Information domain and cognitive domain interaction model construction method and system

Abstract

The application relates to the fields of complex system modeling and cross-domain intelligent decision-making, and discloses a method and system for constructing an information domain and a cognitive domain interaction model. The method constructs a cross-domain heterogeneous graph, aggregates cross-domain evidence to obtain a cognitive posterior, a posterior standard deviation and a lower confidence limit, determines an affected subgraph according to an effective belief increment, a path support degree and the cognitive lower confidence limit, and only executes local edge weight repair on related edges in a subsegment of an original propagation path corresponding to a belief dimension with an effective belief change. In a belief-driven actionable set, multi-agent joint actions are optimized, belief-driven feasible region projection is performed on a state transition result, and feedback gain retraction is executed when a Lyapunov descent condition is not met. The method can improve cross-domain deduction accuracy, cognitive consistency, local repair efficiency and closed-loop stability.

Description

Technical Field

[0001] This invention relates to the fields of complex system modeling, cross-domain intelligent decision-making, and closed-loop control, and particularly to a method and system for constructing an interactive model between the information domain and the cognitive domain. More specifically, it relates to techniques such as cross-domain heterogeneous graph modeling, cognitive posterior estimation, belief-driven edge weight update, multi-agent joint optimization, state feasible domain projection, and closed-loop stability constraints. Background Technology

[0002] In scenarios such as information warfare, situational simulation, complex network collaborative control, and cognitive decision analysis, there is often a significant coupling relationship between the evolution of states in the information domain and the changes in beliefs in the cognitive domain. The propagation structure, influence paths, state perturbations, and policy behaviors in the information domain not only affect the formation and updating of beliefs in the cognitive domain, but also, in turn, influence the allocation of edge weights, action selection, and state transitions in the information domain, based on risk preferences, credibility judgments, and consistency levels. Therefore, how to construct a unified model that can simultaneously characterize the bidirectional interaction between the information and cognitive domains has become a key technical problem in complex system modeling.

[0003] In existing technologies, the following methods are commonly used to address this type of problem: one type models the information propagation process as a static or dynamic graph and uses graph neural networks, graph attention networks, etc., to learn representations of node and edge relationships; another type performs Bayesian updates, probabilistic inference, or sequence modeling on cognitive states, belief states, or public sentiment; and yet another type uses reinforcement learning or multi-agent reinforcement learning to optimize policy behavior. However, existing technologies generally have the following shortcomings:

[0004] First, information domain modeling and cognitive domain modeling are often separated. The results of the information domain's propagation structure update are difficult to directly constrain the evolution of cognitive variables, and the results of the cognitive variable update are also difficult to be fed back to the information domain state transition process in a calculable and executable manner, thus making it difficult to form a truly meaningful two-way closed loop.

[0005] Secondly, even if existing technologies introduce dynamic edge weight update mechanisms, they are mostly based on unified updates of the entire graph or conventional updates based on local features. They lack edge weight repair mechanisms directly driven by cognitive credibility, especially lacking a structured coupling mechanism that "first judges whether the belief is credible, then limits the scope of action, and then performs local repair." This makes edge weight updates susceptible to noise interference and computationally expensive.

[0006] Third, in the strategy optimization stage, existing technologies often only use cognitive variables as ordinary state inputs or reward references, lacking a mechanism to directly transform cognitive posterior into action set constraints and state feasible domain constraints. This results in cognitive feedback remaining at the soft coupling level, making it difficult to form a hard coupling closed loop with engineering constraints.

[0007] Fourth, existing technologies in closed-loop modeling typically lack a unified stability criterion for the entire process of edge weight update, action generation, and state projection, making it difficult to guarantee that the system will maintain convergence, interpretability, and controllability after cognitive feedback enhancement.

[0008] Therefore, there is an urgent need for a new method and system for constructing an interaction model between the information domain and the cognitive domain, in order to solve the problems of weak cognitive feedback, insufficient local repair capabilities, disconnect between action and state constraints, and difficulty in ensuring closed-loop stability in existing technologies. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for constructing an interaction model between the information domain and the cognitive domain, so as to solve the problems in the prior art such as the loose coupling between the information domain and the cognitive domain, the difficulty in directly applying cognitive beliefs to edge weight updates and state transitions, the lack of belief-driven hard constraints in policy optimization, and the insufficient stability of closed loops.

[0010] A further objective of this invention is to construct a data coupling chain of "cognitive posterior—lower confidence bound admission—affected subgraph determination—local edge weight repair—belief-driven action constraint—belief-driven state projection—Lyapunov shrinkage stabilization," so that cognitive beliefs are no longer just auxiliary information participating in modeling, but rather serve as the core driving force determining the edge weight update range, action feasibility, state feasibility, and feedback gain adjustment, thereby improving the accuracy, stability, real-time performance, and engineering feasibility of the interactive model.

[0011] To achieve the aforementioned objective, this invention provides a method for constructing an interaction model between the information domain and the cognitive domain, comprising: constructing a cross-domain heterogeneous graph based on information domain nodes, cognitive domain nodes, and edge types; aggregating cross-domain evidence based on the neighborhood of cognitive domain nodes to obtain a cognitive posterior vector, a posterior standard deviation vector, a cognitive lower confidence bound vector, and an effective belief increment vector; determining the affected subgraph based on the effective belief increment vector, path support, the cognitive lower confidence bound vector, and the correlation indicator between edges and belief dimensions; performing belief-driven local edge weight repair only on edges within the original propagation path segments corresponding to the belief dimensions that have undergone effective belief changes in the affected subgraph to obtain an updated edge weight vector; and within the action set defined by the cognitive lower confidence bound vector, performing edge weight repair based on the cognitive posterior vector and the updated edge weight vector relative to the edge weight before the update. The changes in vectors and the action default degree are used to construct a joint reward, and the multi-agent joint action is optimized to obtain the joint action. The updated edge weight vector and the multi-agent joint action are input into the state transition model to obtain the pre-projection predicted state. The pre-projection predicted state is then projected into the state feasible region defined by the cognitive lower confidence bound vector to obtain the next time-information domain state vector. The cognitive lower confidence bound vector is determined as the next time-information belief vector. Based on the next time-information domain state vector, the next time-information belief vector, and the edge weight changes, the next time-information joint Lyapunov function value is constructed. When the descent condition is not met, the feedback gain is reduced. The local edge weight repair, the multi-agent joint action optimization, and the state feasible region projection are only repeated on the affected subgraph until the descent condition is met, and the closed-loop interaction model is output.

[0012] To achieve the aforementioned objective, this invention also provides a system for constructing an interaction model between the information domain and the cognitive domain, comprising: a cross-domain graph construction module for constructing a cross-domain heterogeneous graph based on information domain nodes, cognitive domain nodes, and edge types; an evidence update module for aggregating cross-domain evidence based on the neighborhood of cognitive domain nodes, and outputting a cognitive posterior vector, a posterior standard deviation vector, a cognitive lower confidence bound vector, and an effective belief increment vector; an affected subgraph determination module for determining the affected subgraph based on the effective belief increment vector, path support, the cognitive lower confidence bound vector, and an association indicator between edges and belief dimensions; a local edge weight repair module for performing belief-driven local edge weight repair only on edges within the original propagation path segments corresponding to the belief dimensions that have undergone effective belief changes in the affected subgraph, and outputting an updated edge weight vector; a policy optimization module for constructing a joint reward within the action set defined by the cognitive lower confidence bound vector, based on the cognitive posterior vector, the change of the updated edge weight vector relative to the original edge weight vector, and the action default degree, and optimizing to obtain a multi-agent joint action; and a state projection module. The system is used to input the updated edge weight vector and the multi-agent joint action into the state transition model to obtain the pre-projection predicted state, and project the pre-projection predicted state into the state feasible domain defined by the cognitive lower confidence bound vector, and output the information domain state vector of the next time step; the stability shrinkage module is used to determine the cognitive lower confidence bound vector as the cognitive belief vector of the next time step, and construct the joint Lyapunov function value of the next time step based on the information domain state vector of the next time step, the cognitive belief vector of the next time step, and the edge weight change, reduce the feedback gain when the descent condition is not met, and trigger the local edge weight repair module, the policy optimization module, and the state projection module to be executed repeatedly until the descent condition is met; wherein, the output of the evidence update module is connected to the affected subgraph determination module, the policy optimization module, the state projection module, and the stability shrinkage module respectively, the output of the local edge weight repair module is connected to the policy optimization module, the state projection module, and the stability shrinkage module respectively, and the output of the state projection module is connected to the stability shrinkage module.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects:

[0014] 1. Form a hard-coupled closed loop driven by cognitive credibility.

[0015] This invention does not treat cognitive variables merely as ordinary feature inputs into the model. Instead, it first calculates the cognitive posterior and posterior standard deviation, then constructs a lower confidence bound, and uses this lower confidence bound as the criterion for determining whether a belief should be included in the feedback chain. This transforms uncertain cognitive quantities into calculable, determinable, and executable closed-loop control variables. Compared to the problems of low participation and weak feedback effects of cognitive variables in existing technologies, this invention significantly enhances the actual constraint capability of the cognitive domain on the information domain.

[0016] 2. Implement local edge weight repair within the affected subgraph to reduce overall graph disturbance.

[0017] This invention does not perform a uniform update on the entire cross-domain graph. Instead, it determines the affected subgraphs based on the effective belief increment and path support, and only performs edge weight repair on the relevant edges within the affected subgraphs. This not only limits the scope of cognitive change to local regions with causal or propagational support relationships, but also reduces fluctuations in irrelevant edges and unnecessary computation, thereby improving the model's real-time performance, interpretability, and robustness.

[0018] 3. Directly transform cognitive beliefs into action feasibility constraints.

[0019] In existing technologies, reinforcement learning or multi-agent optimization typically only uses cognitive states as reward modifiers or state inputs. This invention further constructs a belief-driven action set, ensuring that joint actions by multiple agents are generated within the action space defined by cognitive beliefs. This elevates cognitive feedback from a "soft influence" to a "hard constraint," avoiding action outputs that conflict with cognitive credibility and improving the consistency and feasibility of decision-making results.

[0020] 4. Directly transform cognitive beliefs into feasible domain constraints.

[0021] This invention projects a feasible region onto the state transition results based on a cognitive lower confidence bound, thereby directly constraining the state evolution process in the information domain to the cognitive confidence range. Compared to existing technologies where state updates and cognitive feedback are disconnected, this invention enables the state transition process to synchronously conform to cognitive constraints, improving the physical rationality and engineering constraint consistency of cross-domain coupled modeling.

[0022] 5. Achieve unified optimization of cognitive consistency, structural change, and strategic behavior through joint rewards.

[0023] The joint reward constructed in this invention not only considers task rewards but also cognitive consistency constraints, action violation penalties, and edge weight change penalties, thereby unifying edge weight structure changes, cognitive distribution changes, and multi-agent behavioral changes into a single optimization framework. Compared to the existing technology where edge weight updates, cognitive updates, and policy updates are separated, this invention exhibits stronger overall synergy.

[0024] 6. Improve closed-loop stability through Lyapunov pullback mechanism.

[0025] This invention does not merely achieve one-time feedback coupling, but rather automatically reduces the feedback gain and re-executes local edge weight repair, policy optimization, and state projection when the joint Lyapunov function fails to meet the descent condition, thus enabling the closed-loop process to have adaptive shrinkage capability. This can suppress system oscillations caused by excessive cognitive feedback and improve the model's convergence, controllability, and long-term stability.

[0026] 7. It combines interpretability and engineering applicability.

[0027] In this invention, the cognitive posterior, lower confidence bound, affected subgraph, local edge weight repair, belief-driven action set, belief-driven feasible region, and Lyapunov shrinkage condition are all computable, traceable, and verifiable intermediate variables or constraints. Therefore, this method not only facilitates the explanation of how cognitive beliefs affect the information domain propagation structure, action selection, and state evolution process, but also facilitates modular deployment, making it suitable for applications such as complex network situational simulation, cognitive adversarial analysis, cross-domain decision support, and intelligent control. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method for constructing an interaction model between the information domain and the cognitive domain provided by the present invention.

[0029] Figure 2 This is a block diagram of the information domain and cognitive domain interaction model construction system provided by the present invention. Detailed Implementation

[0030] To enable those skilled in the art to more clearly understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0031] In this specific implementation, the edge is uniformly written as , indicating a connected node With nodes The edges; edge-related parameters are uniformly written as , , , The same letters represent the same meaning, and different letters represent different meanings; the same variable uses the same font in different formulas; the same terminology represents the same technical feature, and different terms represent different technical features.

[0032] In this embodiment, information domain nodes can be information source nodes, propagation channel nodes, control nodes, or status monitoring nodes; cognitive domain nodes can be decision-making entity nodes, group cognitive nodes, risk assessment nodes, or cognitive feedback nodes. The information domain and cognitive domain are coupled and modeled through cross-domain influence edges. For ease of description, For the set of information domain nodes, For the set of cognitive domain nodes, To unify the set of nodes, For a moment The set of edges, It is a set of edge types.

[0033] First Embodiment

[0034] Figure 1 This is a flowchart of the method for constructing an interaction model between the information domain and the cognitive domain provided by this invention. For example... Figure 1 As shown in the figure, the method for constructing the interaction model between the information domain and the cognitive domain provided in this embodiment includes the following steps.

[0035] Step 1: Construction of Cross-Domain Heterogeneous Graphs and Organization of Original Features

[0036] First, raw observation data from the information and cognitive domains are collected and organized into a cross-domain heterogeneous graph. At time... Construct a cross-domain heterogeneous graph: , In the formula, For a moment Cross-domain heterogeneous graph; For a unified set of nodes; For a moment The set of edges; Let be a set of edge types; where, , For the set of information domain nodes, It is a set of nodes in the cognitive domain.

[0037] For any node Construct node observation feature vectors: , In the formula, For nodes At any moment The observed feature vector; For nodes At any moment The Each feature component; ; For observation feature dimensions; This indicates transpose.

[0038] As one implementation method, the information domain node characteristics include one or more of the following: propagation strength, link load, anomaly rate, resource utilization rate, credibility score, latency or interference indicators; the cognitive domain node characteristics include one or more of the following: risk perception value, tendency score, consistency score, credibility judgment value, emotion score or historical belief value.

[0039] Based on the edge set Form an adjacency matrix and edge weight vector : , , In the formula, For a moment adjacency matrix Located in the middle Line number Column elements; For connecting nodes With nodes The edge; For a moment The edge weight vector; For a moment side The right to the side.

[0040] To form a unified node representation required for subsequent evidence aggregation, the nodes... At any moment Observation feature vector Perform embedding encoding to obtain nodes At any moment Embedded vector : , In the formula, For nodes At any moment The embedding vector; It is a non-linear activation function; For nodes The node field type; For node field type The corresponding feature mapping matrix; For nodes The set of neighboring nodes; For nodes At any moment The observed feature vector; edge type The corresponding neighborhood mapping matrix; For the edge Edge type.

[0041] The output of this step is a cross-domain heterogeneous graph. Adjacency matrix Edge weight vector and node embedding vectors Among them, node embedding vectors The adjacency matrix serves as the input for subsequent cross-domain evidence aggregation. and edge weight vector This serves as the data foundation for subsequent path support calculation, local edge weight repair, state transition prediction, and stability assessment.

[0042] This step constructs a cross-domain heterogeneous graph containing information domain nodes, cognitive domain nodes, and edge types, and embeds and encodes the node observation features to form a unified node representation, adjacency relationship representation, and edge weight representation. This provides a consistent data foundation for subsequent cross-domain evidence aggregation, affected subgraph determination, local edge weight repair, action constraint construction, state feasible domain projection, and closed-loop stability determination.

[0043] Step 2: Cross-domain evidence aggregation, cognitive posterior estimation, and lower confidence bound calculation

[0044] In obtaining cross-domain heterogeneous graphs After embedding the vectors of each node, cross-domain evidence aggregation is performed using the embedding vectors of evidence nodes related to the cognitive domain as input to obtain the time-domain evidence. Cross-domain evidence vector : , , In the formula, For a moment Cross-domain evidence vectors; For a moment A set of evidence nodes related to the cognitive domain; For a moment node Evidence aggregation weights; For nodes At any moment The embedding vector; For nodes At any moment The embedding vector; For evidence scoring vectors; It is an exponential function.

[0045] Combination Time Current cognitive belief vector Cross-domain evidence vectors and public opinion sentiment vector The estimated time Cognitive posterior vector and cognitive posterior covariance matrix : , , In the formula, For a moment The cognitive posterior vector; For a moment The current cognitive belief vector; The evidence mapping matrix; The emotional coupling coefficient; For a moment The public sentiment vector; For a moment The cognitive posterior covariance matrix; For covariance estimation network; It is an S-shaped compression mapping function; It is a logarithmic probability transformation function; For diagonalization operators; It is a positive activation function.

[0046] Based on the cognitive posterior covariance matrix Get the time posterior standard deviation vector : , In the formula, For a moment The posterior standard deviation vector; Covariance matrix The diagonal vector; This means taking the square root of each vector component.

[0047] According to cognitive posterior vector and posterior standard deviation vector Construction time The cognitive lower confidence bound vector And further obtain the moment Effective belief increment vector : , , In the formula, For a moment The cognitive confidence bound vector; Confidence coefficient; For a moment The effective belief increment vector.

[0048] In one implementation, the cognitive posterior covariance matrix The output of the covariance estimation network enables a simultaneous characterization of cognitive posterior uncertainty. Therefore, cognitive beliefs are no longer represented solely by the posterior mean, but are further characterized by the cognitive lower confidence bound vector. This generates a conservative, stable feedback driving force that can be used for subsequent constraints.

[0049] The output of this step is the cognitive posterior vector. posterior standard deviation vector Cognitive lower confidence bound vector and effective belief increment vector Among them, the cognitive lower confidence bound vector and effective belief increment vector Simultaneously, it serves as input for subsequent determination of affected subgraphs, repair of local edge weights, construction of action constraints, projection of state feasible regions, and stability judgment, thereby forming a data coupling chain with cognitive credibility and changes in effective beliefs as the core driving forces.

[0050] Step 3: Calculation of path support and determination of affected subgraphs

[0051] In obtaining an effective belief increment vector and cognitive lower confidence bound vector Then, for edges in cross-domain heterogeneous graphs The support of the original propagation path is calculated to obtain the time step. side Path support : , In the formula, For a moment side Path support; For connecting nodes With nodes The set of original propagation paths; For the original propagation path set One of the original propagation paths in it; For path upper connection node With nodes The edge; For a moment side The right of the border; This is for calculating the maximum value.

[0052] Furthermore, based on path support Effective Belief Increment Vector Cognitive lower confidence bound vector and the edge and the first Indicators of correlation between beliefs Determine the time The set of edges of the affected subgraph : , In the formula, For a moment The set of edges of the affected subgraph; Path support threshold; For effective belief increment vector The Dimensional components; For the first Valid belief increment threshold; For cognitive lower confidence bound vector The Dimensional components; For the first Belief admission threshold; For the edge With the Indicators of the correlation between beliefs.

[0053] In one implementation, the associated indicator quantity Determined based on a predefined mapping relationship between edge type and cognitive dimension; when edge The type of the edge belongs to participates in the first When the chain of cognitive belief is established, let Otherwise, let .

[0054] Using the above method, only edges that simultaneously satisfy the following conditions—path support reaching a threshold, effective belief increment reaching a threshold, cognitive lower confidence bound reaching a belief admission threshold, and having a correlation mapping relationship with the corresponding belief dimension—are included in the affected subgraph edge set. Therefore, this step does not determine the local range solely based on the adjacency relationship of the graph structure, but rather uses propagation path support, belief change magnitude, cognitive credibility, and the correlation between edge and belief dimensions as screening criteria, thereby avoiding unnecessary edge weight updates due to low-credibility cognitive changes or irrelevant propagation paths.

[0055] The output of this step is the set of edges in the affected subgraph. The set of edges of the affected subgraph As the scope of subsequent local edge weight repair, the impact of current cognitive changes is limited to a set of local edges that have propagation support relationships, effective belief change foundations, and belief dimension correlation relationships.

[0056] Step 4: Belief-Driven Local Boundary Weight Repair

[0057] Determine the set of edges of the affected subgraph Subsequently, belief-driven local edge weight repair is performed only on edges within the original propagation path segments in the affected subgraph that correspond to the belief dimension that has undergone effective belief change.

[0058] Specifically, the belief dimension for which effective belief change occurs refers to the dimension that satisfies: ,

[0059] The The belief dimension. The original propagation path segment corresponding to the belief dimension that undergoes effective belief change refers to the segment in the original propagation path that satisfies... and The A continuous edge segment with an associated mapping relationship to a belief dimension. For edges that are in the same connected region but do not form an associated mapping relationship with a belief dimension that has undergone a valid belief change, edge weight repair is not performed.

[0060] For the set of edges belonging to the affected subgraph And the edges located within the original propagation path sub-segment The time can be obtained through the following formula. After the update, the rights : , In the formula, For a moment side The rights to update afterwards; edge type Corresponding edge weight interval The projection operator; For the edge The type of edge it belongs to; edge type The lower bound of the boundary weight; edge type The upper limit of the boundary rights; edge type The memory coefficient; For a moment side The baseline edge weight; edge type Feedback gain; For the edge Relational symbol factors; edge type The belief mapping vector; For vectors with vector The inner product of.

[0061] For edges that do not belong to the affected subgraph edge set Edges, or edges that, although in the same connected region, are not located within the original propagation path segment corresponding to the belief dimension where a valid belief change occurred, retain their edge weights unchanged: ,

[0062] In one implementation, when When, it indicates the edge For the corresponding belief dimension, it is a facilitating relationship; when When, it indicates the edge The corresponding belief dimension has an inhibitory relationship. Therefore, changes in beliefs not only affect the magnitude of edge weight updates, but also determine the direction of edge weight repair based on the promoting or inhibiting relationship between edges and beliefs.

[0063] Set of edges Update the rights of each side Update the preceding weight vector. The same edge indices are arranged in order to form the time. Update the edge weight vector : , In the formula, For a moment The updated edge weight vector, and and Arranged in order of the same edge index.

[0064] The output of this step is the updated edge weight vector. The updated edge weight vector On the one hand, it serves as the input for the subsequent joint reward construction; on the other hand, it serves as the input for subsequent state transition prediction and stability judgment, thus forming a vector of effective belief increments. Driven local structural repair link.

[0065] This step involves performing edge weight repair only on edges within the affected subgraph and within the original propagation path segments corresponding to the belief dimensions that have undergone effective belief changes. By imposing edge weight interval projection constraints on the repair results, this step enables cognitive belief changes to act on the propagation structure update process in a controlled manner, while maintaining the edge weight stability of unaffected edges and avoiding perturbations to the entire graph and unbounded changes in edge weights.

[0066] Step 5: Construction of Belief-Driven Action Sets and Optimization of Multi-Agent Joint Actions

[0067] Obtaining the cognitive confidence bound vector And the updated edge weight vector output from step four Then, construct the cognitive lower confidence bound vector. Limited action set : , In the formula, For the confidence bound vector of cognition Limited action set; For joint action vectors; For a moment The information domain state vector; For the state vector in the information domain Lower joint action vector For the The constraint evaluation function for each action constraint; For the first A belief-constrained weight vector; Number of action constraints.

[0068] To quantify moments Multi-agent joint action Deviation from action set The degree of breach of contract is defined by the degree of breach of contract. : , In the formula, For joint operations Compared to action set The degree of breach of contract; For the state vector in the information domain Multi-agent joint action For the The constraint evaluation function for each action constraint; For finding the maximum value operator; This indicates transpose.

[0069] According to cognitive posterior vector Get the time Cognitive distribution vector : , In the formula, For a moment The cognitive distribution vector; This is a cognitive distribution mapping matrix; This is the normalized exponential mapping function.

[0070] In one implementation, the policy network uses the current information domain state vector Cognitive posterior vector Cognitive lower confidence bound vector and the updated edge weight vector As input, first output the time. Candidate joint action Then, the candidate joint action Projected onto the action set Inside, the moment is obtained Multi-agent joint action : , In the formula, For a moment Candidate joint actions; For multi-agent joint actions after projection of the actionable set; To make actionable collection The projection operator.

[0071] Based on task rewards, the deviation between the pre-projection predicted state and the target state, cognitive consistency, action default degree, and edge weight changes, time intervals are constructed. Joint Awards : , In the formula, For a moment Joint rewards; For a moment Task rewards; The weight is the state deviation weight; For a moment Pre-projection predicted state; For a moment The target state; For cognitive consistency weight; Kullback-Leibler divergence; This is the mapping matrix from state to cognitive distribution; Weighting of penalties for breach of contract; Penalty weight for changes in edge weight; It is the Euclidean norm; It is the Frobenius norm.

[0072] in, and Arranged in the same edge index order, therefore the vector difference Representation of edge set The changes in edge weights of the corresponding components of each edge before and after the update; for edges that do not belong to the set of edges in the affected subgraph. Or an edge that is not located within the corresponding original propagation path segment, because it satisfies Therefore, its vector difference The corresponding component in is zero.

[0073] The output of this step is the multi-agent joint action. The multi-agent joint action The updated edge weight vector output from step four These are used together as inputs for subsequent state transition prediction and feasible region projection. Candidate joint actions are then used... Projection is a multi-agent joint action that satisfies belief-driven constraints. It can elevate cognitive feedback from ordinary state input or reward reference to a hard constraint in the action generation process, thereby improving the consistency, compliance and feasibility of joint decision-making.

[0074] Step Six: Belief-Driven State Transition Prediction and Feasible Region Projection

[0075] After obtaining the updated edge weight vector from the output of step four The multi-agent joint action output in step five Then, the subsequent weight vector will be updated. Multi-agent joint action The common input state transition model yields the time step. Pre-projection prediction state It should be noted that this step uses multi-agent joint actions projected from the action set. Instead of candidate joint actions directly output by the policy network. .

[0076] To characterize the coupling effect of updated edge weights on state evolution, time intervals are constructed. coupling matrix and belief-driven items : , , In the formula, For a moment The coupling matrix; For a moment The adjacency matrix; To update the edge weight vector The diagonal matrix formed; This is a mapping matrix from edge type to state dimension; For the confidence bound vector of cognition Generated belief-driven terms; Let be the mapping matrix from belief to state perturbation.

[0077] According to the state transition model Coupling matrix and belief-driven items Get the time Pre-projection prediction state : , In the formula, For a moment Pre-projection predicted state; For parameters State transition model; These are the parameters of the state transition model; For a moment The information domain state vector; For multi-agent joint actions after projection of the action set.

[0078] Based on the cognitive lower confidence bound vector Constructing the feasible region of the state : , In the formula, For the confidence bound vector of cognition The feasible domain of the defined state; This is the information domain state vector; For the first A state safety function; For the first A belief constraint vector; This represents the number of state constraints.

[0079] Predicted state before projection Execution to the feasible domain The projection of the time is obtained. Information domain state vector : , In the formula, For a moment The information domain state vector; To reach the feasible region The projection operator.

[0080] In one implementation, the projection of the state feasible region can be equivalently written as the following quadratic programming process: , In the formula, This represents the values ​​of the variables that minimize the objective function; the quadratic programming problem uses the pre-projection predicted state as an example. Using the reference point, with the state feasible region For the constraint set, the projected information domain state vector While satisfying cognitive constraints and state safety constraints, it should be as close as possible to the pre-projection predicted state. .

[0081] The output of this step is the time. Information domain state vector The information domain state vector On the one hand, it serves as the input for subsequent stability assessment; on the other hand, it serves as the state input for the next round of closed-loop update, thus forming the updated edge weight vector. Multi-agent joint action and cognitive lower confidence bound vector A jointly driven state evolution chain.

[0082] Step 7: Combine Lyapunov stability criterion with feedback gain reduction

[0083] The information domain state vector at the next moment after obtaining the output of step six. Then, the cognitive lower confidence bound vector Determined as the cognitive belief vector for the next moment : , In the formula, For a moment The next moment's cognitive belief vector. This is achieved by using the cognitive lower confidence bound vector. As the cognitive belief vector of the next moment This allows closed-loop stability judgment to be based on conservative and reliable cognitive feedback, avoiding the overly strong feedback caused by directly using the cognitive posterior mean which has high uncertainty.

[0084] Construction time The joint Lyapunov function value : , And construct time The joint Lyapunov function value : , In the formula, For a moment The combined Lyapunov function value; For a moment The combined Lyapunov function value; The target information domain state vector; For vector difference Regarding the semi-positive definite weight matrix The weighted quadratic form; It is a positive semi-definite weight matrix; Weights for cognitive belief bias terms; For the target cognitive belief vector; The weights are the change terms of the edge weights; For a moment The edge weight vector; For a moment The edge weight vector; For a moment Update the edge weight vector; It is the Euclidean norm; It is the Frobenius norm.

[0085] Further constructing time joint state vector Joint state vector with target : , , In the formula, For a moment The joint state vector; Let be the joint state vector of the objective.

[0086] Based on the joint Lyapunov function value , combined Lyapunov function values Joint state vector Joint state vector with target Determine whether the current closed-loop update meets the descent condition: , In the formula, This is the decrease coefficient.

[0087] When the aforementioned descent condition is met, it indicates that the current closed-loop update satisfies stability constraints in terms of state bias, cognitive belief bias, and edge weight changes. In this case, the current update result is output as a valid update result of the closed-loop interaction model. When the aforementioned descent condition is not met, it indicates that the current feedback gain may be too large or local edge weight repair may lead to an overly strong closed-loop response. In this case, the edge type is adjusted according to the following formula. The feedback gain is reduced: , In the formula, For a moment edge type Feedback gain; For a moment edge type Feedback gain; The shrinkage coefficient is, and .

[0088] After performing feedback gain rollback, only the set of edges in the affected subgraph is processed. Step four is re-executed for edges within the original propagation path segment corresponding to the belief dimension where effective belief changes occur; step five is re-executed for the multi-agent joint action; step six is ​​re-executed for the projection of the state feasible region; and the joint Lyapunov function value is recalculated. This continues until the descent condition is met. During this process, the entire graph is not reconstructed, and unaffected edges are not repeatedly updated, thereby reducing the computational cost of the closed-loop shrinkage process.

[0089] The output of this step is the final updated edge weight vector that meets the stability requirements. Multi-agent joint action The state vector of the information domain at the next time step and the cognitive belief vector of the next moment Based on this, a closed-loop interaction model is formed.

[0090] By constructing a joint Lyapunov function and reducing the feedback gain when the descent condition is not met, while re-performing local edge weight repair, action optimization, and state projection only within the affected subgraph, a unified stability criterion can be applied to the closed-loop update process. This allows for the automatic suppression of feedback intensity when cognitive feedback is too strong, edge weight updates are too large, or the state response carries the risk of oscillation, preventing system instability caused by amplified local updates. Furthermore, local recalculation can be completed without reconstructing the entire graph, improving the convergence, stability, and computational efficiency of the closed-loop process.

[0091] Second Embodiment

[0092] Based on the method of the first embodiment described above, such as Figure 2 As shown, this embodiment further provides a system for constructing an interaction model between the information domain and the cognitive domain. The system includes a processor, a storage medium, and a computer program stored in the storage medium and executable on the processor. When executed by the processor, the computer program implements the following functional modules.

[0093] The system includes a cross-domain graph construction module, an evidence update module, an affected subgraph determination module, a local edge weight repair module, a strategy optimization module, a state projection module, and a stability shrinkage module.

[0094] The cross-domain graph construction module is used to construct cross-domain heterogeneous graphs based on information domain nodes, cognitive domain nodes, and edge types, and output the cross-domain heterogeneous graph. Adjacency matrix Edge weight vector and node embedding vector The cross-domain graph construction module executes the cross-domain heterogeneous graph construction and original feature organization process described in step one of the first embodiments.

[0095] The evidence update module is used to aggregate cross-domain evidence based on the neighborhood of cognitive domain nodes and output a cognitive posterior vector. posterior standard deviation vector Cognitive lower confidence bound vector and effective belief increment vector The evidence update module performs the cross-domain evidence aggregation, cognitive posterior estimation, and lower confidence bound calculation processes described in step two of the first embodiment.

[0096] The affected subgraph determination module is used to determine the effective belief increment vector. Path support Cognitive lower confidence bound vector and the correlation indicator between the edge and the belief dimension Identify and output the set of edges of the affected subgraph. The affected subgraph determination module performs the path support calculation and affected subgraph determination process described in step three of the first embodiment.

[0097] The local edge weight repair module performs belief-driven local edge weight repair only on edges within the original propagation path segments corresponding to the belief dimensions that have undergone effective belief changes in the affected subgraph, outputting updated edge weight vectors. The local edge weight repair module executes the belief-driven local edge weight repair process described in step four of the first embodiment.

[0098] The policy optimization module is used to optimize the confidence bound vector under cognition. Limited action set Internally, based on cognitive posterior vectors Update the edge weight vector Compared to updating the previous weight vector Changes and degree of breach of contract Constructing joint rewards And optimize to obtain multi-agent joint actions The strategy optimization module executes the belief-driven action set construction and multi-agent joint action optimization process described in step five of the first embodiment.

[0099] The state projection module is used to update the weight vector of the next edge. Multi-agent joint action Input the state transition model to obtain the predicted state before projection. And the predicted state before projection Projected onto the cognitive lower confidence bound vector Limited feasible domain of states Within, output the state vector of the information domain at the next time step. The state projection module executes the belief-driven state transition prediction and state feasible region projection process described in step six of the first embodiment.

[0100] The stability shrinking module is used to shrink the cognitive lower confidence bound vector. Determined as the cognitive belief vector for the next moment And based on the state vector of the information domain at the next time step Next-moment cognitive belief vector Constructing the joint Lyapunov function value at the next time step using edge weight changes If the descent condition is not met, the feedback gain is reduced, and the local edge weight repair module, policy optimization module, and state projection module are repeatedly executed until the descent condition is met. Specifically, the stability rollback module executes the combined Lyapunov stability criterion and feedback gain rollback process described in step seven of the first embodiment.

[0101] Furthermore, the output of the evidence update module is connected to the affected subgraph determination module, the strategy optimization module, the state projection module, and the stability reduction module, respectively; the output of the local edge weight repair module is connected to the strategy optimization module, the state projection module, and the stability reduction module, respectively; the output of the state projection module is connected to the stability reduction module and fed back into the next round of closed-loop update process. Thus, the system can fully support the information domain and cognitive domain interaction model construction method of the present invention, and form a closed-loop interactive execution chain of "cross-domain graph construction—evidence update—affected subgraph determination—local edge weight repair—strategy optimization—state projection—stability reduction".

[0102] The present invention has been further described in detail above with reference to specific embodiments. It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, several equivalent transformations, substitutions or improvements can be made to the technical solutions of the present invention without departing from the concept and essence of the present invention, and all such equivalent transformations, substitutions or improvements should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for constructing an interaction model between the information domain and the cognitive domain, characterized in that, include: A cross-domain heterogeneous graph is constructed based on information domain nodes, cognitive domain nodes, and edge types. Cross-domain evidence is aggregated based on the neighborhood of cognitive domain nodes to obtain a cognitive posterior vector, a posterior standard deviation vector, a cognitive lower confidence bound vector, and an effective belief increment vector. The affected subgraph is determined based on the effective belief increment vector, path support, the cognitive lower confidence bound vector, and the correlation indicator between edges and belief dimensions. Belief-driven local edge weight repair is performed only on edges within the original propagation path segments corresponding to belief dimensions that have undergone effective belief changes in the affected subgraphs, resulting in an updated edge weight vector. Within the action set defined by the cognitive lower confidence bound vector, a joint reward is constructed based on the cognitive posterior vector, the change in the updated edge weight vector relative to the original edge weight vector, and the action default degree. The system optimizes and obtains the joint action of multiple agents; it inputs the updated edge weight vector and the joint action of multiple agents into the state transition model to obtain the predicted state before projection, and projects the predicted state before projection into the state feasible region defined by the cognitive lower confidence bound vector to obtain the information domain state vector of the next time step; it determines the cognitive lower confidence bound vector as the cognitive belief vector of the next time step, constructs the joint Lyapunov function value of the next time step based on the information domain state vector of the next time step, the cognitive belief vector of the next time step, and the edge weight change, reduces the feedback gain when the descent condition is not met, and repeats the local edge weight repair, the optimization of the joint action of multiple agents, and the projection of the state feasible region only on the affected subgraph until the descent condition is met, and outputs the closed-loop interaction model.

2. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 1, characterized in that, The method of aggregating cross-domain evidence based on the neighborhood of cognitive domain nodes to obtain the cognitive posterior vector, posterior standard deviation vector, cognitive lower confidence bound vector, and effective belief increment vector includes: obtaining the time step using the following formula. Cross-domain evidence vector : And in obtaining the cross-domain evidence vector Then, combined with the current cognitive belief vector and public sentiment vector The cognitive posterior vector is obtained through the following formula. Cognitive posterior covariance matrix posterior standard deviation vector Cognitive lower confidence bound vector and effective belief increment vector : In the formula, For a moment A set of evidence nodes related to the cognitive domain; For a moment node Evidence aggregation weights; For nodes At any moment The embedding vector; For nodes At any moment The embedding vector; For evidence scoring vectors; The evidence mapping matrix; The emotional coupling coefficient; Confidence coefficient; For covariance estimation network; For diagonalization operators; This is the diagonal vector operator; It is a positive activation function; It is a logarithmic probability transformation function; It is an S-shaped compression mapping function.

3. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 2, characterized in that, The step of determining the affected subgraph based on the effective belief increment vector, path support, cognitive lower confidence bound vector, and the correlation indicator between edges and belief dimensions includes: first, for any edge... Calculate path support : Then based on the path support Effective Belief Increment Vector Cognitive lower confidence bound vector and the edge and the first Indicators of correlation between beliefs Determine the time The set of edges of the affected subgraph : In the formula, For connecting nodes With nodes The set of original propagation paths; The original propagation path set One of the original propagation paths in it; For path upper connection node With nodes The edge; For a moment side The right of the border; Path support threshold; For effective belief increment vector The Dimensional components; For the first Valid belief increment threshold; For cognitive lower confidence bound vector The Dimensional components; For the first Belief admission threshold; For the edge With the The correlation indicator between beliefs is a value determined based on a predefined mapping relationship between edge type and cognitive dimension.

4. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 3, characterized in that, The step of performing belief-driven local edge weight repair only on edges within the original propagation path segments corresponding to the belief dimensions that have undergone effective belief changes in the affected subgraph, to obtain updated edge weight vectors, includes: [the following steps are taken for each edge set belonging to the affected subgraph]. edge The time can be obtained through the following formula. After the update, the rights : For edges that do not belong to the set of affected subgraph edges For each edge, keep its weight unchanged: and set the edges The updated edge weights are compared with the updated edge weight vectors. The edge indices are arranged in the same order to form the updated edge weight vector. : In the formula, edge type Corresponding edge weight interval The projection operator; For the edge The type of edge it belongs to; edge type The lower bound of the boundary weight; edge type The upper limit of the boundary rights; edge type The memory coefficient; For a moment side The baseline edge weight; edge type Feedback gain; For the edge Relational symbol factors; edge type The belief mapping vector; For vectors with vector The inner product of; the original propagation path segment corresponding to the belief dimension that undergoes effective belief change refers to the segment in the original propagation path that satisfies... and The A belief is a continuous edge segment with an associated mapping relationship.

5. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 4, characterized in that, Before optimizing the multi-agent joint action within the action set defined by the cognitive confidence bound vector, the process further includes constructing the action set and defining the action default degree, specifically: after obtaining the updated edge weight vector... Then, construct the cognitive lower confidence bound vector. Limited action set : And define time Multi-agent joint action Relative to the aforementioned set of actions Action breach of contract : In the formula, For joint action vectors; For a moment The information domain state vector; For the state vector in the information domain Lower joint action vector For the The constraint evaluation function for each action constraint; For the state vector in the information domain Multi-agent joint action For the The constraint evaluation function for each action constraint; For the first A belief-constrained weight vector; Number of action constraints; For finding the maximum value operator; This indicates transpose.

6. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 5, characterized in that, The process of constructing a joint reward based on the cognitive posterior vector, the change of the updated weight vector relative to the updated weight vector, and the action default, and optimizing it to obtain a multi-agent joint action, includes: first, based on the cognitive posterior vector... Get the moment Cognitive distribution vector : Policy network output time Candidate joint action and the candidate joint action Projected onto the set of movable actions Within, multi-agent joint actions are obtained. : Then, a joint reward is constructed based on task rewards, the deviation between the predicted state before projection and the target state, cognitive consistency, action default degree, and changes in edge weights. : In the formula, This is a cognitive distribution mapping matrix; This is the normalized exponential mapping function; To make actionable collection The projection operator; For a moment Task rewards; The weight is the state deviation weight; For a moment Pre-projection predicted state; For a moment The target state; For cognitive consistency weight; The Kullback-Leibler divergence; This is the mapping matrix from state to cognitive distribution; Weighting of penalties for breach of contract; Penalty weight for changes in edge weight; It is the Euclidean norm; It is the Frobenius norm.

7. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 6, characterized in that, The step of inputting the updated edge weight vector and the multi-agent joint action into the state transition model to obtain the pre-projection predicted state includes: based on the current information domain state vector Multi-agent joint action Update the edge weight vector and cognitive lower confidence bound vector First construct the time. coupling matrix and belief-driven items : The predicted state before projection is then obtained using the following formula. : In the formula, For a moment The adjacency matrix; To update the edge weight vector The diagonal matrix formed; This is a mapping matrix from edge type to state dimension; Let be the mapping matrix from belief to state perturbation; For parameters State transition model; These are the parameters for the state transition model.

8. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 7, characterized in that, The step of projecting the pre-projection predicted state onto the feasible state domain defined by the cognitive lower confidence bound vector to obtain the information domain state vector at the next time step includes: based on the cognitive lower confidence bound vector... Constructing the feasible region of the state : Then, the predicted state before projection... Execution to the feasible state domain The projection of the time is obtained. Information domain state vector : In the formula, This is the information domain state vector; For the first A state safety function; For the first A belief constraint vector; This represents the number of state constraints. To reach the feasible region The projection operator.

9. The method for constructing an interaction model between the information domain and the cognitive domain according to claim 8, characterized in that, The step of determining the cognitive lower confidence bound vector as the cognitive belief vector at the next time step, constructing the joint Lyapunov function value at the next time step based on the information domain state vector at the next time step, the cognitive belief vector at the next time step, and the change in edge weights, and reducing the feedback gain when the descent condition is not met, includes: determining the cognitive lower confidence bound vector as the cognitive belief vector at the next time step, constructing the joint Lyapunov function value at the next time step based on the information domain state vector at the next time step, the cognitive belief vector at the next time step, and reducing the feedback gain when the descent condition is not met. Determined as the cognitive belief vector for the next moment : Construction time The joint Lyapunov function value ,time The joint Lyapunov function value ,time joint state vector and the joint state vector of the target : And determine whether the current closed-loop update meets the descent condition according to the following formula: When the descent condition is not met, the edge type is determined according to the following formula. The feedback gain is reduced: And only for the affected subgraph, the local edge weight repair, the multi-agent joint action optimization, and the state feasible region projection are repeated; where, The target information domain state vector; For vector difference Regarding the semi-positive definite weight matrix The weighted quadratic form; It is a positive semi-definite weight matrix; Weights for cognitive belief bias terms; For the target cognitive belief vector; The weights are the change terms of the edge weights; For a moment The edge weight vector; For a moment The edge weight vector; For a moment The joint state vector; Let this be the joint state vector of the objective. The decreasing coefficient; For a moment edge type Feedback gain; For a moment edge type Feedback gain; The shrinkage coefficient is, and .

10. A system for constructing an interaction model between the information domain and the cognitive domain, characterized in that, include: The cross-domain graph construction module is used to construct cross-domain heterogeneous graphs based on information domain nodes, cognitive domain nodes, and edge types. The system comprises the following modules: an evidence update module, which aggregates cross-domain evidence based on the neighborhood of cognitive domain nodes and outputs a cognitive posterior vector, a posterior standard deviation vector, a cognitive lower confidence bound vector, and an effective belief increment vector; an affected subgraph determination module, which determines the affected subgraph based on the effective belief increment vector, path support, the cognitive lower confidence bound vector, and the correlation indicator between edges and belief dimensions; a local edge weight repair module, which performs belief-driven local edge weight repair only on edges within the original propagation path segments corresponding to the belief dimensions that have undergone effective belief changes in the affected subgraph and outputs an updated edge weight vector; a policy optimization module, which constructs a joint reward based on the cognitive posterior vector, the change of the updated edge weight vector relative to the original edge weight vector, and the action default degree within the action set defined by the cognitive lower confidence bound vector, and optimizes it to obtain a multi-agent joint action; and a state projection module, which inputs the updated edge weight vector and the multi-agent joint action into a state transition model to obtain a pre-projection predicted state, projects the pre-projection predicted state onto the state feasible domain defined by the cognitive lower confidence bound vector, and outputs a state vector in the information domain at the next time step. The stability reduction module is used to determine the cognitive lower confidence bound vector as the cognitive belief vector at the next time step, and construct the joint Lyapunov function value at the next time step based on the information domain state vector at the next time step, the cognitive belief vector at the next time step, and the edge weight change. When the descent condition is not met, the feedback gain is reduced, and the local edge weight repair module, the policy optimization module, and the state projection module are repeatedly executed until the descent condition is met. The output of the evidence update module is connected to the affected subgraph determination module, the policy optimization module, the state projection module, and the stability reduction module, respectively. The output of the local edge weight repair module is connected to the policy optimization module, the state projection module, and the stability reduction module, respectively. The output of the state projection module is connected to the stability reduction module.

Images