Search

CN-122021910-A - Mixed large model cascading intelligent decision method

CN122021910ACN 122021910 ACN122021910 ACN 122021910ACN-122021910-A

Abstract

The application discloses a mixed large model cascading intelligent decision-making method, and relates to the technical field of artificial intelligent decision-making. The method comprises the steps of receiving demand data of a complex decision scene, constructing a decision risk matrix and generating a layered risk type decision demand topological conformation by disassembling decision dimensions and defining accuracy and risk tolerance, adopting a large model cascade-risk guiding matching architecture to call a first large model and a second large model to execute cascade operation, generating a risk controllable type fine decision intermediate feature code table by eliminating redundant links, evaluating risk threshold suitability by a decision risk reciprocating check-optimizing architecture, reversely optimizing an out-of-standard link, fully deriving a logical fault, generating a final intelligent decision associated topological graph, collecting execution effect data of the final intelligent decision associated topological graph, updating the decision risk matrix, precipitating a decision template to a decision feature code knowledge base, and finishing iterative tuning of decision capability. And the decision pertinence and the risk controllability are improved.

Inventors

  • WANG JIAYING
  • CHEN HONGYU
  • LI HUAWEI
  • LI HAIYANG
  • YAO JINTIAN

Assignees

  • 北京甲板智慧科技有限公司

Dates

Publication Date
20260512
Application Date
20260203

Claims (10)

  1. 1. The intelligent decision method for the cascade connection of the mixed large models is characterized by comprising the following steps of: Step 1, receiving requirement data of a complex decision scene, dismantling decision dimensions through a decision requirement layering-risk modeling framework, defining precision and risk tolerance, constructing a decision risk matrix to mark dimension risk association relations, and generating a layering risk type decision requirement topological conformation; Step 2, based on hierarchical risk type decision demand topological conformation, adopting a large model cascade-risk guide matching architecture to call a first large model and a second large model to execute cascade operation, eliminating redundant links through a decision chain pruning operation framework, and generating a risk controllable type refined decision intermediate feature code table; Step 3, evaluating the suitability of a risk threshold value through a decision risk reciprocating check-optimization framework based on a risk controllable type fine decision intermediate feature code table, reversely optimizing an out-of-standard link, fully deriving a logic fault, and generating a final intelligent decision association topological graph; And 4, acquiring execution effect data of a final intelligent decision-making association topological graph, adjusting framework parameters and pruning rules through a cascade model weight-decision chain double-optimization adaptation mechanism, updating a decision risk matrix, precipitating a decision template to a decision feature code knowledge base, and completing iterative tuning of decision making capability.
  2. 2. The hybrid large model cascading intelligent decision method according to claim 1, wherein step 1 comprises: step 11, carrying out dimension disassembly on the requirement data of the complex decision scene, and dividing decision core dimensions, secondary dimensions and risk dimensions to generate a decision dimension topology; Step 12, defining a decision accuracy requirement and a risk tolerance threshold for each decision dimension to generate a dimension decision constraint correlation matrix; And 13, constructing a decision risk matrix, marking risk association relations among all dimensions, restraining the association matrix based on decision dimension topology and dimension decision, and generating a hierarchical risk type decision demand topology conformation by combining the decision risk matrix.
  3. 3. The hybrid large model cascading intelligent decision method according to claim 2, wherein step 11 comprises: Step 111, carrying out semantic analysis and feature extraction on decision targets, constraint conditions and business data in the demand data to generate a decision demand feature conformation; Step 112, dimension clustering is performed based on the service association degree of the feature conformation of the decision requirement, and the core dimension, the secondary dimension and the risk dimension are divided to generate a decision dimension topology, wherein nodes of the decision dimension topology are composed of dimension features, and the edge weight is determined by the dimension association degree.
  4. 4. The hybrid large model cascading intelligent decision method according to claim 2, wherein step 13 comprises: step 131, constructing a risk association quantization matrix, and calculating risk association scores of all dimensions by taking inter-dimension risk conduction probabilities as quantization indexes to generate a risk association score table; Step 132, marking high-association risk dimension pairs based on a risk association score table, and constructing a decision risk matrix; And step 133, generating a hierarchical risk type decision demand topology conformation based on the decision dimension topology, the dimension decision constraint correlation matrix and the decision risk matrix.
  5. 5. The hybrid large model cascading intelligent decision method according to claim 1, wherein step 2 comprises: Step 21, calling a first large model to process core dimension data based on dimension attributes of hierarchical risk type decision demand topology conformation, and generating a basic decision result instruction cluster; Step 22, invoking a second large model, and executing risk assessment and decision optimization by combining the basic decision result instruction cluster, the secondary dimension data and the decision risk matrix to generate a preliminary refined decision fit code table; And step 23, eliminating redundant deduction links in the preliminary fine decision adaptive code table through a decision chain pruning operation framework, and generating a risk controllable fine decision intermediate feature code table based on a pruned decision link and a result.
  6. 6. The hybrid large model cascade intelligent decision method of claim 5, wherein step 21 comprises: step 211, analyzing the core dimension constraint of the hierarchical risk type decision demand topology conformation to generate a core dimension decision instruction; Step 212, inputting the core dimension decision instruction into the first large model, and executing reasoning operation to generate a basic decision result instruction cluster, wherein the instruction dimension of the basic decision result instruction cluster corresponds to the core dimension one by one.
  7. 7. The hybrid large model cascade intelligent decision method of claim 5, wherein step 23 comprises: step 231, constructing a decision chain redundancy judgment rule, marking a redundancy decision link by taking the service contribution of the decision link as a judgment basis, so as to generate a redundancy link marking feature code; step 232, eliminating redundant links in the preliminary refined decision-making adaptive code table based on the redundant link marking feature codes to obtain a pruned decision-making link and a pruned result; and 233, generating a risk-controllable refined decision intermediate feature code table based on the pruned decision link and the result.
  8. 8. The hybrid large model cascading intelligent decision method according to claim 1, wherein step3 comprises: Step 31, inputting a risk controllable type refined decision intermediate feature code table into a decision risk reciprocating check-optimization framework, comparing the decision risk reciprocating check-optimization framework with a preset risk threshold value, and identifying risk exceeding links and decision logic faults to generate a decision risk-logic check topology conformation; step 32, reversely calling a second large model to re-optimize key parameters for the risk exceeding link marked in the decision risk-logic verification topology conformation so as to generate a decision sub-adaptation code table after parameter optimization; And 33, complementing the deduction link for the decision logic fault marked in the decision risk-logic verification topology conformation through a decision chain pruning operation framework, and generating a final intelligent decision correlation topology graph based on the decision sub-adaptation code table and the complementing link after parameter optimization.
  9. 9. The hybrid large model cascading intelligent decision method according to claim 8, wherein step 31 comprises: Step 311, extracting risk index data of a risk controllable type refined decision intermediate feature code table, comparing the risk index data with a risk threshold in a hierarchical risk type decision demand topology conformation, and identifying a risk exceeding link to generate a risk exceeding link feature code; step 312, checking the logical continuity of the decision link, and marking the logical fault nodes to generate a logical fault node association matrix; and step 313, generating a decision risk-logic check topology conformation based on the risk exceeding link feature codes and the logic fault node incidence matrix.
  10. 10. The hybrid large model cascading intelligent decision method according to claim 1, wherein step 4 comprises: step 41, acquiring target achievement degree, execution efficiency, risk control effect and decision logic adaptation degree data of a final intelligent decision-making association topological graph, and generating a decision execution effect feedback adaptation code table; Step 42, inputting the feedback adaptive code table of the decision execution effect into a cascade model weight-decision chain double-optimization adaptive mechanism, adjusting the calling priority and risk weight distribution of the first large model and the second large model, and updating the pruning rules of the decision chain pruning operation frame; And 43, updating a decision risk matrix of a hierarchical risk type decision demand topology conformation based on the risk related data of the decision execution effect feedback adaptation code table, precipitating a successful decision case as an adaptation template, and storing the adaptation template into a decision feature code knowledge base to finish iterative tuning of decision capability.

Description

Mixed large model cascading intelligent decision method Technical Field The application relates to the technical field of artificial intelligence decision making, in particular to a mixed large model cascading intelligent decision making method. Background In complex decision-making scenes such as financial investment combination optimization, large-scale engineering project management, urban intelligent traffic scheduling and the like, the decision-making process often involves multi-objective appeal, multiple constraint conditions and high uncertainty risk factors, and strict requirements are put on the accuracy, risk controllability and execution efficiency of the decision-making scheme. Under such a scene, decision dimensions are complicated and related, core target achievement, secondary index balance and potential risk avoidance are required to be considered, and the traditional decision technology has difficulty in meeting the intelligent decision requirement of a full link. In the prior art, all decision dimension data are processed through a unified model architecture, a fixed model calling strategy is adopted to execute reasoning operation, and after a decision result is generated based on a preset flow template, only a single-dimension risk index is used for carrying out post evaluation. The scheme does not carry out dimension division on decision requirements, all decision links are completed by relying on the same performance model, and a dynamic reasoning path optimization and logic verification mechanism is not set. The technical scheme has the remarkable defects that the layering processing of decision dimensions and a risk association modeling mechanism are not established, the risk conduction relation among different dimensions cannot be quantified, so that a decision result is easy to be locally optimal rather than globally optimal due to dimension coupling conflict, meanwhile, a fixed model calling and reasoning process lacks targeted optimization, not only is caused by the waste of computational resources or the insufficient precision of key links, but also redundant links and logic faults in a decision link are difficult to identify, continuous iteration of decision capability cannot be realized through execution effect feedback, and finally, the decision risk is out of control, the efficiency is low and the adaptability is insufficient. Disclosure of Invention In order to solve the technical problems, the application provides a mixed large model cascading intelligent decision method for at least alleviating the technical problems. The technical scheme provided by the embodiment of the application is as follows: A mixed large model cascading intelligent decision making method comprises the steps of 1, receiving demand data of a complex decision scene, disassembling decision dimension through a decision demand layering-risk modeling framework, defining accuracy and risk tolerance, constructing a decision risk matrix to mark a dimension risk association relationship, generating a layering risk type decision demand topological conformation, 2, calling a first large model and a second large model to execute cascading operation by adopting a large model cascading-risk guiding matching framework based on the layering risk type decision demand topological conformation, eliminating redundant links through a decision chain pruning operation framework, generating a risk controllable type refined decision middle feature code table, 3, evaluating risk threshold suitability through a decision risk reciprocating checking-optimizing framework based on the risk controllable type refined decision middle feature code table, reversely optimizing an exceeding link, fully deducing a logic fault, generating a final intelligent decision association topological graph, 4, collecting execution effect data of the final intelligent decision association topological graph, adjusting framework parameters and pruning rules through a cascading model weight-decision chain double optimization adaptation mechanism, updating the decision matrix, precipitating a decision template to a decision feature knowledge base, and completing iterative optimization of decision template capability. The technical scheme provided by the application has the following technical advantages: Firstly, dismantling decision dimensions through a decision requirement layering-risk modeling framework, defining precision and risk tolerance, constructing a decision risk matrix marking dimension risk association relationship, and generating a layering risk type decision requirement topological conformation. The conventional technology does not perform structural layering on decision requirements, cannot distinguish between cores and secondary dimensions, and also is difficult to quantify risk correlation between dimensions, so that decision is made on one hand. According to the method, the complex requirements are converted into the structured topol