A Multi-Agent Collaborative Decision-Making Method and System Based on Dynamic Cognitive State Modeling

Abstract

This invention discloses a multi-agent collaborative decision-making method and system based on dynamic cognitive state modeling. Addressing the shortcomings of existing multi-agent collaboration methods based on large language models, which rely on unstructured text concatenation and suffer from ambiguous intent reasoning, communication redundancy, and collaborative deadlock, this invention initializes an explicit cognitive space for each agent, including its own intent vector, social cognitive matrix, and cognitive uncertainty covariance. It generates enhanced observation vectors for collaborative partners through a pre-trained large language model encoder combined with multi-head attention, dynamically updates the cognitive state using adaptive Kalman filtering, introduces a second-order belief model to detect implicit cognitive misalignments, and performs pruning-quantization hierarchical collaboration based on Shannon entropy. Finally, it ensures that natural language interaction is triggered only when there is cognitive ambiguity or misalignment. This invention improves the robustness of cognitive updates, effectively suppresses communication storms, and can be applied to complex game theory, distributed task planning, and other scenarios, significantly improving the efficiency and quality of collaborative decision-making.

Description

Technical Field

[0001] This invention relates to the field of distributed artificial intelligence and natural language processing technology, specifically to a multi-agent collaborative decision-making method and system based on dynamic cognitive state modeling. Background Technology

[0002] In recent years, large language models have become a core technology for building open-domain multi-agent systems due to their superior zero-shot reasoning, general task planning, and natural language interaction capabilities. Mainstream multi-agent frameworks generally adopt an interaction mode of natural language context concatenation + round-robin generation: agents use natural language text as a communication medium, and by concatenating historical interaction texts into a context window, they drive the large language model to generate response content, achieving task distribution, information sharing, and collaborative action.

[0003] However, when dealing with highly complex, game-theoretic, or tasks requiring deep strategic collaboration, existing technologies still have the following technical shortcomings:

[0004] (1) Existing systems rely on unstructured natural language context to maintain agent memory. Cognitive state is implicitly embedded in the text stream. It cannot accurately analyze the dynamic intentions, belief boundaries and knowledge reserves of collaborating partners. Agents can only make supplementary responses based on surface text and lack the ability to reason recursively based on "theory of mind". In complex scenarios involving information asymmetry and conflict of objectives, it is easy to misunderstand intentions and semantic drift, causing decisions to deviate from the global optimal goal.

[0005] (2) Due to the lack of explicit cognitive state modeling and behavior prediction mechanism, existing intelligent agents cannot make probabilistic predictions of the intentions of their partners. They can only repeatedly confirm information and align goals through high-frequency and redundant explicit communication, which leads to the phenomenon of "communication storm". In large-scale multi-agent scenarios, the token consumption and communication delay of invalid natural language interaction increase exponentially, and the system efficiency drops sharply with the increase in the number of intelligent agents, making it difficult to support the real-time collaboration requirements.

[0006] (3) Existing systems can only identify collaboration anomalies through explicit language conflicts, and cannot accurately capture implicit cognitive misalignments caused by information asymmetry—that is, neither party is aware of the other’s misunderstanding of the intention, and eventually falls into a non-optimal Nash equilibrium or collaboration deadlock. In scenarios such as strong games and long-term planning, the system’s decision quality and task success rate are extremely low.

[0007] Therefore, there is an urgent need for a new method that can build and maintain explicit and quantitative cognitive models for agents, so as to improve the deep collaborative reasoning ability and decision-making quality of multi-agent systems in complex and dynamic environments while minimizing communication costs. Summary of the Invention

[0008] To address the problems existing in the background technology, this invention proposes a multi-agent collaborative decision-making method and system based on dynamic cognitive state modeling. By constructing an explicit cognitive state space, an adaptive Kalman filter intention reasoning mechanism, second-order belief deadlock detection, and a hierarchical collaboration strategy, it solves core technical problems such as fuzzy agent intention reasoning, communication redundancy, and collaborative deadlock, and achieves joint optimization of communication overhead and decision quality in multi-agent collaboration.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A multi-agent collaborative decision-making method based on dynamic cognitive state modeling includes the following steps:

[0011] S1. Obtain the input task description and the set of agents participating in the collaboration, and initialize the explicit cognitive state for each agent;

[0012] S2. Load the pre-trained large language model, obtain the input data of the cooperative partners received by the agent at the current cooperative time step, encode the semantic features of the large language model, and generate an enhanced observation vector of the cooperative partner's behavior.

[0013] S3. Based on the cognitive state of the previous moment, generate the agent's prior prediction vector of the cooperative partner's behavior, compare it with the current enhanced observation vector, and obtain the observation difference degree.

[0014] S4. Based on the posterior uncertainty covariance obtained in the previous time step and the time interval between the two collaborations, calculate the current prior uncertainty covariance.

[0015] S5. Combining the current enhanced observation vector, prior prediction vector, observation difference, and prior uncertainty covariance, update the cognitive state of the collaborating partners through latent space adaptive Kalman filtering;

[0016] S6. Based on the updated cognitive state and the agent's own historical behavioral characteristics, construct a second-order belief model to detect whether there is a collaborative deadlock caused by cognitive misalignment among agents.

[0017] S7. Calculate the communication necessity entropy value, and perform pruning-quantification hierarchical cooperation based on the communication necessity entropy value and the cooperative deadlock detection results.

[0018] Specifically, step S1 is: obtaining the input task description. With the collection of intelligent agents For each intelligent agent Initialize the explicit cognitive state in its memory structure The process of initializing explicit cognitive states is as follows:

[0019] S11. Set the dimension of the intent vector as follows: The baseline intent space is generated using an orthogonal initialization method. ;

[0020] S12, For each agent , initialize it in Cognitive state at any moment ,in, For intelligent agents Its own intention vector The social cognition matrix is ​​shown, where the first... row vector Represents intelligent agents For intelligent agents Predicting the initial intent To understand the uncertain covariance matrix, it is initially set as , The initial uncertainty coefficient is preset. for 3D identity matrix.

[0021] Specifically, step S2 includes:

[0022] S21. Obtain the pre-trained large language model encoder layer Get collaboration time steps Time-based intelligent agent Received from collaborating partners Input data The input is fed into the encoder layer, where it generates the observation latent vector through semantic mapping. The specific calculation is as follows:

[0023] ;

[0024] Observation latent vector Transform unstructured natural language input into structured feature vectors aligned with the intent space;

[0025] S22. Obtain the data stored in the agent's memory structure. Global context state at time step Enhanced observation vectors are generated through an attention mechanism. The specific calculation is as follows:

[0026] ;

[0027] in, This is a normalization layer used to ensure a reasonable distribution range of feature vectors in the intent space. This represents multi-head attention operations, used to capture the semantic relevance between current observations and historical behaviors.

[0028] Specifically, step S3 includes:

[0029] S31, Obtaining the System Time-based intelligent agent For intelligent agents Cognitive state vector By mapping it to the baseline intent space through a multilayer perceptron, a set of instructions for collaborating partners is generated. Prior prediction vector of behavior ;

[0030] S32. Obtain the current actual augmented observation vector generated in step S2. Compared with the prior prediction vector obtained in step S31 Calculate the observational variability The specific calculation method is as follows:

[0031] ;

[0032] in, These are preset hyperparameters used to balance the weights of the differences between the two items. KL divergence is used to measure the similarity or difference between two probability distributions. Used to quantify the degree of deviation between the current cognitive model and actual observations.

[0033] Specifically, step S4 includes:

[0034] S41. Obtain the current collaboration time step recorded by the system. Compared to the previous moment physical time interval Perform process noise covariance matrix analysis Dynamic calculation:

[0035] ;

[0036] in, The preset base drift rate, It is the identity matrix. Used to provide time-dimensional noise constraints for state extrapolation;

[0037] S42, Obtain System The posterior uncertainty covariance matrix obtained at convergence time 1 Perform state extrapolation calculations to calculate the prior uncertainty covariance at the current time. The specific calculation method is as follows:

[0038] ;

[0039] in, This is the state transition matrix, used to describe the evolution of cognitive states over time. This is used to express the degree of uncertainty in predicting the intentions of collaborating partners at the present moment based solely on historical knowledge.

[0040] Specifically, step S5 includes:

[0041] S51. Obtain the observation difference obtained in step S3. Perform noise scaling based on the exponential decay function to obtain... Dynamic observation noise covariance matrix at time step The specific calculation is as follows:

[0042] ;

[0043] in, As the baseline observation noise matrix, The sensitivity coefficient, Used to quantify the reliability of current observation data; the greater the observational discrepancy, the higher the reliability. The smaller the value, the higher the weight of the current observation in updating the cognitive state;

[0044] S52. Obtain the prior uncertain covariance obtained in step S4. The dynamic observation noise covariance matrix obtained in step S51 The optimal gain matrix is ​​inverted and multiplied to obtain the adaptive Kalman gain matrix. The specific calculation is as follows:

[0045] ;

[0046] S53. Obtain the current actual augmented observation vector obtained in step S2. The prior prediction vector obtained in step S3 The prior uncertainty covariance obtained in step S4 and the adaptive Kalman gain matrix obtained in step S52 After performing linear correction and covariance convergence operations, we obtain... Cognitive state vector updated at each time step With posterior uncertainty covariance :

[0047] ;

[0048] ;

[0049] in for 3D identity matrix.

[0050] Specifically, step S6 includes:

[0051] S61. Obtain the updated cognitive state vector obtained in step S5. and intelligent agents Its own historical behavioral sequence characteristics Perform a perspective transformation operation based on the mirror mapping function to obtain a second-order belief vector. The specific calculation is as follows:

[0052] ;

[0053] in, This is a pre-trained mirror mapping network used to simulate the cognitive logic from the perspective of a collaborating partner. These are the parameters for the mirrored network. This is a vector concatenation operation. Characterizing intelligent agents Anticipated partners The result of one's own perception of one's intentions;

[0054] S62. Obtain the true intent vector defined in step S1. and the second-order belief vector obtained in step S61 Calculate the Euclidean distance between the two to obtain the second-order consistency error. The specific calculation is as follows:

[0055] ;

[0056] like If this is the case, it is determined that there is implicit cognitive misalignment caused by information asymmetry among the agents, and a cooperative deadlock indicator is set. Otherwise, determine that the agents are cognitively aligned and mark them. .

[0057] Specifically, the calculation process for the communication necessity entropy value in step S7 is as follows:

[0058] (1) Obtain the updated cognitive state vector obtained in step S5 to conduct The operation yields the probability distribution of the collaborating partners' intentions. :

[0059] ;

[0060] in The cognitive state vector is mapped to the probability value of each intention dimension, and the sum of the probability values ​​of all dimensions is 1;

[0061] (2) Obtain the intention probability distribution obtained in step (1) The Shannon entropy of this probability distribution is calculated to characterize the necessity of communication, thus obtaining the communication necessity entropy value. The specific calculation is as follows:

[0062] ;

[0063] in, Intent probability distribution The Middle The probability value of each intent dimension. The communication necessity entropy value is the total number of dimensions in the baseline intent space. The smaller the value, the higher the certainty of the agent's understanding of the collaborating partner's intentions.

[0064] Specifically, in step S7, the pruning-quantification hierarchical collaboration involves: obtaining the communication necessity entropy value. The cooperative deadlock flag obtained in step S6, combined with a preset low-entropy threshold... With high entropy threshold Perform three-level hierarchical collaborative operations:

[0065] (1) If ,and If the agent has a clear understanding of its partner and the understanding of both parties is aligned, then natural language generation is suppressed, and the tool execution interface is directly called to execute the target action, with a communication cost of 0.

[0066] (2) If If this is the case, it is determined that there is a implicit cognitive misalignment among the agents, and the agents will then assign their own intention vectors... Decode the text into natural language communication and send it to collaborating partners to correct cognitive misalignments;

[0067] (3) If If the agent's understanding of the collaborating partner's intention is in a state of high uncertainty, the agent will generate natural language queries for multiple intention dimensions whose probability difference in the intention probability distribution is less than a preset fuzziness threshold in order to eliminate cognitive fuzziness.

[0068] The multi-agent collaborative decision-making system based on dynamic cognitive state modeling, applying the aforementioned multi-agent collaborative decision-making method based on dynamic cognitive state modeling, specifically includes the following modules:

[0069] The cognitive state initialization module is used to obtain the input task description and the set of collaborating agents, generate a baseline intent space for a unified multi-agent intent representation, and initialize the explicit cognitive state in the memory structure for each agent.

[0070] The large model semantic encoding module is used to load the pre-trained large model parameter set and encode the semantic features of the input data received by the agent from its collaborating partners.

[0071] The prior prediction and difference calculation module is used to generate prior prediction vectors of the collaborative partner's behavior and quantify the observed differences.

[0072] The adaptive cognitive state update module achieves real-time correction of the cognitive state through latent space adaptive Kalman filtering;

[0073] The second-order deadlock detection module is used to identify implicit cognitive misalignments between agents and detect collaborative deadlocks.

[0074] The hierarchical collaboration decision-making module achieves joint optimization of communication overhead and collaboration efficiency through pruning-quantification hierarchical collaboration.

[0075] In summary, the beneficial technical effects of the present invention are as follows:

[0076] 1. This invention addresses the cognitive lag problem in dynamic environments, enhancing the robustness and real-time performance of the system: By constructing a cognitive state update mechanism based on latent space adaptive Kalman filtering, this invention creatively maps the observation discrepancy based on a large model directly to the dynamic observation noise covariance matrix during the filtering process. This mechanism establishes a negative correlation between "surprise" (observation discrepancy) and "learning rate" (Kalman gain). When the system faces sudden intention drift or abnormal observations, it can automatically reduce its dependence on historical priors and significantly increase the weight of accepting new information. This effectively solves the problem of sluggish response of existing static parameter models in dynamic environments, ensuring that the agent can quickly converge to the true intention distribution and significantly enhancing the adaptive ability of multi-agent systems in complex environments.

[0077] 2. Possessing advanced recursive cognitive capabilities to effectively break collaborative deadlock: This invention introduces a second-order belief model based on a mirror mapping function. By simulating the perspective transformation of collaborating partners, it calculates the second-order consistency error. This mechanism enables the agent to possess deep recursive cognitive capabilities, fundamentally and accurately identifying implicit cognitive misalignments caused by information asymmetry. Through advanced deadlock detection, the system can forcibly trigger interruptions and guide explicit alignment before collaborative deadlock, invalid loops, or non-optimal Nash equilibria form. This greatly improves the decision-making quality and success rate of multi-agent systems in tasks involving strong games, conflicts of interest, or high complexity.

[0078] 3. Achieved joint optimization of communication overhead and collaboration efficiency: This invention establishes a hierarchical collaboration mechanism based on information entropy, namely "pruning-quantization". The system no longer relies on high-frequency redundant dialogue, but instead calculates the communication necessity entropy value of the intent distribution in real time. Only when the cognitive uncertainty is extremely high will the system consume computing power to generate high-cost natural language queries. In the tacit state of low entropy, the language channel is directly "pruned" and physical actions are performed. This mechanism greatly reduces the semantic drift caused by the context window limitation in large language model-based systems, reduces redundant token consumption and communication latency, and effectively suppresses the "communication storm" phenomenon common in multi-agent systems.

[0079] 4. Overcoming the limitations of unstructured text splicing, this invention achieves explicit social cognitive reasoning: Unlike existing technologies that mainly rely on unstructured natural language context splicing to maintain memory, this invention initializes an explicit cognitive state space for each agent, including its own intention vector, social cognitive matrix, and uncertainty covariance matrix. This structured state representation, combined with the semantic feature encoding of a large model, enables the agent to perform accurate mathematical modeling and probabilistic prediction, thereby achieving a leap from shallow text completion to deep logical reasoning based on "theory of mind". Attached Figure Description

[0080] Figure 1 This is a flowchart of the overall method of the present invention;

[0081] Figure 2 This is the logical architecture diagram of the intelligent agent system in this invention. Detailed Implementation

[0082] To make the technical means, creative features, objectives and effects of this invention clearer and easier to understand, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0083] Example

[0084] like Figure 1 , Figure 2 As shown, the multi-agent collaborative decision-making method based on dynamic cognitive state modeling provided by this invention specifically includes the following steps:

[0085] S1. Obtain the input task description With the collection of intelligent agents For each intelligent agent Initialize the explicit cognitive state in its memory structure The process of initializing explicit cognitive states is as follows:

[0086] S11. Set the dimension of the intent vector as follows: The baseline intent space is generated using an orthogonal initialization method. ;

[0087] S12, For each agent , initialize it in Cognitive state at any moment ,in, For intelligent agents Its own intention vector The social cognition matrix is ​​shown, where the first... row vector Represents intelligent agents For intelligent agents Predicting the initial intent To understand the uncertain covariance matrix, it is initially set as , The initial uncertainty coefficient is preset. for 3D identity matrix.

[0088] S2. Load the pre-trained large language model, obtain the input data of the collaborating partners received by the agent at the current collaborative time step, encode the semantic features of the data based on the large language model, and generate enhanced observation vectors of the collaborating partners' behavior. The specific execution steps include:

[0089] S21. Obtain the pre-trained large language model encoder layer Get collaboration time steps Time-based intelligent agent Received from collaborating partners Input data The input is fed into the encoder layer, where it generates the observation latent vector through semantic mapping. The specific calculation is as follows:

[0090] ;

[0091] Observation latent vector Transform unstructured natural language input into structured feature vectors aligned with the intent space;

[0092] S22. Obtain the data stored in the agent's memory structure. Global context state at time step Enhanced observation vectors are generated through an attention mechanism. The specific calculation is as follows:

[0093] ;

[0094] in, This is a normalization layer used to ensure a reasonable distribution range of feature vectors in the intent space. This represents multi-head attention operations, used to capture the semantic relevance between current observations and historical behaviors.

[0095] S3. Based on the cognitive state of the previous moment, generate a priori prediction vector of the agent's behavior towards the collaborating partner, compare it with the current enhanced observation vector, and obtain the observation difference degree. The specific execution steps include:

[0096] S31, Obtaining the System Time-based intelligent agent For intelligent agents Cognitive state vector By mapping it to the baseline intent space through a multilayer perceptron, a set of instructions for collaborating partners is generated. Prior prediction vector of behavior ;

[0097] S32. Obtain the current actual augmented observation vector generated in step S2. Compared with the prior prediction vector obtained in step S31 Calculate the observational variability The specific calculation method is as follows:

[0098] ;

[0099] in, These are preset hyperparameters used to balance the weights of the differences between the two items. KL divergence is used to measure the similarity or difference between two probability distributions. Used to quantify the degree of deviation between the current cognitive model and actual observations.

[0100] S4. Based on the posterior uncertainty covariance obtained in the previous time step and the time interval between the two collaborations, calculate the current prior uncertainty covariance. The specific steps include:

[0101] S41. Obtain the current collaboration time step recorded by the system. Compared to the previous moment physical time interval Perform process noise covariance matrix analysis Dynamic calculation:

[0102] ;

[0103] in, The preset base drift rate, It is the identity matrix. Used to provide time-dimensional noise constraints for state extrapolation;

[0104] S42, Obtain System The posterior uncertainty covariance matrix obtained at convergence time 1 Perform state extrapolation calculations to calculate the prior uncertainty covariance at the current time. The specific calculation method is as follows:

[0105] ;

[0106] in, This is the state transition matrix, used to describe the evolution of cognitive states over time. This is used to express the degree of uncertainty in predicting the intentions of collaborating partners at the present moment based solely on historical knowledge.

[0107] S5. Combining the current enhanced observation vector, prior prediction vector, observation discrepancy, and prior uncertainty covariance, update the cognitive state of the collaborating partners through latent space adaptive Kalman filtering. Specific execution steps include:

[0108] S51. Obtain the observation difference obtained in step S3. Perform noise scaling based on the exponential decay function to obtain... Dynamic observation noise covariance matrix at time step The specific calculation is as follows:

[0109] ;

[0110] in, As the baseline observation noise matrix, The sensitivity coefficient, Used to quantify the reliability of current observation data; the greater the observational discrepancy, the higher the reliability. The smaller the value, the higher the weight of the current observation in updating the cognitive state;

[0111] S52. Obtain the prior uncertain covariance obtained in step S4. The dynamic observation noise covariance matrix obtained in step S51 The optimal gain matrix is ​​inverted and multiplied to obtain the adaptive Kalman gain matrix. The specific calculation is as follows:

[0112] ;

[0113] S53. Obtain the current actual augmented observation vector obtained in step S2. The prior prediction vector obtained in step S3 The prior uncertainty covariance obtained in step S4 and the adaptive Kalman gain matrix obtained in step S52 After performing linear correction and covariance convergence operations, we obtain... Cognitive state vector updated at each time step With posterior uncertainty covariance :

[0114] ;

[0115] ;

[0116] in for 3D identity matrix.

[0117] Step S5 establishes a negative correlation coupling mechanism between "surprise" and "learning rate" by directly mapping the observation difference to the dynamic observation noise parameter in Kalman filtering. This enables the system to automatically reduce its reliance on historical priors and significantly increase the weight of accepting new information when faced with sudden intention drift or abnormal observations. This solves the problem of sluggish response of existing static parameter models in dynamic environments, ensures that the agent can quickly converge to the true intention distribution, and significantly improves the robustness and real-time performance of cognitive state updates.

[0118] S6. Based on the updated cognitive state and the agent's own historical behavioral characteristics, construct a second-order belief model to detect whether there is a collaborative deadlock caused by cognitive misalignment among agents. Specific steps include:

[0119] S61. Obtain the updated cognitive state vector obtained in step S5. and intelligent agents Its own historical behavioral sequence characteristics Perform a perspective transformation operation based on the mirror mapping function to obtain a second-order belief vector. The specific calculation is as follows:

[0120] ;

[0121] in, This is a pre-trained mirror mapping network used to simulate the cognitive logic from the perspective of a collaborating partner. These are the parameters for the mirrored network. This is a vector concatenation operation. Characterizing intelligent agents Anticipated partners The result of one's own perception of one's intentions;

[0122] S62. Obtain the true intent vector defined in step S1. and the second-order belief vector obtained in step S61 Calculate the Euclidean distance between the two to obtain the second-order consistency error. The specific calculation is as follows:

[0123] ;

[0124] like If this is the case, it is determined that there is implicit cognitive misalignment caused by information asymmetry among the agents, and a cooperative deadlock indicator is set. Otherwise, determine that the agents are cognitively aligned and mark them. .

[0125] Step S6 introduces a mirror mapping function to construct a second-order belief model and calculates the second-order consistency error, enabling the agent to possess the recursive cognitive ability to simulate the perspective of the collaborating partner. This mechanism can fundamentally and accurately identify implicit cognitive misalignments caused by information asymmetry, thereby forcibly triggering interruptions and guiding explicit alignment before collaborative deadlock or invalid loops form. This effectively breaks the "prisoner's dilemma" or collaborative deadlock in complex game scenarios, greatly improving the decision-making quality and success rate of multi-agent systems in highly conflicting tasks.

[0126] S7. Calculate the communication necessity entropy value, and perform pruning-quantification hierarchical cooperation based on the communication necessity entropy value and the cooperative deadlock detection results.

[0127] The calculation process for the communication necessity entropy value is as follows:

[0128] (1) Obtain the updated cognitive state vector obtained in step S5 to conduct The operation yields the probability distribution of the collaborating partners' intentions. :

[0129] ;

[0130] in The cognitive state vector is mapped to the probability value of each intention dimension, and the sum of the probability values ​​of all dimensions is 1;

[0131] (2) Obtain the intention probability distribution obtained in step (1) The Shannon entropy of this probability distribution is calculated to characterize the necessity of communication, thus obtaining the communication necessity entropy value. The specific calculation is as follows:

[0132] ;

[0133] in, Intent probability distribution The Middle The probability value of each intent dimension. The communication necessity entropy value is the total number of dimensions in the baseline intent space. The smaller the value, the higher the certainty of the agent's understanding of the collaborating partner's intentions.

[0134] The pruning-quantitative hierarchical collaboration process specifically involves: obtaining the communication necessity entropy value. The cooperative deadlock flag obtained in step S6, combined with a preset low-entropy threshold... With high entropy threshold Perform three-level hierarchical collaborative operations:

[0135] (1) If ,and If the agent has a clear understanding of its partner and the understanding of both parties is aligned, then natural language generation is suppressed, and the tool execution interface is directly called to execute the target action, with a communication cost of 0.

[0136] (2) If If this is the case, it is determined that there is a implicit cognitive misalignment among the agents, and the agents will then assign their own intention vectors... Decode the text into natural language communication and send it to collaborating partners to correct cognitive misalignments;

[0137] (3) If If the agent's understanding of the collaborating partner's intention is in a state of high uncertainty, the agent will generate natural language queries for multiple intention dimensions whose probability difference in the intention probability distribution is less than a preset fuzziness threshold in order to eliminate cognitive fuzziness.

[0138] Step S7 establishes a "pruning-quantization" hierarchical execution mechanism. The system only consumes computing power to generate high-cost natural language queries when the cognitive uncertainty is extremely high. In the tacit state of low entropy, the language channel is directly "pruned" and physical actions are executed. This greatly reduces redundant token consumption and communication latency, effectively suppresses the "communication storm" phenomenon common in multi-agent systems, and achieves joint optimization of communication overhead and collaboration efficiency.

[0139] Based on the above-mentioned multi-agent collaborative decision-making method using dynamic cognitive state modeling, this invention provides a multi-agent collaborative decision-making system based on dynamic cognitive state modeling, which specifically includes the following modules:

[0140] The cognitive state initialization module is used to obtain the input task description and the set of collaborating agents, generate a baseline intent space for a unified multi-agent intent representation, and initialize the explicit cognitive state in the memory structure for each agent.

[0141] The large model semantic encoding module is used to load the pre-trained large model parameter set and encode the semantic features of the input data received by the agent from its collaborating partners.

[0142] The prior prediction and difference calculation module is used to generate prior prediction vectors of the collaborative partner's behavior and quantify the observed differences.

[0143] The adaptive cognitive state update module achieves real-time correction of the cognitive state through latent space adaptive Kalman filtering;

[0144] The second-order deadlock detection module is used to identify implicit cognitive misalignments between agents and detect collaborative deadlocks.

[0145] The hierarchical collaboration decision-making module achieves joint optimization of communication overhead and collaboration efficiency through pruning-quantification hierarchical collaboration.

[0146] Therefore, the multi-agent collaborative decision-making method and system based on dynamic cognitive state modeling provided by this invention directly maps observational discrepancies to dynamic observational noise parameters using a latent space adaptive Kalman filter mechanism to establish adaptive coupling adjustment of "surprise" and "learning rate," utilizes a second-order belief model based on mirror mapping functions for recursive cognitive reasoning and high-order deadlock detection of collaborative partners' perspectives, and finally constructs a core process based on the pruning-quantification hierarchical execution architecture of communication necessity entropy. This achieves a closed-loop collaborative process of "explicit state modeling—dynamic cognitive calibration—on-demand hierarchical execution," solving the problem of... Existing multi-agent technologies based on large language models suffer from pain points such as shallow intent reasoning due to reliance on unstructured text splicing, and communication redundancy and collaborative deadlock caused by the lack of deep cognitive modeling. This paper improves the cognitive convergence speed of agents in dynamic environments, the decision quality in strong game scenarios, and the communication efficiency of the system. It significantly suppresses the "communication storm" phenomenon and semantic drift risk, and enhances the model's ability to capture implicit cognitive misalignment and complex intent distribution. It provides a robust, real-time, and low-overhead solution for multi-agent systems that require deep strategic collaboration in distributed artificial intelligence and complex game environments.

[0147] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-agent collaborative decision-making method based on dynamic cognitive state modeling, applied to multi-agent collaborative decision-making scenarios based on large language models, characterized in that... Includes the following steps: S1. Obtain the input task description and the set of agents participating in the collaboration, and initialize the explicit cognitive state for each agent; S2. Load the pre-trained large language model, obtain the input data of the cooperative partners received by the agent at the current cooperative time step, encode the semantic features of the large language model, and generate an enhanced observation vector of the cooperative partner's behavior. S3. Based on the cognitive state of the previous moment, generate the agent's prior prediction vector of the cooperative partner's behavior, compare it with the current enhanced observation vector, and obtain the observation difference degree. S4. Based on the posterior uncertainty covariance obtained in the previous time step and the time interval between the two collaborations, calculate the current prior uncertainty covariance. S5. Combining the current enhanced observation vector, prior prediction vector, observation difference, and prior uncertainty covariance, update the cognitive state of the collaborating partners through latent space adaptive Kalman filtering; S6. Based on the updated cognitive state and the agent's own historical behavioral characteristics, construct a second-order belief model to detect whether there is a collaborative deadlock caused by cognitive misalignment among agents. S7. Calculate the communication necessity entropy value, and perform pruning-quantification hierarchical cooperation based on the communication necessity entropy value and the cooperative deadlock detection results.

2. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 1, characterized in that, Step S1 specifically involves obtaining the input task description. With the collection of intelligent agents For each intelligent agent Initialize the explicit cognitive state in its memory structure The process of initializing explicit cognitive states is as follows: S11. Set the dimension of the intent vector as follows: The baseline intent space is generated using an orthogonal initialization method. ; S12, For each agent , initialize it in Cognitive state at any moment ,in, For intelligent agents Its own intention vector The social cognition matrix is ​​shown, where the first... row vector Represents intelligent agents For intelligent agents Predicting the initial intent To understand the uncertain covariance matrix, it is initially set as , The initial uncertainty coefficient is preset. for 3D identity matrix.

3. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 2, characterized in that, Step S2 specifically includes: S21. Obtain the pre-trained large language model encoder layer Get collaboration time steps Time-based intelligent agent Received from collaborating partners Input data The input is fed into the encoder layer, where it generates the observation latent vector through semantic mapping. The specific calculation is as follows: ； Observation latent vector Transform unstructured natural language input into structured feature vectors aligned with the intent space; S22. Obtain the data stored in the agent's memory structure. Global context state at time step Enhanced observation vectors are generated through an attention mechanism. The specific calculation is as follows: ； in, This is a normalization layer used to ensure a reasonable distribution range of feature vectors in the intent space. This represents multi-head attention operations, used to capture the semantic relevance between current observations and historical behaviors.

4. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 3, characterized in that, Step S3 specifically includes: S31, Obtaining the System Time-based intelligent agent For intelligent agents Cognitive state vector By mapping it to the baseline intent space through a multilayer perceptron, a set of instructions for collaborating partners is generated. Prior prediction vector of behavior ; S32. Obtain the current actual augmented observation vector generated in step S2. Compared with the prior prediction vector obtained in step S31 Calculate the observational variability The specific calculation method is as follows: ； in, These are preset hyperparameters used to balance the weights of the differences between the two items. KL divergence is used to measure the similarity or difference between two probability distributions. Used to quantify the degree of deviation between the current cognitive model and actual observations.

5. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 4, characterized in that, Step S4 specifically includes: S41. Obtain the current collaboration time step recorded by the system. Compared to the previous moment physical time interval Perform process noise covariance matrix analysis Dynamic calculation: ； in, The preset base drift rate, It is the identity matrix. Used to provide time-dimensional noise constraints for state extrapolation; S42, Obtain System The posterior uncertainty covariance matrix obtained at convergence time 1 Perform state extrapolation calculations to calculate the prior uncertainty covariance at the current time. The specific calculation method is as follows: ； in, This is the state transition matrix, used to describe the evolution of cognitive states over time. This is used to express the degree of uncertainty in predicting the intentions of collaborating partners at the present moment based solely on historical knowledge.

6. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 5, characterized in that, Step S5 specifically includes: S51. Obtain the observation difference obtained in step S3. Perform noise scaling based on the exponential decay function to obtain... Dynamic observation noise covariance matrix at time step The specific calculation is as follows: ； in, As the baseline observation noise matrix, The sensitivity coefficient, Used to quantify the reliability of current observation data; the greater the observational discrepancy, the higher the reliability. The smaller the value, the higher the weight of the current observation in updating the cognitive state; S52. Obtain the prior uncertain covariance obtained in step S4. The dynamic observation noise covariance matrix obtained in step S51 The optimal gain matrix is ​​inverted and multiplied to obtain the adaptive Kalman gain matrix. The specific calculation is as follows: ； S53. Obtain the current actual augmented observation vector obtained in step S2. The prior prediction vector obtained in step S3 The prior uncertainty covariance obtained in step S4 and the adaptive Kalman gain matrix obtained in step S52 After performing linear correction and covariance convergence operations, we obtain... Cognitive state vector updated at each time step With posterior uncertainty covariance : ； ； in for 3D identity matrix.

7. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 6, characterized in that, Step S6 specifically includes: S61. Obtain the updated cognitive state vector obtained in step S5. and intelligent agents Its own historical behavioral sequence characteristics Perform a perspective transformation operation based on the mirror mapping function to obtain a second-order belief vector. The specific calculation is as follows: ； in, This is a pre-trained mirror mapping network used to simulate the cognitive logic from the perspective of a collaborating partner. These are the parameters for the mirrored network. This is a vector concatenation operation. Characterizing intelligent agents Anticipated partners The result of one's own perception of one's intentions; S62. Obtain the true intent vector defined in step S1. and the second-order belief vector obtained in step S61 Calculate the Euclidean distance between the two to obtain the second-order consistency error. The specific calculation is as follows: ； like If this is the case, it is determined that there is implicit cognitive misalignment caused by information asymmetry among the agents, and a cooperative deadlock indicator is set. Otherwise, determine that the agents are cognitively aligned and mark them. .

8. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 7, characterized in that, The calculation process for the communication necessity entropy value in step S7 is as follows: (1) Obtain the updated cognitive state vector obtained in step S5 to conduct The operation yields the probability distribution of the collaborating partners' intentions. : ； in The cognitive state vector is mapped to the probability value of each intention dimension, and the sum of the probability values ​​of all dimensions is 1; (2) Obtain the intention probability distribution obtained in step (1) The Shannon entropy of this probability distribution is calculated to characterize the necessity of communication, thus obtaining the communication necessity entropy value. The specific calculation is as follows: ； in, Intent probability distribution The Middle The probability value of each intent dimension. The communication necessity entropy value is the total number of dimensions in the baseline intent space. The smaller the value, the higher the certainty of the agent's understanding of the collaborating partner's intentions.

9. The multi-agent collaborative decision-making method based on dynamic cognitive state modeling according to claim 8, characterized in that, Step S7, pruning-quantification hierarchical collaboration, specifically involves obtaining the communication necessity entropy value. The cooperative deadlock flag obtained in step S6, combined with a preset low-entropy threshold... With high entropy threshold Perform three-level hierarchical collaborative operations: (1) If ,and If the agent has a clear understanding of its partner and the understanding of both parties is aligned, then natural language generation is suppressed, and the tool execution interface is directly called to execute the target action, with a communication cost of 0. (2) If If this is the case, it is determined that there is a implicit cognitive misalignment among the agents, and the agents will then assign their own intention vectors... Decode the text into natural language communication and send it to collaborating partners to correct cognitive misalignments; (3) If If the agent's understanding of the collaborating partner's intention is in a state of high uncertainty, the agent will generate natural language queries for multiple intention dimensions whose probability difference in the intention probability distribution is less than a preset fuzziness threshold in order to eliminate cognitive fuzziness.

10. A multi-agent collaborative decision-making system based on dynamic cognitive state modeling, employing the multi-agent collaborative decision-making method based on dynamic cognitive state modeling as described in any one of claims 1-9, characterized in that, Includes the following modules: The cognitive state initialization module is used to obtain the input task description and the set of collaborating agents, generate a baseline intent space for a unified multi-agent intent representation, and initialize the explicit cognitive state in the memory structure for each agent. The large model semantic encoding module is used to load the pre-trained large model parameter set and encode the semantic features of the input data received by the agent from its collaborating partners. The prior prediction and difference calculation module is used to generate prior prediction vectors of the collaborative partner's behavior and quantify the observed differences. The adaptive cognitive state update module achieves real-time correction of the cognitive state through latent space adaptive Kalman filtering; The second-order deadlock detection module is used to identify implicit cognitive misalignments between agents and detect collaborative deadlocks. The hierarchical collaboration decision-making module achieves joint optimization of communication overhead and collaboration efficiency through pruning-quantification hierarchical collaboration.

Images