, constructing dynamic capability profiles, generating candidate solutions by inputting pre-trained models, and selecting the optimal scheduling. It can dynamically match the optimal decision branch according to task characteristics and system status, adapt to dynamic business environments, improve resource utilization efficiency, and ensure stable decision quality. Attached Figure Description

[0056] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0057] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating an embodiment of the adaptive scheduling method for an intelligent decision engine based on multi-objective optimization in this application.

[0059] Figure 2 For this application Figure 1 A detailed flowchart of step S200;

[0060] Figure 3 For this application Figure 1 Detailed flowchart of step S300;

[0061] Figure 4 For this application Figure 1 Detailed flowchart of step S400;

[0062] Figure 5 For this application Figure 1 Detailed flowchart of step S500;

[0063] Figure 6 This is a schematic diagram of the structure of an embodiment of the intelligent decision engine adaptive scheduling system based on multi-objective optimization according to this application.

[0064] Explanation of icon numbers:

[0065] 10. Memory; 20. Processor.

[0066] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] The technical solutions 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. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0068] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0069] In existing technologies, intelligent decision engines often rely on a single technology stack when handling complex business scenarios. This prevents the system from making optimal choices and coordinating scheduling among different technology branches based on task characteristics. As a result, the system suffers from low utilization of computing resources, high decision-making costs, rigid decision-making processes, limited overall quality, and a lack of adaptive degradation and fault tolerance capabilities.

[0070] Based on this, embodiments of this application provide an adaptive scheduling method for an intelligent decision engine based on multi-objective optimization, referring to... Figure 1 The adaptive scheduling method for intelligent decision engines based on multi-objective optimization includes steps S100 to S600, wherein:

[0071] Step S100: Receive a decision request from the business system, and obtain task characteristics and current system status from the decision request;

[0072] Step S200: Based on historical decision-making tasks, construct dynamic capability profiles for decision-making branches;

[0073] Step S300: Input the task features, the current system state, and the dynamic capability profile into a pre-trained Pareto front prediction model based on deep learning;

[0074] Step S400: The task features, the current system state, and the dynamic capability profile are processed by the deep learning-based pre-trained Pareto front prediction model to generate a set of candidate scheduling schemes that have been calibrated by capability.

[0075] Step S500: Based on the current business preferences, select a scheduling scheme from the set of candidate scheduling schemes;

[0076] Step S600: According to the scheduling scheme, the decision request is distributed to the corresponding decision branch for processing.

[0077] In this embodiment, a decision request refers to a processing instruction issued by the business system to the intelligent decision engine, which contains specific task information requiring a decision. Task characteristics describe the attributes of the task corresponding to the decision request, such as task type, complexity, priority, and data volume, used to characterize the task's inherent requirements. The current system state reflects the real-time status of the intelligent decision engine and its operating environment, such as CPU utilization, memory usage, and network latency, used to assess the system's carrying capacity. Historical decision tasks refer to all decision requests processed by the intelligent decision engine in the past, along with their corresponding processing results, resource consumption, performance metrics, etc., used for model training and capability profile construction. Decision branches refer to different technical modules or strategies within the intelligent decision engine used to process decision tasks, such as rule engines, statistical models, and large language models; each branch has different capabilities and applicable scenarios.

[0078] In this embodiment, dynamic capability profiling refers to the real-time or near-real-time evaluation and modeling of the performance, cost, latency, confidence level, and robustness of each decision branch under different task complexities and system loads, forming a capability description. The pre-trained Pareto front prediction model based on deep learning technology, trained with sufficient historical data, can predict the Pareto fronts of different scheduling schemes under multiple optimization objectives (such as cost, latency, and confidence level), i.e., a series of optimal non-dominated solutions. The candidate scheduling scheme set for capability calibration refers to a set of potential scheduling schemes generated by the prediction model. These schemes have been verified and adjusted based on the actual capabilities of the decision branches to ensure their feasibility and effectiveness, and constitute a set of non-dominated solutions under multi-objective optimization. Current business preference refers to the priority or weight settings for different optimization objectives (such as cost, latency, confidence level, and robustness) in a specific business scenario, used to guide the selection of the final scheme from the candidate scheduling scheme set. The scheduling scheme refers to the specific decision branch selected for a specific decision request and its corresponding processing strategy, which is the final decision instruction output by the intelligent decision engine.

[0079] In this embodiment, the intelligent decision engine adaptive scheduling method based on multi-objective optimization first receives a decision request from the business system and extracts task characteristics and the current system state from the decision request. Specifically, the decision request can be received by the system through a standard API interface or message queue. Task characteristics may include task type, priority, data size, etc., while the current system state may include CPU load, memory usage, network latency, etc. For example, the system can be configured with a listener to continuously monitor incoming decision requests and parse the structured data in the request body to extract task characteristics, while simultaneously obtaining real-time system operation indicators as the current system state through system monitoring tools. In one implementation, task characteristics and system state can be simply recorded in a log file for subsequent processing modules to read. In another implementation, this information can be directly passed to the next processing module, but may lack a unified format or preprocessing.

[0080] Furthermore, based on historical decision-making tasks, dynamic capability profiles are constructed for each decision branch. The data from historical decision-making tasks can include information such as the task types, processing times, resource consumption, and success rates handled by each decision branch. For example, historical data can be periodically statistically analyzed to generate indicators such as average processing time and average resource consumption for each decision branch, which are then used as part of its capability profile. In one implementation, the capability profile can be manually configured with static parameters; for example, the capability profile of a rule engine can be set to low cost and high determinism, while the capability profile of a large language model can be set to high cost and high semantic understanding capability. In another implementation, the capability profile can be periodically updated based on a simple statistical model; for example, the average performance metrics of each branch can be calculated at fixed intervals.

[0081] In this embodiment, the task features, the current system state, and the dynamic capability profile are then input into a pre-trained Pareto front prediction model based on deep learning. For example, this information can be simply concatenated or encoded to form an input vector, which is then fed into the pre-trained model. In one implementation, the task features, system state, and capability profile can be directly used as the model's raw input features without complex preprocessing or feature engineering. In another implementation, these inputs can be fed into different sub-networks for preliminary processing, and then the processing results are simply fused.

[0082] In this embodiment, the task characteristics, the current system state, and the dynamic capability profile are processed by a deep learning-based pre-trained Pareto front prediction model to generate a set of capability-calibrated candidate scheduling schemes. The model analyzes and predicts the input information through its complex internal logic, outputting a series of potential scheduling schemes. For example, the model can predict the performance metrics (such as cost, latency, and confidence) that different combinations of decision branches may achieve given the task and system state. In one implementation, the model may directly output a single "optimal" scheduling scheme without considering the trade-offs under multi-objective optimization. In another implementation, the model may output multiple scheduling schemes, but these schemes may not have undergone rigorous capability calibration or form a true Pareto front set.

[0083] In this embodiment, a scheduling scheme is selected from the candidate scheduling scheme set based on current business preferences. Business preferences can be reflected in the priority settings for different optimization objectives; for example, prioritizing low cost in some scenarios and low latency in others. For instance, each scheme in the candidate scheduling scheme set can be scored according to preset business rules, and then the scheme with the highest score can be selected. In one implementation, business preferences can be hard-coded as fixed weights; for example, always setting the cost weight to 0.5, the latency weight to 0.3, and the confidence weight to 0.2. In another implementation, a human expert can manually select a scheme from the candidate set based on experience.

[0084] In this embodiment, the decision request is finally distributed to the corresponding decision branch for processing according to the scheduling scheme. For example, if the scheduling scheme specifies a rules engine branch, the decision request is routed to the rules engine for processing. In one implementation, distribution can be achieved through simple conditional judgments and direct calls to the target branch's API. In another implementation, a simple load balancer can be used to distribute requests evenly to all available decision branches, regardless of the specific scheduling scheme.

[0085] In this embodiment, by dynamically constructing a capability profile of decision branches and combining it with a deep learning model to predict the Pareto front, the system can adaptively select the optimal scheduling scheme from multiple technology branches based on task characteristics, system state, and business preferences. This effectively solves the "precise scheduling dilemma" inherent in traditional single-technology stacks in complex business scenarios, significantly improves computing resource utilization, reduces decision-making costs, and enhances the flexibility and overall quality of the decision-making process, while also possessing better adaptive degradation and fault tolerance capabilities.

[0086] In one feasible implementation, refer to Figure 2 Step S200 includes steps S210 to S240, wherein:

[0087] Step S210: Collect the actual processing results of the decision branches to obtain the actual result set;

[0088] Step S220: Based on the actual result set, analyze the confidence change pattern of the decision branch under multiple task complexities, generate relationship curve data, analyze the cost change pattern of the decision branch in processing multiple feature tasks, generate correlation model data, and analyze the latency change pattern of the decision branch under multiple system loads, generate response model data.

[0089] Step S230: Generate a branch robustness score based on the historical anomaly handling records of the decision branches;

[0090] Step S240: Integrate the relationship curve data, the association model data, the response model data, and the branch robustness score to generate a dynamic capability profile.

[0091] In this embodiment, the actual processing results refer to the various performance indicators and behavioral data generated by the decision branch after executing a specific decision task. These results may include, but are not limited to, task completion time, resource consumption (such as CPU and memory usage), decision accuracy, output quality, and errors or anomalies during task execution. By continuously collecting actual processing result data, for example by deploying monitoring agents or logging modules when the decision branch executes tasks, and capturing and storing these key indicators in real time, a comprehensive set of actual results can be formed, providing a foundation for subsequent analysis of the decision branch's true capabilities.

[0092] In this embodiment, based on the actual result set, the confidence level variation pattern of the decision branch under multiple task complexities is analyzed to generate relationship curve data. The confidence level variation pattern reflects the reliability or success rate of the decision branch when handling tasks of different complexities. Task complexity can be quantified according to dimensions such as the amount of input data, computational load, and number of logical branches. By performing statistical analysis on the actual result set, for example, calculating the success rate, error rate, or prediction accuracy of the decision branch within different complexity ranges, a relationship curve of confidence level versus task complexity can be plotted. These relationship curve data can intuitively demonstrate the performance boundaries and stability of the decision branch when facing different challenges. Simultaneously, the cost variation pattern of the decision branch handling multiple characteristic tasks is analyzed to generate correlation model data. The cost variation pattern refers to the resources consumed or expenses incurred by the decision branch when handling tasks with specific characteristics. Characteristic tasks can refer to tasks with specific business attributes, data types, or computational patterns. Costs can include computational costs (such as CPU time and memory usage), storage costs, network transmission costs, etc. By analyzing resource consumption data for tasks with different characteristics in actual result sets, cost-related models can be constructed. For example, regression analysis or machine learning models can be used to establish a mapping relationship between task characteristics and costs. This correlation model data helps predict the resource requirements of future tasks. Furthermore, analyzing the latency variation patterns of decision branches under multiple system loads generates response model data. The latency variation patterns describe the response speed of the decision branch under different system load conditions. System load can be measured by metrics such as the number of concurrent requests, CPU utilization, memory utilization, and network bandwidth. By monitoring the task processing latency (such as request response time and processing queue waiting time) of the decision branch under different load levels, latency response model data can be generated. These models can reveal the performance bottlenecks of the decision branch and predict its performance under high load conditions, for example, by fitting a functional relationship between load and latency.

[0093] In this embodiment, a branch robustness score is generated based on the historical anomaly handling records of the decision branch. The branch robustness score measures the recovery capability and stability of the decision branch in the face of abnormal situations (such as input data errors, external service failures, internal logic errors, etc.). Historical anomaly handling records include the frequency of anomalies, anomaly types, anomaly durations, and the time and resources required for the system to recover from the anomalies. By quantitatively analyzing these records—for example, calculating the anomaly rate, average recovery time, or the severity of the anomaly's impact on business—a comprehensive branch robustness score can be generated. The higher the score, the more stable and reliable the performance of the decision branch under abnormal conditions.

[0094] In this embodiment, the relationship curve data, the association model data, the response model data, and the branch robustness score are finally integrated to generate a dynamic capability profile. The integration process involves uniformly modeling and representing the aforementioned multi-dimensional and multi-type data. This can be achieved using various techniques, such as inputting these data as feature vectors into a machine learning model for training, or constructing a multi-dimensional knowledge graph to represent the comprehensive capabilities of the decision branches. The dynamic capability profile is a comprehensive description that is updated in real-time and fully reflects the performance, cost, latency, and robustness of the decision branches. Its "dynamic" nature is reflected in the continuous collection and updating of data; the profile can reflect the latest state and trends of the decision branches in real time, rather than static fixed values.

[0095] In this embodiment, by collecting actual processing results and deeply analyzing the confidence levels of decision branches under different task complexities, the costs of different characteristic tasks, and the latency variation patterns under different system loads, and combining historical anomaly handling records to generate robustness scores, the generated dynamic capability profile not only includes static performance indicators but also incorporates the behavioral patterns and reliability characteristics of decision branches under various dynamic environments. This multi-dimensional, real-time updated profile greatly improves the accuracy of characterizing the true capabilities of decision branches. When these refined dynamic capability profiles are input into a pre-trained Pareto front prediction model based on deep learning, the model can more accurately understand the strengths and weaknesses of each decision branch, thereby generating a more realistic and feasible set of candidate scheduling schemes during multi-objective optimization. This avoids scheduling biases caused by insufficient or distorted capability profile information, ensuring that the intelligent decision engine can adaptively select a scheduling scheme that achieves the best balance between performance, cost, latency, and robustness based on the current task characteristics and system state, thus significantly improving the scheduling efficiency and decision quality of the entire intelligent decision engine.

[0096] In one feasible implementation, refer to Figure 3 Step S300 includes steps S310 to S340, wherein:

[0097] Step S310: Perform multi-scale feature encoding processing on the task features to generate task feature codes; perform temporal state encoding processing on the current system state to generate system state codes; perform structured profile encoding processing on the dynamic capability profile to generate capability profile codes.

[0098] Step S320: Based on the cross-attention mechanism, calculate the interaction weights among the task feature encoding, the system state encoding, and the capability profile encoding;

[0099] Step S330: Based on the interaction weight, perform weighted fusion processing on the task feature encoding, the system state encoding, and the capability profile encoding to generate a fused encoding;

[0100] Step S340: Input the fused encoding into the multi-target decoder, process the fused encoding through multi-task parallel learning, and generate a prediction of the candidate scheduling scheme.

[0101] In this embodiment, multi-scale feature encoding is performed on the task features to transform the original task features (such as task type, priority, data volume, computational requirements, etc.) into a numerical representation with multi-level semantic information that can be effectively processed by deep learning models. This can be achieved in various ways. For example, for discrete features, an embedding layer can be used to map them to a low-dimensional dense vector space; for continuous features, standardization or normalization can be performed; for structured or sequential task descriptions, convolutional neural networks (CNNs) or recurrent neural networks (RNNs) can be used to extract feature patterns at different scales, thereby generating a unified task feature encoding.

[0102] In this embodiment, temporal state encoding is performed on the current system state to capture the dynamic information of system state (such as CPU utilization, memory usage, network latency, queue length, etc.) changing over time. System state typically changes continuously, and its historical trends are crucial for predicting future system behavior. Temporal state encoding can process a sequence of system states over a period of time using recurrent neural network models such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), or mechanisms such as Temporal Convolutional Networks (TCNs) and Transformer encoders, thereby generating a system state code that reflects the current system state and its historical evolution trend.

[0103] In this embodiment, the dynamic capability profile is subjected to structured profile encoding processing, aiming to convert the dynamic capability profile of the decision branch (including relationship curve data, correlation model data, response model data, and branch robustness scores, etc.) into a unified, structured numerical representation. Since the capability profile may contain multiple types of data (such as curve data, model parameters, and scalar scores), structured encoding needs to effectively integrate this heterogeneous information. This can be achieved by designing specialized sub-networks (e.g., using small convolutional networks for curve data and multilayer perceptrons for model parameters) to encode different components of the profile, and then concatenating or further fusing these codes to ultimately generate a structured profile encoding that comprehensively represents the capabilities of the decision branch.

[0104] In this embodiment, the cross-attention mechanism allows the model to dynamically focus on the most relevant parts of different types of encodings. For example, task feature encoding can be used as a query, and system state encoding and capability profile encoding can be used as keys and values, thereby calculating the attention weights between task features and system state, and between task features and capability profiles. Similarly, the attention weights between system state and capability profiles can also be calculated. These interaction weights reflect the strength and importance of information flow between different encodings, helping the model understand the deep dependencies between them.

[0105] In this embodiment, the task feature encoding, system state encoding, and capability profile encoding are weighted and fused according to the interaction weights to generate a fused encoding. After calculating the interaction weights, these weights are used to combine the original or attention-weighted task feature encoding, system state encoding, and capability profile encoding. The fusion process can be performed in various ways. For example, the attention-weighted encodings can be concatenated and then dimensionality reduced and integrated through a fully connected layer; or, the encodings can be weighted and summed according to the interaction weights to generate a single fused encoding that contains all input information and highlights key interactions. This fused encoding serves as the unified input for subsequent prediction tasks.

[0106] In this embodiment, the fused encoding is input into a multi-objective decoder. The fused encoding is processed through multi-task parallel learning to generate predictions of candidate scheduling schemes. The multi-objective decoder receives the fused encoding as input and is designed to simultaneously predict multiple objectives related to the candidate scheduling scheme. For example, it predicts the probability of each decision branch being selected, the confidence level of the scheduling scheme, its execution cost, and its latency. Multi-task parallel learning means that the decoder simultaneously optimizes these different prediction tasks. By sharing underlying parameters or parts of the network structure, the model can learn the common features and mutually reinforcing relationships between different tasks, thereby generating a comprehensive prediction of the candidate scheduling scheme.

[0107] In this embodiment, by performing multi-scale, temporal, and structured encoding on task features, system states, and dynamic capability profiles, heterogeneous raw data can be transformed into unified feature representations rich in semantic information, effectively solving the processing challenges brought about by data diversity. Furthermore, by calculating the interaction weights between these encodings based on a cross-attention mechanism, the model can deeply understand the complex relationships between tasks, systems, and decision branch capabilities, accurately capturing key interaction information and avoiding information loss or noise interference that may result from simple splicing. Subsequently, weighted fusion is performed based on these interaction weights to generate a highly condensed fusion encoding containing global context, providing high-quality input for subsequent predictions. Finally, the multi-objective decoder processes this fusion encoding through multi-task parallel learning, enabling simultaneous prediction of multiple performance indicators of candidate scheduling schemes. This generates more comprehensive and accurate scheduling scheme predictions that consider multi-dimensional optimization objectives, significantly improving the accuracy and robustness of the intelligent decision engine's adaptive scheduling.

[0108] In one feasible implementation, the step of calculating the interaction weights among the task feature encoding, the system state encoding, and the capability profile encoding based on the cross-attention mechanism includes: simultaneously calculating attention weight data between task features and system state, between task features and capability profile, and between system state and capability profile based on the task feature encoding, system state encoding, and capability profile encoding through the cross-attention mechanism; constructing a three-dimensional interaction attention matrix based on the attention weight data between task features and system state, between task features and capability profile, and between system state and capability profile; and performing attention normalization processing on the three-dimensional interaction attention matrix to generate an interaction weight distribution.

[0109] In this embodiment, based on the task feature encoding, the system state encoding, and the capability profile encoding, a cross-attention mechanism is used to simultaneously calculate attention weight data between task features and system state, between task features and capability profile, and between system state and capability profile. Specifically, the cross-attention mechanism is a neural network mechanism that can effectively capture the correlation between different input sequences or features. Here, it is configured to calculate the attention weight data between three pairs of data in parallel. For example, when calculating the attention weight data between task features and system state, the task feature encoding can act as a query, and the system state encoding as a key and value. The degree of attention the task features pay to the system state is determined by calculating the similarity between the query and the key, and vice versa. This simultaneous pairwise calculation ensures that all direct interactions between the task, the system, and the decision branch capabilities are fully considered and quantified.

[0110] Building upon this foundation, a three-dimensional interactive attention matrix is ​​constructed based on the attention weight data between task features and system state, between task features and capability profiles, and between system state and capability profiles. This three-dimensional interactive attention matrix is ​​an integrated representation of all the aforementioned pairwise attention weight data. It provides a unified, global perspective to examine the complex interaction patterns among task feature encoding, system state encoding, and capability profile encoding. For example, this matrix can be a tensor, with different dimensions corresponding to the interaction strength between task features, system state, and capability profiles, thereby more comprehensively characterizing their mutual influence.

[0111] In this embodiment, the three-dimensional interactive attention matrix is ​​then subjected to attention normalization to generate an interaction weight distribution. Attention normalization aims to convert the original attention weight data into a probability distribution, ensuring that the sum of all weights is a specific value (e.g., 1), so that these weights can be directly used in subsequent weighted fusion processing. Common normalization methods include the Softmax function, which can convert any real value into a probability value between 0 and 1, guaranteeing that the sum of all output values ​​is 1. Through this normalization process, a clear and interpretable interaction weight distribution can be obtained, accurately indicating the degree of influence of task characteristics, system state, and capability profile on the generation of the final decision scheme in the current context.

[0112] In this embodiment, through the above technical solution, this application can systematically calculate the attention weight data between each pair of task feature encoding, system state encoding, and capability profile encoding via a cross-attention mechanism, thereby comprehensively capturing the complex interaction relationships between them. Furthermore, by integrating these pairwise interaction attention weight data into a three-dimensional interaction attention matrix and performing attention normalization, an accurate and globally-oriented interaction weight distribution can be generated. This refined interaction weight calculation method ensures that subsequent fusion encoding can more accurately reflect the deep correlation between the task, system, and decision branch capabilities, thus laying the foundation for the multi-objective decoder to generate more valuable candidate scheduling scheme predictions, significantly improving the accuracy and effectiveness of the intelligent decision engine's adaptive scheduling.

[0113] In one feasible implementation, refer to Figure 4 Step S400 includes steps S410 to S450, wherein:

[0114] Step S410: Based on the relationship curve data, association model data, and response model data in the dynamic capability profile, process and generate benchmark confidence data, benchmark cost data, and benchmark latency data;

[0115] Step S420: Use the benchmark confidence data, the benchmark cost data, and the benchmark delay data as prior constraints;

[0116] Step S430: Perform joint constraint coding processing on the task feature code, the system state code, the capability profile code, and the prior constraints to generate constraint-aware code;

[0117] Step S440: Based on the constraint-aware coding, generate prediction results for multiple candidate scheduling schemes;

[0118] Step S450: Perform non-dominated sorting on the prediction results of the multiple candidate scheduling schemes, filter out the candidate scheduling schemes on the Pareto front, and generate a set of candidate scheduling schemes.

[0119] In this embodiment, based on the relationship curve data, association model data, and response model data in the dynamic capability profile, baseline confidence data, baseline cost data, and baseline latency data are first generated. These baseline data aim to extract specific, quantifiable performance benchmarks from the dynamic capability profile of the decision branch. The dynamic capability profile (as described above) includes the confidence variation patterns of the decision branch under different task complexities (relationship curve data), the cost variation patterns for handling different characteristic tasks (association model data), and the latency variation patterns under different system loads (response model data). By analyzing and processing this historical data, the expected baseline confidence, baseline cost, and baseline latency of each decision branch in the current context can be inferred based on the task characteristics of the currently received decision request and the current system state. For example, based on the complexity of the task characteristics, interpolation or querying can be performed on the relationship curve data to obtain the corresponding baseline confidence; based on the task type, the corresponding baseline cost can be found in the association model data; and based on the current system load, the corresponding baseline latency can be obtained from the response model data. These baseline data reflect the performance boundaries and typical performance of the decision branch in actual operation.

[0120] In this embodiment, the baseline confidence data, baseline cost data, and baseline latency data are then used as prior constraints. Treating the generated baseline performance data as prior constraints means that these data will serve as important references or limitations in the subsequent prediction process. They provide the deep learning model with "hard" boundaries or "soft" guidance regarding the actual operational capabilities of decision branches, ensuring that the model fully considers these actual performance limitations when generating candidate scheduling schemes. These prior constraints can be represented as numerical vectors or tensors and input into specific modules of the model along with task feature encoding, system state encoding, and capability profile encoding to influence the model's internal state and prediction logic.

[0121] Building upon this, the task feature encoding, system state encoding, capability profile encoding, and prior constraints are jointly constrained and encoded to generate constraint-aware encoding. This step is crucial for capability calibration. It deeply integrates the abstract task feature encoding, system state encoding, and capability profile encoding with specific prior constraints. The goal of the joint constraint encoding process is to generate a new, context-rich representation, namely constraint-aware encoding. This encoding not only includes the original information about the task, system, and branch capabilities but also inherently incorporates the actual performance boundaries of the decision branches. This encoding enables the model to "sense" and respect these constraints during subsequent predictions, thereby avoiding the generation of unrealistic scheduling schemes. Specifically, as described above, this can be achieved through a gated attention mechanism, where the prior constraint vector can adjust the observed feature vector (task feature encoding, system state encoding, and capability profile encoding), dynamically adjusting the model's attention to different features to generate a feature representation sensitive to constraints.

[0122] Furthermore, based on the constraint-aware encoding, predictions for multiple candidate scheduling schemes are generated. Once the constraint-aware encoding is obtained, the deep learning model can use this encoding to predict multiple potential scheduling schemes. Since the constraint-aware encoding has internalized the actual performance limitations of the decision branches, the predictions generated based on it will be more realistic and feasible. These predictions may include multiple dimensions such as the branch selection probability, expected confidence level, expected cost, and expected latency for each candidate scheme. As described above, this can be achieved by performing parallel processing on the constraint-aware encoding, such as branch selection feature extraction, confidence level prediction feature extraction, cost assessment feature extraction, and latency estimation feature extraction, thereby generating comprehensive prediction results.

[0123] In this embodiment, the prediction results of the multiple candidate scheduling schemes are finally subjected to non-dominated ranking to filter out candidate scheduling schemes on the Pareto front, generating a set of candidate scheduling schemes. In multi-objective optimization problems, there is usually no single optimal solution, but rather a set of solutions that achieve the best trade-offs among different objectives; this set of solutions is called the Pareto front. Non-dominated ranking (also known as Pareto ranking) is used to identify these non-dominated solutions from the prediction results of multiple candidate scheduling schemes. For each candidate scheme, its prediction confidence, prediction cost, and prediction delay can be used as optimization objectives. A scheme that cannot be completely surpassed by another scheme on all objectives is called a non-dominated solution. Through this ranking, a set containing multiple optimal trade-off schemes can be obtained, namely, the set of candidate scheduling schemes on the Pareto front. This set provides diverse and capability-calibrated high-quality options for subsequent selection based on business preferences.

[0124] In this embodiment, the above technical solution effectively addresses the problem that deep learning models may deviate from the actual capabilities of decision branches when predicting scheduling schemes. Specifically, by extracting baseline confidence data, baseline cost data, and baseline latency data from dynamic capability profiles as prior constraints, and performing joint constraint encoding processing with task feature encoding, system state encoding, and capability profile encoding, a constraint-aware encoding sensitive to actual performance boundaries is generated. This encoding ensures that the prediction results of subsequently generated candidate scheduling schemes are naturally integrated with the actual operational capabilities of the decision branches, thereby ensuring the "capability calibration" and practical feasibility of the predicted schemes. Furthermore, by performing non-dominated ranking processing on these prediction results, a set of candidate scheduling schemes located on the Pareto front can be selected. These schemes achieve optimal trade-offs among multiple optimization objectives (such as confidence, cost, and latency) and all conform to the actual capability range of the decision branches. This not only significantly improves the practicality and reliability of adaptive scheduling by the intelligent decision engine but also provides high-quality and diverse optimization options for subsequent flexible selection based on business preferences, thereby avoiding the generation of unrealistic or unexecutable scheduling decisions.

[0125] In one feasible implementation, the step of jointly constraining the task feature encoding, the system state encoding, the capability profile encoding, and the prior constraints to generate a constraint-aware encoding includes: converting the baseline confidence data, the baseline cost data, and the baseline latency data into prior constraint vectors; using the task feature encoding, the system state encoding, and the capability profile encoding as observation feature vectors; calculating the adjustment weights of the prior constraint vectors on the observation feature vectors based on a gating attention mechanism; and performing feature adjustment processing on the observation feature vectors according to the adjustment weights to generate a constraint-aware feature representation, which is then used as the constraint-aware encoding.

[0126] In this embodiment, the baseline confidence data, baseline cost data, and baseline latency data are generated based on the relationship curve data, correlation model data, and response model data in the dynamic capability profile of the decision branch. They represent the expected performance baseline of the decision branch under specific task complexity, feature tasks, and system loads. To uniformly input these heterogeneous baseline data into the deep learning model and enable them to serve as signals for adjusting other features, they need to be converted into standardized prior constraint vectors. For example, these baseline data can be numerically normalized, mapped to the interval [0, 1] or [-1, 1], and then concatenated into a fixed-dimensional numerical vector. Each dimension of this vector represents a specific prior constraint, providing a quantitative basis for subsequent feature adjustment.

[0127] In this embodiment, the task feature encoding, system state encoding, and capability profile encoding are obtained after processing with multi-scale feature encoding, temporal state encoding, and structured profile encoding. They respectively characterize the characteristics of the current decision request, the real-time status of the system operation, and the historical capability performance of the decision branch. These encodings are the core inputs for the model to make decision predictions, reflecting the "observational" information of the current scheduling scenario. To facilitate subsequent unified processing and constraint adjustment, these independent encodings can be concatenated or fused to form a comprehensive observation feature vector. This vector contains all feature information directly related to the current scheduling decision and requiring prior constraint adjustment.

[0128] Building upon this, this application introduces a gated attention mechanism to calculate the adjustment weight of the prior constraint vector on the observed feature vector. The gated attention mechanism is one that dynamically adjusts the importance of features based on the input signal. Specifically, the prior constraint vector can be transformed through one or more fully connected layers, and then a gate vector with the same dimension as the observed feature vector can be generated using an activation function (e.g., the sigmoid function). Each element of this gate vector has a value between 0 and 1, representing the adjustment strength or importance of the prior constraint on the corresponding dimension of the observed feature vector. This mechanism allows the model to dynamically "gated" or "focused" on the constraint-related parts of the observed feature vector based on a preset performance baseline (prior constraints), thereby ensuring that subsequent decision-making processes fully consider these constraints.

[0129] In this embodiment, after obtaining the adjustment weights of the prior constraint vector on the observed feature vector, these weights are applied to the observed feature vector through feature adjustment processing. Specifically, the adjustment weights generated by the gating attention mechanism can be multiplied element-wise with the observed feature vector (or other forms of interaction such as weighted summation). This multiplication operation allows the parts of the observed feature vector that do not conform to the constraints or need to be emphasized to be enhanced or suppressed accordingly based on the adjustment weights. The adjusted feature vector is the constraint-aware feature representation, which not only contains the original task features, system state, and capability profile information, but more importantly, this information has been calibrated and optimized according to preset prior constraints such as baseline confidence, cost, and latency. Finally, this constraint-aware feature representation is used as constraint-aware encoding as input for the subsequent generation of candidate scheduling scheme prediction results.

[0130] In this embodiment, through the above technical solution, this application effectively solves the problem of how to deeply and dynamically fuse baseline performance constraints (such as confidence, cost, and latency) with core observation features (task features, system state, and capability profile) when generating constraint-aware codes. Traditional simple concatenation or linear combination methods are difficult to capture the subtle adjustment effect of constraints on features, which may lead to insufficient sensitivity of the generated codes to constraints. This application introduces a gated attention mechanism, which enables the prior constraint vector to dynamically calculate and generate adjustment weights for the observed feature vector. This means that the model no longer passively receives constraint information, but actively uses constraint information to "gated" and "calibrate" the observed features, thereby achieving deep perception of constraints at the feature level. Through this feature adjustment processing, the generated constraint-aware code not only contains rich original feature information, but more importantly, this information has been optimized and adjusted according to the preset performance baseline, ensuring that the subsequently generated candidate scheduling schemes can better meet these prior constraints, improving the compliance and robustness of scheduling decisions. This mechanism enables the model to understand and apply constraints more intelligently, avoiding suboptimal or non-compliant scheduling results caused by improper processing of constraint information, thereby improving the adaptive scheduling capability of the intelligent decision engine.

[0131] In one feasible implementation, the step of generating prediction results for multiple candidate scheduling schemes based on the constraint-aware coding includes: performing branch selection feature extraction processing based on the constraint-aware coding to generate a branch selection probability distribution for the candidate scheduling schemes; performing confidence prediction feature extraction processing based on the constraint-aware coding to generate a prediction confidence of the candidate scheduling schemes; performing cost evaluation feature extraction processing based on the constraint-aware coding to generate a prediction cost of the candidate scheduling schemes; performing delay estimation feature extraction processing based on the constraint-aware coding to generate a prediction delay of the candidate scheduling schemes; and integrating the branch selection probability distribution, the prediction confidence, the prediction cost, and the prediction delay to generate prediction results for the candidate scheduling schemes.

[0132] In this embodiment, branch selection feature extraction is performed based on the constraint-aware encoding to generate a branch selection probability distribution for candidate scheduling schemes. This process aims to predict the probability of each possible decision branch being selected based on the current comprehensive information (i.e., the constraint-aware encoding). Specifically, a dedicated branch selection prediction module can be designed. This module receives the constraint-aware encoding as input and processes it through a series of neural network layers (e.g., fully connected layers), ultimately outputting a vector. Each element of this vector corresponds to a specific decision branch, representing the probability of that branch being selected in the current context. To ensure the validity of the probabilities, a Softmax activation function is typically used to normalize the output, ensuring that the sum of the probabilities of all branches is 1. This step provides a quantitative basis for subsequent scheduling scheme selection regarding which branch is most likely to be selected.

[0133] In this embodiment, confidence prediction feature extraction is performed based on the constraint-aware encoding to generate the predicted confidence of candidate scheduling schemes. This process aims to evaluate the probability of each candidate scheduling scheme succeeding or achieving the expected results when performing a specific task. A separate confidence prediction module, which also takes constraint-aware encoding as input and processes it through a neural network, outputs a scalar value between 0 and 1, representing the predicted confidence of the scheduling scheme. A higher confidence indicates that the model has stronger confidence in predicting the performance of the scheme, and vice versa. This helps the system consider not only performance indicators but also the reliability of the scheme when selecting a scheduling scheme.

[0134] Furthermore, based on the constraint-aware encoding, cost evaluation feature extraction processing is performed to generate predicted costs for candidate scheduling schemes. This processing is used to predict the resource consumption or economic overhead required to execute each candidate scheduling scheme. A dedicated cost evaluation module receives the constraint-aware encoding and uses techniques such as regression neural networks to output a value representing the predicted cost. This cost may include, but is not limited to, computational resource consumption, storage overhead, network bandwidth usage, or actual economic costs. By predicting costs, the system can balance performance and resource efficiency when optimizing scheduling schemes, avoiding unnecessary resource waste.

[0135] Furthermore, based on the constraint-aware encoding, latency estimation feature extraction is performed to generate predicted latency for candidate scheduling schemes. This process aims to predict the time delay required to execute each candidate scheduling scheme. A separate latency estimation module takes the constraint-aware encoding as input and outputs a numerical value representing the predicted latency through regression models and other methods. Latency can encompass multiple aspects such as task processing time, data transmission time, and queuing time. Accurate latency estimation is crucial for business scenarios requiring real-time response or time sensitivity, helping the system select scheduling schemes that meet timeliness requirements.

[0136] In this embodiment, the branch selection probability distribution, prediction confidence, prediction cost, and prediction delay are finally integrated to generate candidate scheduling scheme prediction results. This step aggregates the four independent prediction results into a unified and comprehensive evaluation information. The integration process can encapsulate these prediction values ​​into a structured data unit, such as a data structure containing multiple fields or a multidimensional vector. Each candidate scheduling scheme will correspond to such an integrated prediction result, which includes detailed prediction data on branch selection, confidence, cost, and delay. This integration allows subsequent non-dominated ranking and multi-objective optimization to be performed from a unified, multi-dimensional perspective, thereby more comprehensively evaluating and selecting the optimal scheduling scheme.

[0137] In this embodiment, through the above technical solution, this application can effectively transform complex constraint-aware encoding into multi-dimensional, quantifiable candidate scheduling scheme prediction results. Specifically, by performing branch selection feature extraction, confidence prediction, cost assessment, and latency estimation in parallel, the system can generate a branch selection probability distribution, prediction confidence, prediction cost, and prediction latency for each potential scheduling scheme. This multi-dimensional prediction capability overcomes the limitations of single-index evaluation, enabling the system to more comprehensively consider the reliability, resource efficiency, and timeliness of the schemes in the subsequent scheduling scheme selection process. Integrating these independent prediction results to form a unified candidate scheduling scheme prediction result provides rich and accurate input for subsequent non-dominated ranking based on Pareto fronts, thereby significantly improving the accuracy and decision quality of the intelligent decision engine's adaptive scheduling method, ensuring that the selected scheduling scheme meets business preferences while taking into account multiple optimization objectives such as performance, cost, and latency.

[0138] In one feasible implementation, refer to Figure 5 Step S500 includes steps S510 to S550, wherein:

[0139] Step S510: Generate business preference weights through reinforcement learning dynamic weight adjustment processing;

[0140] Step S520: Obtain the constraint data for the current business scenario;

[0141] Step S530: Based on the business preference weights and the constraint data, calculate and process each candidate scheduling scheme in the candidate scheduling scheme set to generate a constraint satisfaction score;

[0142] Step S540: Based on the business preference weights, calculate and process the candidate scheduling schemes that meet the constraints to generate a comprehensive utility score;

[0143] Step S550: Perform a weighted comprehensive processing on the constraint satisfaction score and the comprehensive utility score, select the candidate scheduling scheme with the highest score based on the weighted comprehensive result, and generate a scheduling scheme.

[0144] In this embodiment, business preference weights are first generated through reinforcement learning dynamic weight adjustment. This process aims to dynamically determine the relative importance of different optimization objectives (e.g., cost, latency, confidence, robustness) when evaluating candidate scheduling schemes. This is an adaptive learning method, rather than using statically preset weights. Besides generating system state representation data based on system operating parameters and historical scheduling decision effects, then calculating dynamic weight configuration strategy data, generating weight allocation data based on task characteristics and system state, and finally adjusting weights based on decision execution feedback to generate business preference weights, other reinforcement learning models can also be used, such as agents based on Q-learning or Actor-Critic architectures. These agents learn through interaction with the environment and optimize weight allocation strategies based on long-term cumulative rewards. The agent can observe system states such as scheduling results, business metrics, and resource utilization, and adjust its weight outputs based on this feedback to maximize predefined business objectives. These weights can be represented as a vector, where each element corresponds to a specific optimization objective.

[0145] In this embodiment, the method further includes acquiring constraint data for the current business scenario. This feature ensures that practical operational limitations and business requirements are explicitly considered during the scheduling process. These constraints are crucial for ensuring the feasibility and acceptability of the selected solution. For example, constraint data may include hard constraints such as the maximum allowed processing time for a specific task type, the minimum confidence level required for critical decisions, budgetary limits on resource consumption, or compliance requirements. This data can be obtained from a configuration database, a real-time monitoring system, or directly from the business request system, and can be structured into a set of rules, thresholds, or logical expressions.

[0146] Based on this, this method calculates and processes each candidate scheduling scheme in the aforementioned candidate scheduling scheme set using the business preference weights and constraint data described above, generating a constraint satisfaction score. This step is used to quantitatively evaluate the degree to which each potential scheduling scheme complies with predefined business constraints. For each candidate scheduling scheme, the system compares its prediction performance metrics (e.g., prediction cost, prediction latency, prediction confidence) with the acquired constraint data. For example, if a constraint specifies a maximum latency of 100 milliseconds, and a candidate scheme predicts a latency of 120 milliseconds, then the satisfaction score of that scheme on that constraint will be low (or zero). Business preference weights can be used to distinguish the importance of different constraints or to measure the severity of violations of different constraints. The constraint satisfaction score can be a binary value (0 or 1 for hard constraints) or a continuous value (e.g., a penalty function for soft constraints).

[0147] In this embodiment, the method calculates and processes candidate scheduling schemes that meet the constraints based on the aforementioned business preference weights, generating a comprehensive utility score. This step aims to evaluate the overall expected value or value of candidate schemes that have passed the constraint screening and considers multiple optimization objectives based on the current business priority. For candidate schemes that pass the constraint satisfaction check (or obtain a sufficiently high constraint satisfaction score), their prediction performance on various optimization objectives (e.g., prediction confidence, prediction cost, prediction latency, prediction robustness) will be comprehensively considered. The business preference weights generated by reinforcement learning will be applied here. For example, if the business currently prioritizes low latency, the utility weight of the latency component of the scheme will be higher than the weight of its cost component. The comprehensive utility score can be a weighted sum or a more complex multi-attribute utility function.

[0148] In this embodiment, the method finally performs a weighted summation of the constraint satisfaction score and the comprehensive utility score, and selects the candidate scheduling scheme with the highest score based on the weighted summation result to generate a scheduling scheme. This final selection step integrates feasibility (constraint satisfaction) and expected value (utility) into a single decision index, ensuring that the selected scheme is both feasible and in line with current business objectives. The constraint satisfaction score and comprehensive utility score can be combined using weighted summation, hierarchical methods (e.g., filtering by constraints first, then ranking by utility), or multi-criteria decision-making methods. For example, a simple method could be "Final Score = Constraint Weight × Constraint Satisfaction Score + Utility Weight × Comprehensive Utility Score", where the constraint weight and utility weight reflect the relative importance of satisfying constraints and maximizing utility. The scheme with the highest score will be selected as the optimal scheduling scheme. This ensures that even if a scheme has very high utility, it will not be selected if it violates key constraints; conversely, a scheme that barely satisfies constraints but provides excellent utility may be prioritized based on its weights.

[0149] In this embodiment, through the above technical solution, this application can adaptively generate business preference weights based on system operating parameters and historical scheduling decision effects, thereby enabling scheduling decisions to respond in real time to changes in the business environment and avoiding the limitations of static preference settings. Simultaneously, by acquiring constraint data of the current business scenario and calculating the constraint satisfaction score of candidate scheduling schemes, the actual feasibility of the selected scheduling scheme is ensured, avoiding the selection of schemes that do not conform to business specifications or resource limitations. Furthermore, by calculating the comprehensive utility score of candidate scheduling schemes that satisfy the constraints and combining it with the constraint satisfaction score for weighted comprehensive processing, this application can comprehensively evaluate candidate schemes from multiple dimensions, maximizing business value while ensuring the feasibility of the schemes. This dynamic and comprehensive evaluation and selection mechanism enables the intelligent decision engine to perform adaptive scheduling more flexibly and efficiently, significantly improving the quality of decisions and the robustness of the system, ensuring decision optimization and effective resource utilization in complex and ever-changing business scenarios.

[0150] In one feasible implementation, the step of generating business preference weights through reinforcement learning dynamic weight adjustment processing includes: generating system state representation data based on system operating parameters and historical scheduling decision effects; calculating and generating dynamic weight configuration strategy data based on the system state representation data; generating weight allocation data based on the current task characteristics and the current system state, according to the dynamic weight configuration strategy data; and adjusting the weight allocation data based on decision execution feedback to generate business preference weights.

[0151] In this embodiment, firstly, system state characterization data is generated based on system operating parameters and historical scheduling decision effects. System operating parameters may include, but are not limited to, hardware and software resource indicators such as current system CPU utilization, memory usage, network bandwidth, concurrent request count, and queue length. Historical scheduling decision effects encompass the actual execution results of past scheduling decisions, such as the actual processing time, resource consumption, decision accuracy, and user satisfaction of each decision branch. By collecting, cleaning, and aggregating this multi-dimensional data, a comprehensive system state characterization data reflecting the current operating status and historical performance of the system can be formed. This data provides crucial environmental awareness input for subsequent reinforcement learning models.

[0152] In this embodiment, secondly, dynamic weight configuration strategy data is calculated and generated based on the system state representation data. After obtaining the system state representation data, reinforcement learning (RL) algorithms, such as Q-learning, SARSA, and Actor-Critic, can be used to take the system state representation data as the state input of the RL environment. The RL model learns how to configure business preference weights to maximize long-term cumulative rewards under different system states through interaction with the environment. The reward function can be designed according to business objectives (such as minimizing cost, minimizing latency, maximizing confidence, etc.). After training, the RL model can output a set of dynamic weight configuration strategy data, which describes how to allocate weights to different decision objectives under a given system state to achieve optimal scheduling performance.

[0153] In this embodiment, weight allocation data is then generated based on the current task characteristics and the current system state, according to the dynamic weight configuration strategy data. Current task characteristics refer to the attributes of the decision request to be processed, such as task complexity, priority, and data volume. The current system state refers to the real-time system operating parameters at the time of scheduling. This information is combined with the previously learned dynamic weight configuration strategy data, which may be a multidimensional lookup table, a neural network model, or a set of rules. Based on the current task characteristics and system state, weight allocation data for the current scheduling scenario is queried or calculated from the strategy data. This weight allocation data specifically indicates the relative importance that each decision objective (such as confidence, cost, latency, etc.) should be assigned in the current specific context.

[0154] In this embodiment, finally, based on decision execution feedback, the weight allocation data is adjusted to generate business preference weights. Decision execution feedback refers to the actual effects produced by the system after the scheduling scheme is executed, such as actual decision latency, resource consumption, accuracy of decision results, and user satisfaction with the decision results. This feedback data is used as the reward signal for the reinforcement learning model to evaluate the merits of the current weight allocation. If the actual effect is unsatisfactory, the reinforcement learning model will adjust its internal parameters according to the feedback signal, thereby optimizing the dynamic weight configuration strategy data and influencing the generation of future weight allocation data. Through this continuous feedback loop and learning process, the weight allocation data is continuously optimized and adjusted, ultimately generating business preference weights that more accurately reflect the current business preferences and the actual situation of the system.

[0155] In this embodiment, by continuously monitoring system operating parameters and historical scheduling decision effects, the system can comprehensively perceive the current operating environment and historical performance, and transform this into system state representation data. Based on this, utilizing a reinforcement learning mechanism, the system can learn and extract the optimal dynamic weight configuration strategy from massive amounts of historical data and real-time feedback. When a new decision request arrives, combined with the current task characteristics and system state, the system can generate accurate weight allocation data based on the learned strategy. More importantly, by introducing decision execution feedback, the system can correct and optimize the weight allocation data in real time, forming a closed-loop adaptive learning process. This makes business preference weights no longer static preset values, but dynamically adjusted according to the actual system operating conditions, task characteristics, and business objectives, thereby ensuring that scheduling decisions are always highly aligned with the actual needs of the current business, significantly improving the decision quality, resource utilization efficiency, and user satisfaction of the intelligent decision engine in complex and ever-changing environments.

[0156] In the embodiments of this application, the intelligent decision engine adaptive scheduling method based on multi-objective optimization achieves adaptive decision scheduling by receiving decision requests, constructing dynamic capability profiles, inputting pre-trained models to generate candidate solutions, and selecting the optimal scheduling. It can dynamically match the optimal decision branch according to task characteristics and system status, adapt to dynamic business environments, improve resource utilization efficiency, and ensure stable decision quality.

[0157] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the adaptive scheduling method of the intelligent decision engine based on multi-objective optimization of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0158] This application also provides an adaptive scheduling system based on a multi-objective optimization intelligent decision engine, referencing... Figure 6 The intelligent decision engine adaptive scheduling system based on multi-objective optimization includes: a memory 10, a processor 20, and an intelligent decision engine adaptive scheduler based on multi-objective optimization stored in the memory 10 and executable on the processor 20. The intelligent decision engine adaptive scheduler based on multi-objective optimization is configured to implement the steps of the intelligent decision engine adaptive scheduling method based on multi-objective optimization.

[0159] The adaptive scheduling system for intelligent decision engines based on multi-objective optimization provided in this application adopts the adaptive scheduling method for intelligent decision engines based on multi-objective optimization in the above embodiments, which can adapt to dynamic business environments, improve resource utilization efficiency, and ensure stable decision quality. Compared with the prior art, the beneficial effects of the adaptive scheduling system for intelligent decision engines based on multi-objective optimization provided in this application are the same as those of the adaptive scheduling method for intelligent decision engines based on multi-objective optimization provided in the above embodiments, and other technical features of the adaptive scheduling system for intelligent decision engines based on multi-objective optimization are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0160] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0161] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. An adaptive scheduling method for an intelligent decision engine based on multi-objective optimization, characterized in that, The method includes: Receive decision requests from business systems, and obtain task characteristics and current system status from the decision requests; Based on historical decision-making tasks, construct dynamic capability profiles for decision-making branches; The task characteristics, the current system state, and the dynamic capability profile are input into a pre-trained Pareto front prediction model based on deep learning. The task features, the current system state, and the dynamic capability profile are processed by the deep learning-based pre-trained Pareto front prediction model to generate a set of candidate scheduling schemes that have been calibrated by capability. Based on current business preferences, a scheduling scheme is selected from the set of candidate scheduling schemes; According to the scheduling scheme, the decision request is distributed to the corresponding decision branch for processing; The steps for building dynamic capability profiles for decision branches based on historical decision-making tasks include: Collect the actual processing results of the decision branches to obtain the actual result set; Based on the actual result set, the confidence change pattern of the decision branch under multiple task complexities is analyzed to generate relation curve data, and the cost change pattern of the decision branch in processing multiple feature tasks is analyzed to generate correlation model data. In addition, the latency change pattern of the decision branch under multiple system loads is analyzed to generate response model data. Generate a branch robustness score based on historical anomaly handling records of decision branches; The relationship curve data, the association model data, the response model data, and the branch robustness score are integrated and processed to generate a dynamic capability profile; The steps of inputting the task features, the current system state, and the dynamic capability profile into a pre-trained Pareto front prediction model based on deep learning include: The task features are subjected to multi-scale feature encoding processing to generate task feature codes; the current system state is subjected to temporal state encoding processing to generate system state codes; and the dynamic capability profile is subjected to structured profile encoding processing to generate capability profile codes. Based on the cross-attention mechanism, the interaction weights among the task feature encoding, the system state encoding, and the capability profile encoding are calculated. Based on the interaction weights, the task feature encoding, the system state encoding, and the capability profile encoding are weighted and fused to generate a fused encoding. The fused code is input into a multi-target decoder, and the fused code is processed through multi-task parallel learning to generate predictions of candidate scheduling schemes; The steps of processing the task features, the current system state, and the dynamic capability profile using the deep learning-based pre-trained Pareto front prediction model to generate a set of capability-calibrated candidate scheduling schemes include: Based on the relationship curve data, association model data and response model data in the dynamic capability profile, benchmark confidence data, benchmark cost data and benchmark latency data are generated. The benchmark confidence data, the benchmark cost data, and the benchmark delay data are used as prior constraints. The task feature encoding, the system state encoding, the capability profile encoding, and the prior constraints are jointly constrained and encoded to generate constraint-aware encoding. Based on the constraint-aware coding, prediction results for multiple candidate scheduling schemes are generated; The prediction results of the multiple candidate scheduling schemes are sorted by non-dominated ordering to select candidate scheduling schemes on the Pareto front and generate a set of candidate scheduling schemes.

2. The adaptive scheduling method for intelligent decision engines based on multi-objective optimization as described in claim 1, characterized in that, The steps for calculating the interaction weights among the task feature encoding, the system state encoding, and the capability profile encoding based on the cross-attention mechanism include: Based on the task feature encoding, the system state encoding, and the capability profile encoding, attention weight data between task features and system state, attention weight data between task features and capability profile, and attention weight data between system state and capability profile are calculated simultaneously through a cross-attention mechanism. A three-dimensional interactive attention matrix is ​​constructed based on the attention weight data between task features and system state, the attention weight data between task features and capability profile, and the attention weight data between system state and capability profile. The three-dimensional interactive attention matrix is ​​subjected to attention normalization processing to generate an interactive weight distribution.

3. The adaptive scheduling method for intelligent decision engines based on multi-objective optimization as described in claim 1, characterized in that, The steps of jointly constraining and encoding the task feature code, the system state code, the capability profile code, and the prior constraints to generate a constraint-aware code include: The baseline confidence data, the baseline cost data, and the baseline delay data are converted into a priori constraint vector; The task feature encoding, the system state encoding, and the capability profile encoding are used as observation feature vectors; Based on the gated attention mechanism, the adjustment weight of the prior constraint vector on the observed feature vector is calculated; Based on the adjustment weights, the observed feature vector is subjected to feature adjustment processing to generate a constraint-aware feature representation, which is then used as a constraint-aware encoding.

4. The adaptive scheduling method for intelligent decision engines based on multi-objective optimization as described in claim 1, characterized in that, The steps for generating prediction results for multiple candidate scheduling schemes based on the constraint-aware coding include: Based on the constraint-aware coding, branch selection feature extraction processing is performed to generate the branch selection probability distribution of candidate scheduling schemes; Based on the constraint-aware coding, confidence prediction feature extraction is performed to generate the predicted confidence of candidate scheduling schemes; Based on the constraint-aware coding, cost evaluation feature extraction processing is performed to generate the predicted cost of candidate scheduling schemes; Based on the constraint-aware coding, delay estimation feature extraction is performed to generate candidate scheduling schemes for delay prediction. The branch selection probability distribution, the prediction confidence, the prediction cost, and the prediction delay are integrated and processed to generate the prediction results of the candidate scheduling scheme.

5. The adaptive scheduling method for an intelligent decision engine based on multi-objective optimization as described in claim 1, characterized in that, The steps for selecting a scheduling scheme from the candidate scheduling scheme set based on current business preferences include: Business preference weights are generated through dynamic weight adjustment processing using reinforcement learning. Obtain the constraint data for the current business scenario; Based on the business preference weights and the constraint data, each candidate scheduling scheme in the candidate scheduling scheme set is calculated and processed to generate a constraint satisfaction score. Based on the aforementioned business preference weights, candidate scheduling schemes that meet the constraints are calculated and processed to generate a comprehensive utility score. The constraint satisfaction score and the comprehensive utility score are weighted and combined, and the candidate scheduling scheme with the highest score is selected based on the weighted combination result to generate a scheduling scheme.

6. The adaptive scheduling method for intelligent decision engines based on multi-objective optimization as described in claim 5, characterized in that, The steps involved in generating business preference weights through reinforcement learning dynamic weight adjustment include: Based on system operating parameters and the effects of historical scheduling decisions, system state characterization data is generated. Based on the system state characterization data, dynamic weight configuration strategy data is calculated and generated; Based on the current task characteristics and the current system state, weight allocation data is generated according to the dynamic weight configuration strategy data. Based on the decision execution feedback, the weight allocation data is adjusted to generate business preference weights.

7. An adaptive scheduling system based on a multi-objective optimization intelligent decision engine, characterized in that, The intelligent decision engine adaptive scheduling system based on multi-objective optimization includes: a memory, a processor, and an intelligent decision engine adaptive scheduler based on multi-objective optimization stored in the memory and executable on the processor. The intelligent decision engine adaptive scheduler based on multi-objective optimization is configured to implement the steps of the intelligent decision engine adaptive scheduling method based on multi-objective optimization as described in any one of claims 1 to 6.

Images