Intelligent scheduling method for large language model in simulation deduction
By constructing a dynamic causal graph and graph neural network to evaluate the intervention effect of causal edges, the problem of efficient calling and deep analysis of large language models in simulation simulation systems is solved, and forward-looking intelligent scheduling is realized, which improves the decision support capability and resource utilization efficiency of simulation simulation systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-19
AI Technical Summary
The direct access to large language models in existing simulation and inference systems leads to high computational costs and response delays. Traditional triggering methods lack in-depth causal logic insights, resulting in fragmented and inefficient analysis results and low resource utilization.
By constructing a dynamic causal graph and using graph neural networks to evaluate the intervention effect of causal edges, key scenario situations can be identified in real time, triggering the intervention of large language models to optimize resource allocation and analysis guidance.
It enables accurate invocation and efficient analysis of large language models, improves the effectiveness of decision support and system resource utilization, and has adaptive optimization capabilities.
Smart Images
Figure CN122065978A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and simulation technology, specifically relating to an intelligent scheduling method for large language models in simulation and deduction. Background Technology
[0002] As large language models demonstrate their powerful potential in complex decision-making and analysis tasks, their application in simulation and simulation systems is gaining increasing attention. Simulation and simulation systems continuously generate massive amounts of high-dimensional, dynamic entity state time-series data during operation. Directly connecting large language models to real-time data streams for continuous analysis would result in extremely high computational costs and response delays, failing to meet the timeliness requirements of simulation and simulation. Conversely, relying solely on fixed rules or simple thresholds (such as "entity damage" or "indicator exceeding limits") to trigger large language model analysis has significant drawbacks: these methods are essentially "post-hoc," passive response mechanisms. They can only capture superficial, already-occurring anomalies, failing to penetrate the deeper causal logic and forward-looking risks of situational evolution, leading to a lack of interpretability and predictive decision-making in the analysis. Furthermore, this triggering method cannot provide high-quality, structured analytical guidance information to the invoked large language model, making it difficult for the powerful reasoning capabilities of the large language model to focus on high-value issues, resulting in fragmented and inefficient analysis results and low overall system resource utilization.
[0003] Existing technologies lack a method capable of automatically and in real-time extracting deep causal relationships from dynamic simulation data and intelligently determining when and how to guide large language models for in-depth analysis. Therefore, there is an urgent need for a technical solution that enables "proactive intelligent scheduling based on causal inference" to achieve precise timing of large language model invocation, structured analysis guidance, and efficient system resource utilization in simulation scenarios, thereby truly enhancing the decision support capabilities of intelligent simulation systems. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides an intelligent scheduling method for large language models in simulation and deduction. By constructing a dynamic causal graph and utilizing graph neural networks to evaluate the intervention effect of causal edges, it identifies key scenario situations requiring in-depth analysis in real time, thereby triggering the intervention of large language models, thus optimizing resource allocation and improving decision-making efficiency.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A method for intelligent scheduling of large language models in simulation and inference includes:
[0007] Step 1: Collect the state time-series data of each entity in the simulation system in real time and organize it into an entity state matrix sequence;
[0008] Step 2: Based on the entity state matrix sequence, dynamically calculate the intensity of directional causal influence between entities, and combine it with predefined domain relationship weighting factors to construct a weighted directed causal graph and form a sparse causal graph;
[0009] Step 3: Input the sparse causal graph from multiple consecutive time points into a pre-trained directed spatiotemporal graph neural network model to predict future trends, and calculate the intervention sensitivity of each edge in the sparse causal graph based on the directed spatiotemporal graph neural network model:
[0010] Step 4: Perform burst detection on the intervention sensitivity of each side, and calculate the comprehensive confidence level that triggers the intervention of the large language model by fusing it with the entity importance.
[0011] Step 5: When the overall confidence level exceeds the dynamic threshold, extract the local causal subgraph centered on the corresponding edge, generate a structured enhanced trigger information package containing causal chain descriptions and analysis suggestions, call the large language model for analysis, and optimize the triggering strategy online based on the feedback.
[0012] Furthermore, step 2, which involves constructing a weighted directed causal graph, specifically includes: calculating the basic causal influence strength between entities based on the entity state matrix sequence using a partially oriented coherence algorithm, wherein the calculation is limited to spatially adjacent entity pairs or entities with predefined logical associations; multiplying the basic causal influence strength by an asymmetric domain relation weight factor to obtain the final directed edge weights used to construct the causal graph, wherein the domain relation weight factor is determined based on a predefined directed relation type between entities; constructing a weighted directed causal graph based on the edge weights, and filtering out edges with weights lower than a preset significance threshold to form a sparse causal graph.
[0013] Furthermore, the directed spatiotemporal graph neural network model in step 3 includes a directed graph convolutional layer for processing the directed graph structure. When updating node features, the directed graph convolutional layer aggregates the in-neighbor node information and out-neighbor node information of the node, respectively.
[0014] Furthermore, the calculation of intervention sensitivity in step 3 specifically involves: feeding the edge weight matrix of the sparse causal graph at the current moment as a differentiable input into the directed spatiotemporal graph neural network model; calculating the absolute value of the partial derivative of the future situation prediction value output by the model with respect to each element in the edge weight matrix using automatic differentiation technology; and using the absolute value as the real-time intervention sensitivity of the corresponding edge.
[0015] Furthermore, the sudden increase detection in step 4 adopts the Z-score method based on a sliding window, specifically: for each edge, calculate the mean and standard deviation of its intervention sensitivity values over the most recent M consecutive time steps; if the result of subtracting the mean from the sensitivity value at the current moment and dividing by the standard deviation is greater than a preset threshold, then it is determined that the sensitivity of that edge has suddenly increased.
[0016] Furthermore, in step 4, the comprehensive confidence score is calculated as follows: for an edge that is determined to have a sudden increase in sensitivity, the magnitude of the increase is weighted and fused with the preset importance scores of the two entities connected to the edge to obtain a preliminary confidence score; based on the preliminary confidence score, a decay factor based on the recent triggering time interval is introduced for correction to suppress repeated triggering of the same problem, and the corrected result is the comprehensive confidence score.
[0017] Furthermore, the structured enhanced trigger information package generated in step 5 includes at least the following: trigger timestamp and corresponding comprehensive confidence level, natural language description of the key causal chain with the trigger edge as the core, and one or more specific questions suggesting that the large language model conduct in-depth analysis of the causal chain; the natural language description is automatically generated by filling in keywords of the entities, relationships and state changes involved through a predefined text template.
[0018] Furthermore, the online optimization in step 5 specifically involves: the system recording the complete context information for each trigger and the subsequent human feedback evaluation; using the collected feedback data, fine-tuning the key parameters in the trigger model, wherein the key parameters include at least the weight parameters of the confidence fusion formula and / or the threshold parameters for sensitivity surge detection.
[0019] Secondly, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned intelligent scheduling method for large language models in simulation and deduction.
[0020] Thirdly, the present invention provides a computer-readable storage medium storing executable instructions thereon, which, when executed by a processor, enable the processor to implement the aforementioned intelligent scheduling method for large language models in simulation and deduction.
[0021] The beneficial effects of this invention are as follows:
[0022] This invention achieves forward-looking intelligent scheduling based on causal inference: It abandons the traditional passive triggering mechanism based on surface-level anomalies. By constructing a dynamic causal graph and utilizing graph neural networks to quantify the intervention sensitivity of causal edges, it can proactively identify key causal relationships that, if intervened upon, will have a significant impact on the overall future situation. This endows the timing judgment of large language models with inherent causal explanatory power and decision-making foresight, transforming analytical decision-making from "post-event response" to "pre-event warning and in-event intervention," significantly improving the effectiveness of decision support.
[0023] This invention significantly improves the efficiency of large language model invocation and analysis: Through a lightweight front-end causal perception and computation module, it performs online screening and value assessment of massive real-time simulation data. The large language model is only triggered when a high-value, high-confidence causal situation change is detected, providing a structured "enhanced trigger information package." This greatly reduces unnecessary model invocation frequency, allowing expensive large language model computational resources to focus on the key analytical tasks requiring the most complex reasoning. Simultaneously, the structured guidance information enhances the analytical depth and relevance of the large language model, thereby optimizing the overall system resource utilization efficiency.
[0024] The system's adaptability and continuous optimization capabilities are enhanced: This invention establishes a complete online learning loop. The system records the context of each trigger, the generated causal information, and subsequent feedback evaluations (such as human scoring), and uses this feedback data to dynamically fine-tune key parameters of the triggering model (such as confidence fusion weights and detection thresholds). This allows the system to continuously optimize the accuracy and value orientation of its triggering judgments as it is used more extensively, gradually approaching the expectations of expert decision-making, thus possessing continuously evolving intelligent characteristics and improving the system's practicality and long-term effectiveness. Attached Figure Description
[0025] Figure 1 This is a schematic diagram illustrating the principle of an intelligent scheduling method for large language models in simulation and deduction according to the present invention. Detailed Implementation
[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0027] like Figure 1 As shown, this invention relates to an intelligent scheduling method for large language models in simulation and inference. By constructing a dynamic causal graph and using a graph neural network to evaluate the intervention effect of causal edges, it achieves real-time identification of key scene situations and triggers the intervention of large language models accordingly. The specific implementation method is as follows:
[0028] Step 1: Collect the state time-series data of each entity in the simulation system in real time and organize it into an entity state matrix sequence;
[0029] The system connects to the data output interface of the simulation platform in real time to collect the state time-series data of all combat entities during the simulation. The simulation platform mentioned here can be various simulation simulation systems (such as transportation, supply chain, emergency drills, etc.), suitable for dynamic simulation scenarios with multiple entities and multiple causal relationships. The entity state data includes at least location coordinates and the currently executed task label. The d-dimensional state data of N entities collected at each simulation time step t are organized into an entity state matrix. It also maintains a sliding time window of length L, storing the state matrix sequence of the most recent L time steps. The window length L needs to balance the sufficiency and real-time performance of the causal analysis. It can be set according to the duration of typical tactical actions, usually between 20 and 50 steps, to balance real-time performance and statistical reliability.
[0030] Step 2: Based on the entity state matrix sequence, dynamically calculate the intensity of directional causal influence between entities, and combine it with predefined domain relationship weighting factors to construct a weighted directed causal graph and form a sparse causal graph;
[0031] ① Based on the state matrix sequence within the sliding window, a partially directional coherence algorithm is used. By fitting a multivariate autoregressive model and calculating the directional coherence between entities in the frequency domain after excluding the influence of other variables, the strength of direct causal influence between entities is estimated. This dynamically calculates the direct causal influence between any two entities. and The strength of the causal influence between them at the current time step t Specifically, the calculation involves selecting time-series data from key dimensions of the state matrix. For candidate entity pairs, a low-order vector autoregressive (VAR) model is fitted based on data within a sliding window. The model parameters can be estimated using the least squares method. After performing a Fourier transform on the VAR coefficient matrix (the core parameter set of the low-order VAR model), the calculation is performed in the frequency domain from the entity... To the entity The partial directional coherence values are obtained by integrating the partial directional coherence values in the low-frequency band relevant to the decision. The calculation is performed only on spatially adjacent or logically related entity pairs to reduce computational complexity.
[0032] ② Context-weight the basic causal strengths to generate the final entities used to construct the causal graph. and edge weights The weighted formula is: ,in As a domain relationship weighting factor, taking a simulation system with attack and defense relationships between multiple entities as an example, this domain relationship weighting factor is comprehensively assigned based on predefined relationship types such as command and affiliation, sensor detection, and direct combat between entities. If there is a command relationship, a higher weight is assigned, and if it is only spatial proximity, a lower weight is assigned, thereby injecting domain knowledge into the graph structure.
[0033] ③ Construct an asymmetric edge weight matrix using the calculated directed edge weights. Construct a weighted directed causal graph at time t. The causal graph is stored in memory as an adjacency list. By traversing the edge weight matrix, the weights are compared with the quantiles based on the edge weight distribution to filter out edges with weights lower than a preset significance threshold, thus forming a sparse causal graph representation.
[0034] Step 3: Input the sparse causal graph from multiple consecutive time points into a pre-trained directed spatiotemporal graph neural network (GNN) model to predict future situations, and calculate the intervention sensitivity of each edge in the sparse causal graph based on the GNN model:
[0035] ① The causal graph sequence at T consecutive time points As input, a pre-trained graph neural network model for directed graphs is fed in. Its training objective is to predict the performance of K key metrics at P future time steps based on a historical causal graph. This model is a lightweight predictor, its core consisting of two directed graph convolutional layers to encode the asymmetric causal structure at each time step. The directed graph convolutional operation in each layer updates node features by aggregating information from the in-neighbor and out-neighbor communities, respectively. Taking node v at the Pth time step as an example... Taking feature update of a layer as an example, its simplified operation can be represented as:
[0036] ,in Indicates that node v is at the th Feature vectors (hidden states) of a layer. Indicates the first The feature vector of the layer, i.e., the updated hidden state. , Let V be the set of causal predecessors (sources of incoming edges) and successors (targets of outgoing edges) of node v. and For edge weight Normalized weights, For a trainable parameter matrix, The ReLU activation function is used, and CONCAT is the vector concatenation operation. The node feature sequence obtained after T time steps of this layer is then captured by a GRU layer to capture the temporal dynamics, ultimately outputting a K-dimensional vector prediction of the overall situation at the next P time steps. These correspond to K key metrics for different scenarios. The following explanations will use a single metric as an example.
[0037] ② Use the GNN model to output the encoding of the current situation and future predictions. Based on this model, calculate the value of each edge in the causal graph G(t). Intervention sensitivity The intervention sensitivity is calculated using automatic differentiation techniques and is defined as the predicted value. Border rights The absolute value of the partial derivative, i.e. The computation is efficiently performed using the automatic differentiation function of modern deep learning frameworks. During inference, the asymmetric edge weight matrix W(t) at the current time step is set as the differentiable input tensor of the model. A complete forward propagation is then performed to obtain... Then, the result can be calculated through a single backpropagation. The gradient matrix of all elements in W(t) Taking the absolute value of each element in the gradient matrix yields the real-time intervention sensitivity of all directed edges. .
[0038] Step 4: Perform burst detection on the intervention sensitivity of each side, and calculate the comprehensive confidence level that triggers the intervention of the large language model by fusing it with the entity importance.
[0039] ① For each edge, calculate the rate of change of its sensitivity over a short period of time. The Z-score method based on a sliding window is used to detect sudden increases in sensitivity: the mean of the M nearest sensitivity values for that edge is calculated. with standard deviation ,like If it is a sudden increase, then it is determined to be a sudden increase. The preset threshold;
[0040] ② For edges where a sudden increase in sensitivity is detected, measure the magnitude of the increase, i.e., the rate of change. The edge is determined by the preset importance scores of the two connected entity types and the current task role. We perform weighted fusion to calculate the initial confidence level for triggering LLM intervention for this edge. The specific fusion formula is a linear weighted average: ,in The min-max normalization function is used to normalize... and Mapped to the interval [0,1] respectively. The relative weights used to control data anomalies and prior importance are one of the key adjustable parameters of the system. This represents a weighted sum of the normalized results of the intervention sensitivity values for the k-th indicator. Let be the weighting coefficient for the k-th indicator. Finally, a decay factor is introduced for similar patterns that have been triggered recently. ( It can be simply designed as ,in For similar trigger intervals, the attenuation factor can be obtained by subtracting timestamps. Alternatively, other functions based on time intervals or trigger frequencies can be used to refine the initial confidence level and obtain the overall confidence level. This is to prevent the same issue from being triggered repeatedly within a short period of time.
[0041] Step 5: When the overall confidence exceeds the dynamic threshold, extract the local causal subgraph centered on the corresponding edge, generate a structured enhanced trigger information package containing causal chain description and analysis suggestions, call the large language model for analysis, and optimize the trigger strategy online based on the feedback.
[0042] ① When there is a certain edge Overall confidence level Exceeding the dynamic trigger threshold When the system is triggered, it first performs causal chain tracing and key subgraph extraction. (Based on edge...) Using the central node, a breadth-first search with finite depth is performed in the causal graph G(t) to extract local causal subgraphs containing key antecedents and consequences, thereby achieving the effect of visualizing the direct context that leads to a sudden increase in sensitivity.
[0043] ② Based on the extracted causal subgraph, an enhanced trigger information package is automatically generated. This package should contain at least the following structured content: trigger timestamp and confidence level; natural language description of the key causal chain, such as "abnormal fluctuations occur in the command link between entity A and entity B"; and one to three specific questions for the causal chain that suggest in-depth LLM analysis.
[0044] ③ Using the "enhanced trigger information package" generated in the previous step as the core context, format it into an input instruction that the large language model can understand through the prompt word template, and asynchronously call the large language model interface to request it to generate a deep analysis report. The system records the complete context of this trigger, including the feature vector at the time of triggering, confidence level, generated causal chain description, etc., as well as the subsequent human feedback evaluation. An online learning loop is established, and the parameters of the trigger model are fine-tuned using the collected feedback data. For example, if a certain type of high-confidence trigger continues to receive low value scores, the weight parameter α of the confidence fusion formula in step 4 is fine-tuned using the gradient descent method, or the threshold for sensitivity surge detection is adjusted. This allows the system's triggering decisions to evolve in a direction that better aligns with expert value expectations.
[0045] In summary, the intelligent scheduling method for large language models in simulation and deduction disclosed in this invention effectively addresses the shortcomings in the coordination between existing simulation systems and large language models by constructing a dynamic causal graph and utilizing graph neural networks to evaluate the intervention effect of causal edges. This achieves causal identification of key scenario situations and provides interpretable decision-making basis for the intervention of large language models. It is of great significance for realizing efficient and reliable intelligent command and decision support.
[0046] Secondly, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned intelligent scheduling method for large language models in simulation and deduction.
[0047] Thirdly, the present invention provides a computer-readable storage medium storing executable instructions thereon, which, when executed by a processor, enable the processor to implement the aforementioned intelligent scheduling method for large language models in simulation and deduction.
[0048] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for intelligent scheduling of large language models in simulation and deduction, characterized in that, include: Step 1: Collect the state time-series data of each entity in the simulation system in real time and organize it into an entity state matrix sequence; Step 2: Based on the entity state matrix sequence, dynamically calculate the intensity of directional causal influence between entities, and combine it with predefined domain relationship weighting factors to construct a weighted directed causal graph and form a sparse causal graph; Step 3: Input the sparse causal graph from multiple consecutive time points into a pre-trained directed spatiotemporal graph neural network model to predict future trends, and calculate the intervention sensitivity of each edge in the sparse causal graph based on the directed spatiotemporal graph neural network model: Step 4: Perform burst detection on the intervention sensitivity of each side, and calculate the comprehensive confidence level that triggers the intervention of the large language model by fusing it with the entity importance. Step 5: When the overall confidence level exceeds the dynamic threshold, extract the local causal subgraph centered on the corresponding edge, generate a structured enhanced trigger information package containing causal chain descriptions and analysis suggestions, call the large language model for analysis, and optimize the triggering strategy online based on the feedback.
2. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, Step 2, constructing a weighted directed causal graph, specifically includes: calculating the basic causal influence strength between entities based on the entity state matrix sequence using a partially oriented coherence algorithm; the calculation is limited to spatially adjacent entity pairs or entities with predefined logical associations; multiplying the basic causal influence strength by an asymmetric domain relationship weight factor to obtain the final directed edge weights used to construct the causal graph, wherein the domain relationship weight factor is determined based on a predefined directed relationship type between entities; constructing a weighted directed causal graph based on the edge weights, and filtering out edges with weights lower than a preset significance threshold to form a sparse causal graph.
3. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, The directed spatiotemporal graph neural network model in step 3 includes a directed graph convolutional layer for processing directed graph structures. When updating node features, the directed graph convolutional layer aggregates the in-neighbor node information and out-neighbor node information of the node, respectively.
4. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, The specific steps for calculating intervention sensitivity in step 3 are as follows: the edge weight matrix of the sparse causal graph at the current moment is fed into the directed spatiotemporal graph neural network model as a differentiable input; the absolute value of the partial derivative of the future situation prediction value output by the model with respect to each element in the edge weight matrix is calculated using automatic differentiation technology; and the absolute value is used as the real-time intervention sensitivity of the corresponding edge.
5. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, The sudden increase detection in step 4 adopts the Z-score method based on a sliding window. Specifically, for each edge, the mean and standard deviation of its intervention sensitivity values over the most recent M consecutive time steps are calculated. If the result of subtracting the mean from the sensitivity value at the current moment and dividing by the standard deviation is greater than a preset threshold, it is determined that the sensitivity of the edge has suddenly increased.
6. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, In step 4, the comprehensive confidence score is calculated as follows: for an edge that is determined to have a sudden increase in sensitivity, the magnitude of the increase is weighted and fused with the preset importance scores of the two entities connected to the edge to obtain a preliminary confidence score; based on the preliminary confidence score, a decay factor based on the recent triggering time interval is introduced for correction to suppress repeated triggering of the same problem, and the corrected result is the comprehensive confidence score.
7. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, The structured enhanced trigger information package generated in step 5 includes at least the following: trigger timestamp and corresponding comprehensive confidence level, natural language description of the key causal chain with the trigger edge as the core, and one or more specific questions suggesting that the large language model conduct in-depth analysis of the causal chain; the natural language description is automatically generated by filling in keywords of the entities, relationships and state changes involved through a predefined text template.
8. The intelligent scheduling method for large language models in simulation and deduction according to claim 1, characterized in that, The online optimization in step 5 specifically involves: the system recording the complete context information for each trigger and the subsequent human feedback evaluation; using the collected feedback data, fine-tuning the key parameters in the trigger model, wherein the key parameters include at least the weight parameters of the confidence fusion formula and / or the threshold parameters for sensitivity surge detection.
9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the intelligent scheduling method for large language models in simulation and deduction as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, It stores executable instructions, which, when executed by a processor, enable the processor to implement the intelligent scheduling method for large language models in simulation and deduction as described in any one of claims 1-8.