A hybrid open multi-agent system and method of three-layer component architecture and scene constraint

Abstract

The application discloses a kind of three-layer component architecture and scene constraint hybrid open multi-agent system and method, wherein the method comprises: by preparing component group configuration target, meaning, signal and prior shared understanding, form scalable initial semantic basis;Through interactive component group, the interactive task of single focusing on single concept attribute is executed, and demand expression, scheme feedback and consensus verification are completed under the rule of physical concise communication;The purpose of the application is to solve the problem that in the prior art, in the hybrid open multi-agent system containing human and artificial agent, due to the ambiguity of human communication mode, the limited cognitive load, the sensitive interaction cost, and the system needs to support dynamic agent composition and unpredictable task scene, the traditional fixed communication protocol and formal semantic alignment mechanism are difficult to adapt, and then it is impossible to efficiently establish shared task-oriented understanding under the double constraints of "less assumption and less interaction".

Description

Technical Field

[0001] This invention relates to the technical field of information technology, specifically to a hybrid open multi-agent system and method with a three-layer component architecture and scenario constraints. Background Technology

[0002] With the rapid development of artificial intelligence technology, the collaborative relationship between computers and humans is becoming increasingly close, and multi-agent systems are being widely used in scenarios such as negotiation, task delegation, and resource allocation. Among them, open multi-agent systems, due to the dynamic change of the agent population, cannot pre-fix all communication protocols. They require agents to gradually build shared understanding through incremental interaction, and must minimize prior assumptions to maintain the openness of the system.

[0003] In existing technologies, various methods have emerged for establishing shared understanding in Open Multi-Agent Systems (OMAS), such as lexical semantic alignment methods based on language games, semantic mapping methods based on ontology matching, and interaction specification methods based on dialogue protocols. These methods have achieved certain results in purely AI agent interaction scenarios, enabling semantic coordination and task collaboration among agents. However, when OMAS is expanded to include hybrid populations of humans and AI agents, existing technologies reveal significant limitations:

[0004] On the one hand, humans do not follow strict formal semantic rules, and their communication methods are ambiguous and context-dependent. Existing methods based on fixed protocols or formal semantics are insufficient in adapting to human communication habits. On the other hand, the number of interactions between humans and intelligent agents is strictly limited by factors such as time and cognitive load. Existing methods often require a large number of interaction iterations to establish a stable shared understanding, and they do not take into account the balance between "minimizing prior assumptions" and "minimizing interaction costs", resulting in low efficiency in practical applications.

[0005] Furthermore, existing research largely focuses on proposing specific communication implementation methods, lacking a systematic description and unified guiding framework for the shared understanding establishment process. The design of agents in different scenarios lacks unified standards, resulting in poor compatibility and scalability, making it difficult to meet the interaction needs of unforeseen scenarios in hybrid systems. Therefore, there is an urgent need for a shared task-oriented understanding establishment scheme that can adapt to the characteristics of hybrid open multi-agent systems and provide unified design guidance, in order to improve the adaptability, systematicity, and interaction efficiency of existing technologies in hybrid scenarios. Summary of the Invention

[0006] This invention provides a hybrid open multi-agent system and method with a three-layer component architecture and scenario constraints. The purpose is to solve the problem in the prior art that in hybrid open multi-agent systems containing both humans and artificial intelligence agents, the ambiguity of human communication methods, the limited cognitive load, the sensitivity to interaction costs, and the need for the system to support dynamic agent composition and unpredictable task scenarios make it difficult to adapt traditional fixed communication protocols and formal semantic alignment mechanisms, thus making it impossible to efficiently establish shared task-oriented understanding under the dual constraints of "few assumptions and few interactions".

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A hybrid open multi-agent approach with a three-layer component architecture and scenario constraints includes: configuring goals, meanings, signals, and prior shared understanding through a preparatory component group to form a scalable initial semantic foundation; executing single-session interactive tasks focused on a single conceptual attribute through an interaction component group, completing requirement expression, solution feedback, and consensus verification under physically concise communication rules; dynamically sensing goal relevance, updating semantic understanding, evaluating the overall sharing level, and quantifying cognitive load through a strategy component group to drive adaptive optimization of the interaction process; forcibly applying a triple hybrid scenario constraint of physically concise communication, task-goal consistency, and efficiency priority during the interaction process to ensure that the human-machine collaboration process meets the requirements of low cognitive load, high interaction efficiency, and strong scenario adaptability; terminating the interaction process and outputting the final shared task-oriented understanding result when the overall shared understanding reaches a preset goal completion threshold and the interaction efficiency and cognitive load indicators meet the constraint conditions.

[0009] In one aspect of the invention, the provision of a scalable initial semantic foundation through preparatory component group configuration of goals, meanings, signals, and prior shared understanding includes:

[0010] Define shared understanding of target type, target completion threshold and target priority, and generate target component P1;

[0011] The core concepts related to the task are broken down into a three-dimensional structure of concept-attribute-initial value range, and a meaningful component P2 is constructed.

[0012] The signal forms that can be used by humans and AI are limited to natural language or body movements, and at least one adaptive signal form is bound to each concept attribute to generate signal component P3;

[0013] Three categories of minimal prior content are set: basic language consensus, core interaction rules, and basic attribute consensus, which constitute the prior shared understanding component P4.

[0014] The parameters of components P1 to P4 are written into the system configuration file as the initial basis for subsequent interactions and policy execution.

[0015] In one aspect of the invention, the step of performing a single interactive task focusing on a single conceptual attribute through an interactive component group, and completing requirement expression, solution feedback, and consensus verification under physically concise communication rules, includes:

[0016] Based on the core concept attribute list in P2, assign a unique corresponding concept attribute to the current interaction round and generate the interaction task component I1.

[0017] Determine whether the current interaction is an open-ended or restricted scenario, set the initial information input method, and generate the interaction scenario component I2;

[0018] Receive natural language or action signals input by the initiator, verify whether they conform to the signal length and meaning uniqueness rules defined in P3, and if they do not conform, return a simplified prompt and generate the initiator behavior component I3;

[0019] The respondent generates a solution signal or clarification signal based on its current understanding. The signal content includes a matching solution and attribute quantification value, or specific questions and no more than three options, generating the respondent behavior component I4.

[0020] Monitor the termination conditions of the interaction, including the initiator outputting acknowledgment feedback, the task performance score being ≥90%, the initiator requesting a pause, or the duration of a single interaction exceeding 5 minutes, and generate the termination behavior component I5;

[0021] The task performance score is calculated using the following formula: Where R is the attribute value range finally approved by the initiator, S is the attribute value of the responder's plan, R_max is the upper limit of the initial value range of the attribute in P2, and the task performance component I6 is generated.

[0022] In one aspect of the invention, the step of dynamically sensing target relevance, updating semantic understanding, evaluating overall sharing level, and quantifying cognitive load through a strategy component group to drive adaptive optimization of the interaction process includes:

[0023] Calculate the ratio of the number of unconsensus concept attributes involved in the current interaction to the total number of attributes to obtain the target relevance C. If C < 30%, switch to high relevance attribute interaction and generate target-aware interaction component A1.

[0024] Based on the task performance score P output by I6, if P≥80%, S is stored in the shared value range library; if 50%≤P<80%, the value range is narrowed to (R∩S); if P<50%, the original range is retained and three optional clarifications are inserted before the next interaction to generate the understanding update component A2.

[0025] The overall shared understanding is calculated by combining the p-values ​​of each round of interaction with the corresponding attribute weights. If G≥P1 target completion threshold, the process terminates and an understanding evaluation component A3 is generated.

[0026] Statistically calculate the number of human output signals n_s and the number of complex inferences n_c, and substitute them into the formula. Calculate the cognitive load L, and simultaneously calculate the interaction efficiency. If L≥80% or E<0.5, a simplification strategy is triggered, generating the efficiency evaluation component A4.

[0027] In one aspect of the invention, the forced application of a triple hybrid scenario constraint of physical concise communication, task-goal consistency, and efficiency priority during the interaction process includes:

[0028] All signals must be carried out in the form of text, voice or body movements, and purely digital encoding is prohibited; human output signals should not exceed two sentences per signal and the vocabulary difficulty should not be higher than the junior high school level; AI output signals should not exceed two sentences per signal, should not contain professional terminology, and should not have more than five visual elements.

[0029] Each round of I1 interactive task must correspond to a core concept attribute in P2, and the target relevance C ≥ 30%; the correlation coefficient between the I6 task performance score and the overall shared understanding G is calculated using the Pearson formula and is required to be ≥ 0.8.

[0030] The total human interaction time T_total ≤ 15 minutes, the proportion of invalid interaction rounds ≤ 20%, and the cognitive load L ≤ 80%.

[0031] In another aspect, the present invention also relates to a hybrid open multi-agent system with a three-layer component architecture and scene constraints, the system comprising:

[0032] The preparatory component configuration module is used to define the target, meaning, signal and prior shared understanding, and generate the parameters of components P1-P4;

[0033] The interaction execution module is used to control the single-round interaction task, scenario, behavior and performance evaluation, and generate I1-I6 component outputs;

[0034] The strategy optimization module is used to dynamically adjust the interaction focus, update semantic understanding, evaluate the overall level, and adjust cognitive load to generate A1-A4 component decisions.

[0035] The constraint verification module is used to monitor the triple constraints of physical simplification communication, task-goal consistency, and efficiency priority in real time, intercept violation signals, and trigger simplification mechanisms.

[0036] The process termination determination module is used to output the final shared task-oriented understanding result when the overall shared understanding meets the standard and the constraints are satisfied.

[0037] In one aspect of the invention, the preparatory component configuration module is configured as follows:

[0038] Receive the task description text, extract the shared understanding of the target type, target completion threshold and target priority, and generate the target component P1;

[0039] Semantic parsing of the task description is performed to identify core concepts and decompose them into a three-dimensional structure of concept-attribute-initial value range, and to construct the meaning component P2;

[0040] Based on the capability list of human-computer interaction devices, bind at least one signal form to each attribute to generate signal component P3;

[0041] Establish consensus on basic language, core interaction rules, and basic attributes to form the prior shared understanding component P4;

[0042] The parameters of components P1 to P4 are serialized and written to the system configuration file.

[0043] In one aspect of the invention, the interactive execution module is configured as follows:

[0044] Read the core concept attribute list from P2, sort it by priority, and assign it to the interaction rounds one by one to generate the interaction task component I1;

[0045] Based on the task context, determine whether the interaction mode is an open scene or a limited scene, and generate the interaction scene component I2;

[0046] Receive human input signals, verify whether they conform to the signal length and vocabulary difficulty rules, and if they do not conform, return a simplified prompt and generate the initiator behavior component I3;

[0047] The AI ​​agent generates a solution signal or clarification signal based on its current understanding, and then generates the responder behavior component I4.

[0048] Monitor interaction termination conditions, including positive feedback, task performance score ≥90%, pause request or single interaction timeout of 5 minutes, and generate end behavior component I5;

[0049] Calculate task performance score Generate task performance component I6.

[0050] In one aspect of the invention, the strategy optimization module is configured as follows:

[0051] Calculate the target correlation If C < 30%, switch to high-relevance attribute interaction and generate target-aware interaction component A1;

[0052] Based on the task performance score P, a semantic update strategy is executed: when P≥80%, the value is stored in the shared value range library; when 50%≤P<80%, the intersection is taken; when P<50%, the original range is retained and three optional clarifications are inserted to generate the understanding update component A2.

[0053] Calculate overall shared understanding If G ≥ the target completion threshold, a termination preparation signal is sent and an understanding evaluation component A3 is generated.

[0054] Calculate cognitive load With interaction efficiency If L≥80% or E<0.5, a simplification strategy is triggered, generating the efficiency evaluation component A4.

[0055] In one aspect of the invention, the constraint verification module is configured as follows:

[0056] Verify that all signals are carried by text, voice, or body movements, and prohibit purely digital encoding; verify that each human signal is no more than two sentences and the vocabulary difficulty is no higher than junior high school level; verify that each AI signal is no more than two sentences, contains no professional terminology, and has no more than five visual elements.

[0057] Verify whether each round of interactive tasks corresponds to a core concept attribute in P2 and whether the target relevance C ≥ 30%; verify whether the Pearson correlation coefficient between the task performance score and the overall shared understanding is ≥ 0.8;

[0058] Verify that the total human interaction time T_total ≤ 15 minutes, the percentage of invalid interaction rounds ≤ 20%, and the cognitive load L ≤ 80%; if any constraint is violated, intercept the signal and return a simplified prompt.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] First, this invention, through a hybrid scenario constraint and strategy component, strictly limits signal length, vocabulary difficulty, and the number of complex inferences, avoiding the mental burden of complex reasoning chains in pure AI systems. Simultaneously, through an informal semantic clarification sub-process, it transforms vague human expressions into intuitive feedback with ≤3 options, eliminating the need for humans to actively quantify requirements, significantly lowering the cognitive threshold, adapting to human-machine collaboration characteristics, and reducing the cost of human participation.

[0061] Through the preparatory component P2 (initial configuration of the fuzzy value range of the meaning set), the interaction components I3-I4 (natural language adaptation of demand / solution signals), and the strategy component A2 (incremental semantic update), it conforms to the natural communication logic of humans "preliminary description - feedback adjustment - consensus locking", without requiring humans to learn professional communication rules.

[0062] Secondly, the system minimizes prior assumptions, ensuring its openness: prior shared understanding only includes core consensus such as basic language and interaction rules, without the need for pre-defined fixed communication protocols or full semantic mapping. This allows for dynamic changes in the composition of agents, such as the addition of new AI / humans, and allows for unpredictable scenarios, such as switching between different tasks. This aligns with the essential requirements of open multi-agent systems, solves the problem of poor adaptability of traditional fixed protocols, and adapts to the dynamic characteristics of open systems.

[0063] Furthermore, through the closed-loop process of the strategy component A2 (which iteratively narrows the range of concept values ​​based on interaction results) and the interaction component, semantic understanding is supported to deepen gradually with interaction, without the need to achieve full consensus all at once. Even if the agent's cognition changes in the system, it can be quickly realigned through multiple rounds of iteration to adapt to the dynamic evolution of semantics.

[0064] Finally, this invention extends the shared understanding framework from a pure AI system to a hybrid open multi-agent system combining humans and AI, solving the core shortcoming of existing technologies that cannot adapt to human characteristics. It can be widely applied in service, industrial collaboration, and government scenarios that require human-machine collaboration to complete tasks. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0066] Figure 1 This is a flowchart of a hybrid open multi-agent method based on a three-layer component architecture and scene constraints according to the present invention.

[0067] Figure 2 This is a framework diagram of a hybrid open multi-agent system with a three-layer component architecture and scene constraints according to the present invention. Detailed Implementation

[0068] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0069] Example 1

[0070] Please see Figure 1 as well as Figure 2As shown, this embodiment discloses a hybrid open multi-agent method with a three-layer component architecture and scenario constraints, including: configuring goals, meanings, signals, and prior shared understanding through a preparatory component group to form a scalable initial semantic foundation; executing single-time interactive tasks focusing on a single conceptual attribute through an interaction component group, completing requirement expression, solution feedback, and consensus verification under physically concise communication rules; dynamically sensing goal relevance, updating semantic understanding, evaluating the overall sharing level, and quantifying cognitive load through a strategy component group to drive adaptive optimization of the interaction process; forcibly applying a triple hybrid scenario constraint of physically concise communication, task-goal consistency, and efficiency priority during the interaction process to ensure that the human-machine collaboration process meets the requirements of low cognitive load, high interaction efficiency, and strong scenario adaptability; and terminating the interaction process and outputting the final shared task-oriented understanding result when the overall shared understanding reaches a preset goal completion threshold and the interaction efficiency and cognitive load indicators meet the constraint conditions.

[0071] First, this invention, through a hybrid scenario constraint and strategy component, strictly limits signal length, vocabulary difficulty, and the number of complex inferences, avoiding the mental burden of complex reasoning chains in pure AI systems. Simultaneously, through an informal semantic clarification sub-process, it transforms vague human expressions into intuitive feedback with ≤3 options, eliminating the need for humans to actively quantify requirements, significantly lowering the cognitive threshold, adapting to human-machine collaboration characteristics, and reducing the cost of human participation.

[0072] Through the preparatory component P2 (initial configuration of the fuzzy value range of the meaning set), the interaction components I3-I4 (natural language adaptation of demand / solution signals), and the strategy component A2 (incremental semantic update), it conforms to the natural communication logic of humans "preliminary description - feedback adjustment - consensus locking", without requiring humans to learn professional communication rules.

[0073] Secondly, the system minimizes prior assumptions, ensuring its openness: prior shared understanding only includes core consensus such as basic language and interaction rules, without the need for pre-defined fixed communication protocols or full semantic mapping. This allows for dynamic changes in the composition of agents, such as the addition of new AI / humans, and allows for unpredictable scenarios, such as switching between different tasks. This aligns with the essential requirements of open multi-agent systems, solves the problem of poor adaptability of traditional fixed protocols, and adapts to the dynamic characteristics of open systems.

[0074] Furthermore, through the closed-loop process of the strategy component A2 (which iteratively narrows the range of concept values ​​based on interaction results) and the interaction component, semantic understanding is supported to deepen gradually with interaction, without the need to achieve full consensus all at once. Even if the agent's cognition changes in the system, it can be quickly realigned through multiple rounds of iteration to adapt to the dynamic evolution of semantics.

[0075] Finally, this invention extends the shared understanding framework from a pure AI system to a hybrid open multi-agent system combining humans and AI, solving the core shortcoming of existing technologies that cannot adapt to human characteristics. It can be widely applied in service, industrial collaboration, and government scenarios that require human-machine collaboration to complete tasks.

[0076] In some embodiments, the step of configuring the target, meaning, signal, and prior shared understanding through the preparatory component group to form a scalable initial semantic foundation includes:

[0077] Define shared understanding of target type, target completion threshold and target priority, and generate target component P1;

[0078] The core concepts related to the task are broken down into a three-dimensional structure of concept-attribute-initial value range, and a meaningful component P2 is constructed.

[0079] The signal forms that can be used by humans and AI are limited to natural language or body movements, and at least one adaptive signal form is bound to each concept attribute to generate signal component P3;

[0080] Three categories of minimal prior content are set: basic language consensus, core interaction rules, and basic attribute consensus, which constitute the prior shared understanding component P4.

[0081] The parameters of components P1 to P4 are written into the system configuration file as the initial basis for subsequent interactions and policy execution.

[0082] Among them, the meaning component P2 is defined through a three-dimensional structure of concept-attribute-value range, as shown in the following formula:

[0083]

[0084] in, M represents the set of meanings, n represents the number of core concepts, and k represents the number of attributes for each concept. For the i-th core concept, For the j-th attribute of the i-th concept, Let be the initial value range for the j-th attribute of the i-th concept.

[0085] The configuration of signal component P3 includes signal mapping rules and each concept attribute. It corresponds to at least one human-understandable signal form F ij The specific formula is as follows:

[0086]

[0087] The prior shared understanding component P4 defines the pre-agreed consensus U0 of the agents, with the following formula:

[0088]

[0089] in, To establish a consensus on basic languages ​​(such as commonly used Chinese vocabulary and the meaning of numbers). For consensus on interaction rules, Consensus on basic attributes (such as the meaning of units like "minutes" and "yuan", and a preliminary qualitative understanding of "near").

[0090] In some embodiments, the step of performing a single interactive task focused on a single conceptual attribute through an interactive component group, and completing requirement expression, solution feedback, and consensus verification under physically concise communication rules, includes:

[0091] Based on the core concept attribute list in P2, assign a unique corresponding concept attribute to the current interaction round and generate the interaction task component I1.

[0092] Determine whether the current interaction is an open-ended or restricted scenario, set the initial information input method, and generate the interaction scenario component I2;

[0093] Receive natural language or action signals input by the initiator, verify whether they conform to the signal length and meaning uniqueness rules defined in P3, and if they do not conform, return a simplified prompt and generate the initiator behavior component I3;

[0094] The respondent generates a solution signal or clarification signal based on its current understanding. The signal content includes a matching solution and attribute quantification value, or specific questions and no more than three options, generating the respondent behavior component I4.

[0095] Monitor the termination conditions of the interaction, including the initiator outputting acknowledgment feedback, the task performance score being ≥90%, the initiator requesting a pause, or the duration of a single interaction exceeding 5 minutes, and generate the termination behavior component I5;

[0096] The task performance score is calculated using the following formula: Where R is the attribute value range finally approved by the initiator, S is the attribute value of the responder's plan, R_max is the upper limit of the initial value range of the attribute in P2, and the task performance component I6 is generated.

[0097] Among them, the interactive task component I1 limits each interactive task to only one core concept attribute, avoiding semantic confusion caused by multiple concepts running in parallel. The specific formula is as follows:

[0098]

[0099] in, For the i-th interaction task, These are the corresponding core concept attributes.

[0100] The task performance component I6 uses consensus achievement as its core metric, as shown in the following formula:

[0101]

[0102] Where P is the task performance score, R is the attribute value range finally approved by the initiator, S is the attribute value of the responder's solution, and R_max is the maximum value of the initial value range in P2.

[0103] In some embodiments, the step of dynamically sensing target relevance, updating semantic understanding, evaluating overall sharing level, and quantifying cognitive load through a strategy component group to drive adaptive optimization of the interaction process includes:

[0104] Calculate the ratio of the number of unconsensus concept attributes involved in the current interaction to the total number of attributes to obtain the target relevance C. If C < 30%, switch to high relevance attribute interaction and generate target-aware interaction component A1.

[0105] Based on the task performance score P output by I6, if P≥80%, S is stored in the shared value range library; if 50%≤P<80%, the value range is narrowed to (R∩S); if P<50%, the original range is retained and three optional clarifications are inserted before the next interaction to generate the understanding update component A2.

[0106] The overall shared understanding is calculated by combining the p-values ​​of each round of interaction with the corresponding attribute weights. If G≥P1 target completion threshold, the process terminates and an understanding evaluation component A3 is generated.

[0107] Statistically calculate the number of human output signals n_s and the number of complex inferences n_c, and substitute them into the formula. Calculate the cognitive load L, and simultaneously calculate the interaction efficiency. If L≥80% or E<0.5, a simplification strategy is triggered, generating the efficiency evaluation component A4.

[0108] The formula for calculating the target relevance C is as follows:

[0109]

[0110] in, This represents the number of unresolved conceptual attributes involved in the current interaction. This represents the total number of conceptual attributes in P2.

[0111] The formula for calculating the overall shared understanding level G is as follows:

[0112]

[0113] in, The concept attribute weight corresponding to the i-th interaction is set by the target priority of P1.

[0114] The formula for calculating cognitive load L is as follows:

[0115]

[0116] in, The number of signals output to humans. For complex reasoning times, This represents the maximum acceptable cognitive load threshold for humans.

[0117] The formula for calculating interaction efficiency E is as follows:

[0118]

[0119] in, To reach a consensus on the number of conceptual attributes, This represents the total interaction time.

[0120] In some embodiments, the forced application of a triple hybrid scenario constraint of physical concise communication, task-goal consistency, and efficiency priority during the interaction process includes:

[0121] All signals must be carried out in the form of text, voice or body movements, and purely digital encoding is prohibited; human output signals should not exceed two sentences per signal and the vocabulary difficulty should not be higher than the junior high school level; AI output signals should not exceed two sentences per signal, should not contain professional terminology, and should not have more than five visual elements.

[0122] Each round of I1 interactive task must correspond to a core concept attribute in P2, and the target relevance C ≥ 30%; the correlation coefficient between the I6 task performance score and the overall shared understanding G is calculated using the Pearson formula and is required to be ≥ 0.8.

[0123] The total human interaction time T_total ≤ 15 minutes, the proportion of invalid interaction rounds ≤ 20%, and the cognitive load L ≤ 80%.

[0124] In some different embodiments, the present invention also relates to a hybrid open multi-agent system with a three-layer component architecture and scene constraints, the system comprising:

[0125] The preparatory component configuration module is used to define the target, meaning, signal and prior shared understanding, and generate the parameters of components P1-P4;

[0126] The interaction execution module is used to control the single-round interaction task, scenario, behavior and performance evaluation, and generate I1-I6 component outputs;

[0127] The strategy optimization module is used to dynamically adjust the interaction focus, update semantic understanding, evaluate the overall level, and adjust cognitive load to generate A1-A4 component decisions.

[0128] The constraint verification module is used to monitor the triple constraints of physical simplification communication, task-goal consistency, and efficiency priority in real time, intercept violation signals, and trigger simplification mechanisms.

[0129] The process termination determination module is used to output the final shared task-oriented understanding result when the overall shared understanding meets the standard and the constraints are satisfied.

[0130] In some embodiments, the preparatory component configuration module is configured as follows:

[0131] Receive the task description text, extract the shared understanding of the target type, target completion threshold and target priority, and generate the target component P1;

[0132] Semantic parsing of the task description is performed to identify core concepts and decompose them into a three-dimensional structure of concept-attribute-initial value range, and to construct the meaning component P2;

[0133] Based on the capability list of human-computer interaction devices, bind at least one signal form to each attribute to generate signal component P3;

[0134] Establish consensus on basic language, core interaction rules, and basic attributes to form the prior shared understanding component P4;

[0135] The parameters of components P1 to P4 are serialized and written to the system configuration file.

[0136] In some embodiments, the interactive execution module is configured to:

[0137] Read the core concept attribute list from P2, sort it by priority, and assign it to the interaction rounds one by one to generate the interaction task component I1;

[0138] Based on the task context, determine whether the interaction mode is an open scene or a limited scene, and generate the interaction scene component I2;

[0139] Receive human input signals, verify whether they conform to the signal length and vocabulary difficulty rules, and if they do not conform, return a simplified prompt and generate the initiator behavior component I3;

[0140] The AI ​​agent generates a solution signal or clarification signal based on its current understanding, and then generates the responder behavior component I4.

[0141] Monitor interaction termination conditions, including positive feedback, task performance score ≥90%, pause request or single interaction timeout of 5 minutes, and generate end behavior component I5;

[0142] Calculate task performance score Generate task performance component I6.

[0143] In some embodiments, the policy optimization module is configured to:

[0144] Calculate the target correlation If C < 30%, switch to high-relevance attribute interaction and generate target-aware interaction component A1;

[0145] Based on the task performance score P, a semantic update strategy is executed: when P≥80%, the value is stored in the shared value range library; when 50%≤P<80%, the intersection is taken; when P<50%, the original range is retained and three optional clarifications are inserted to generate the understanding update component A2.

[0146] Calculate overall shared understanding If G ≥ the target completion threshold, a termination preparation signal is sent and an understanding evaluation component A3 is generated.

[0147] Calculate cognitive load With interaction efficiency If L≥80% or E<0.5, a simplification strategy is triggered, generating the efficiency evaluation component A4.

[0148] In some embodiments, the constraint verification module is configured as follows:

[0149] Verify that all signals are carried by text, voice, or body movements, and prohibit purely digital encoding; verify that each human signal is no more than two sentences and the vocabulary difficulty is no higher than junior high school level; verify that each AI signal is no more than two sentences, contains no professional terminology, and has no more than five visual elements.

[0150] Verify whether each round of interactive tasks corresponds to a core concept attribute in P2 and whether the target relevance C ≥ 30%; verify whether the Pearson correlation coefficient between the task performance score and the overall shared understanding is ≥ 0.8;

[0151] Verify that the total human interaction time T_total ≤ 15 minutes, the percentage of invalid interaction rounds ≤ 20%, and the cognitive load L ≤ 80%; if any constraint is violated, intercept the signal and return a simplified prompt.

[0152] Example 2

[0153] Please see Figure 1 as well as Figure 2 As shown, this embodiment is a further optimization based on embodiment one. In this embodiment, the present invention provides a hybrid open multi-agent shared task-oriented understanding establishment system and method based on a three-layer component architecture and hybrid scenario constraints. The specific implementation method is as follows.

[0154] In this embodiment, the system consists of five functional modules: a pre-component configuration module, an interactive execution module, a strategy optimization module, a constraint verification module, and a process termination determination module. These modules are connected via an internal data bus, share a unified memory space, and employ an event-driven mechanism for state synchronization and information transmission. All modules are deployed on the same computing node or a logical node in a distributed cluster, exchanging parameters and issuing control commands through standardized interfaces.

[0155] The preparatory component configuration module is initialized and run first. This module receives task description text input from the system administrator or the upper-level task planner and extracts the shared understanding target type, such as "collaborative assembly," "joint navigation," or "joint decision-making." Subsequently, the module generates a target completion threshold (e.g., overall shared understanding G ≥ 0.85) and a target priority list (e.g., attribute A weight 0.4, attribute B weight 0.3, etc.) according to preset rules, and encapsulates the above parameters into target component P1. Next, the module performs semantic parsing on the task description, identifies several core concepts (e.g., "tool," "location," "time"), and decomposes several attributes for each core concept (e.g., the attributes of "tool" include "type," "size," and "availability"). Each attribute is further bound to an initial value range (e.g., "type" ∈ {wrench, screwdriver, pliers}), thereby constructing semantic component P2. The generation process of signal component P3 depends on the capability list of the human-computer interaction device: if the system is equipped with a speech recognition and synthesis module, natural language is allowed; if equipped with motion capture equipment, a predefined set of body movements is allowed. The module binds at least one signal form to each attribute in P2; for example, the "location" attribute is bound to both voice coordinate description and gesture pointing. The prior shared understanding component P4 consists of three parts: basic language consensus (e.g., "left" refers to the operator's left side), core interaction rules (e.g., "only one attribute is discussed at a time"), and basic attribute consensus (e.g., "red represents danger"). The parameters of the components P1 to P4 are ultimately serialized and written to the system configuration file config.json, serving as the initial basis for subsequent module execution.

[0156] The interaction execution module starts after the pre-configured components are complete. This module first reads the complete list of core concept attributes from P2, sorts them by priority, and assigns them to each interaction round, generating the interaction task component I1. Each interaction round focuses on a single attribute; for example, the first round handles "tool-type," and the second round handles "location-coordinates." The interaction scenario component I2 determines the interaction mode based on the current task context: if the task scenario is being introduced for the first time and has no history, it is set to an open scenario, allowing free human input; if the task is already in progress or there is some consensus, it is set to a restricted scenario, guiding the human to choose from existing options. The initiator behavior component I3 is responsible for receiving human input signals, which are collected through the front-end interface (such as a microphone, touchscreen, or motion sensor) and sent to the module. The module first verifies whether the signal conforms to the physical conciseness rules defined by P3: natural language signals must be no longer than two sentences, and the vocabulary difficulty must be confirmed by a word frequency database to be no higher than the junior high school level; body movement signals must match entries in a predefined action template library. If not, the module returns a simplified prompt (such as "Please describe what tool you need in one sentence") and waits for re-entry. The responder behavior component I4 is executed by the AI ​​agent, with its input being the current shared understanding state (including agreed-upon attribute values ​​and unagreed-upon parts) and the valid signal output from I3. The AI ​​generates response signals based on its internal knowledge graph and task model: if it has a high degree of confidence in the signal's meaning, it outputs a solution signal containing a matching scheme and attribute quantification values ​​(e.g., "I will use a 10mm wrench"); if there is ambiguity, it outputs a clarification signal containing a specific question and no more than three options (e.g., "Do you mean a Phillips screwdriver or a flathead screwdriver? Options: A. Phillips B. Flathead C. Unsure"). The termination behavior component I5 continuously monitors the interaction termination conditions: termination is triggered when the human outputs acknowledgment such as "confirm" or "okay"; or automatically terminates when the task performance score P ≥ 90%; or when the human actively requests a pause; or when a single interaction lasts more than 5 minutes (recorded by a system timer). The task performance component I6 calculates the performance score P after each round of interaction, using the formula: Where R is the attribute value range ultimately accepted by humans (e.g., "wrench size ∈ {8mm, 10mm}"), S is the value in the AI ​​solution (e.g., "10mm"), and R_max is the upper limit of the initial value range of this attribute in P2 (e.g., if the size set is 5, then R_max = 5). The calculation results are stored in the interaction log database.

[0157] The strategy optimization module and the interaction execution module run in parallel, reading interaction log data in real time through shared memory. The goal-aware interaction component A1 first counts the number of concept attributes for which consensus has not yet been reached. With the total number of attributes Calculate the target correlation degree If C < 30%, it indicates that most attributes have reached a consensus. At this time, the module will switch the interaction focus to the remaining highly relevant attributes (i.e., attributes strongly related to the target task, whose weights are derived from the target priority mapping in P1) and notify the interaction execution module to adjust the allocation of I1 in the next round. Understanding update component A2 executes different update strategies based on the P value output by I6: if P ≥ 80%, it indicates high consistency, and S is directly stored in the shared value range library; if 50% ≤ P < 80%, it indicates partial consistency, and the value range is narrowed to the intersection of R and S (e.g., R = {8, 10, 12}, S = {10, 12}, then the new range is {10, 12}); if P < 50%, it indicates serious disagreement, the original range is retained, and three optional clarifications are inserted by the responder behavior component I4 before the next interaction. Understanding evaluation component A3 integrates the results of each round of interaction and calculates the overall shared understanding degree. , where w_i is the weight of the i-th attribute (obtained by normalization of the target priority in P1), and P_i is the task performance score corresponding to that attribute. If G ≥ the target completion threshold set in P1 (e.g., 0.85), a pre-termination signal is sent to the process termination determination module. The efficiency evaluation component A4 simultaneously monitors human cognitive load and interaction efficiency: it counts the number of signals output by humans in this round of interaction n_s (each independent voice or action is counted as 1) and the number of complex inferences n_c (determined by detecting interrogative sentences, conditional sentences, etc., through the natural language processing module), and substitutes them into the formula. Calculate cognitive load L, where L_max is the preset maximum load value (e.g., 10); interaction efficiency. Where N_c is the number of consensus attributes completed, and T_total is the total human participation time (in minutes). If L≥80% or E<0.5 (i.e., the number of consensus attributes per minute is less than 0.5), a simplification strategy is triggered, including shortening the signal length, adding option guidance, skipping low-priority attributes, etc., and the constraint verification module is notified to strengthen signal filtering.

[0158] The constraint verification module is embedded as middleware in all signal transmission paths. This module verifies three mixed scenario constraints in real time: First, the physical simplicity communication constraint requires all signals to be carried by text, voice, or body movements, prohibiting the use of pure numerical codes (such as "0101"); human output signals should not exceed two sentences per line and the vocabulary difficulty should not exceed the junior high school level (achieved through comparison with the built-in dictionary); AI output signals should not exceed two sentences per line, should not contain professional terminology (filtered through a terminology blacklist), and the number of visual elements (such as charts and icons) should not exceed five. Second, the task-goal consistency constraint requires that each round of I1 interaction task must strictly correspond to a core concept attribute in P2, and the goal relevance C ≥ 30%; at the same time, the module periodically calculates the Pearson correlation coefficient between the P value in each round of I6 and the G value in A3. If it is lower than 0.8, it is determined that there is a comprehension deviation, triggering a recalibration process. Third, the efficiency priority constraint requires that the total human participation interaction time T_total ≤ 15 minutes (monitored by the system's global timer), the proportion of invalid interactions (P < 50%) ≤ 20% (statistics through interaction logs), and the cognitive load L ≤ 80%. Once any constraint is violated, the module immediately intercepts the violation signal, returns a standardized simplified prompt to the initiator or responder, and records the violation event for subsequent analysis.

[0159] The process termination determination module continuously monitors the G value from the understanding evaluation component A3 and the compliance status from the constraint verification module. When G ≥ the target completion threshold and all efficiency priority constraints are met (i.e., T_total ≤ 15 minutes, invalid interaction ratio ≤ 20%, L ≤ 80%), the module terminates all interaction processes and serializes the final shared task-oriented understanding result (including consensus values ​​of all attributes, consensus establishment path, and performance indicators) to the task execution engine to drive subsequent human-machine collaborative actions. If the termination conditions are not met before T_total reaches 15 minutes, the module forcibly terminates and outputs the current best understanding result, while marking the status as "timeout termination" for system log recording.

[0160] In practical applications, such as in smart manufacturing workshops, human operators collaborate with multiple AI robots to complete equipment maintenance tasks. The system first loads a maintenance task template through the pre-configuration module, defining the goal as "accurately identifying faulty components and selecting appropriate tools," with a target completion threshold of 0.9 and "tool type" as the highest priority attribute. The semantic component P2 includes attributes such as "faulty component-location," "tool-type," and "operation-sequence." After the interaction begins, the human says, "I need a tool that can tighten M6 screws." This signal is verified by I3 to conform to the conciseness rule and then sent to I4. The AI, based on its knowledge base, determines that the desired tool might be a "5mm Allen wrench" or a "Phillips head screwdriver." Due to ambiguity, a clarification signal is output: "Do you need an Allen wrench or a Phillips head screwdriver? Options: A. Allen B. Phillips C. Other." After the human selects A, the P-value is calculated to be 1.0, and this consensus is stored in the shared library by A2. Subsequent interactions process attributes such as position and order sequentially. The strategy optimization module dynamically adjusts the focus based on the increase of the G value, while the constraint verification module ensures that each round of interaction does not exceed two sentences and the total duration is controlled within 12 minutes. Ultimately, G reaches 0.92, and the system outputs a complete shared understanding result, guiding the robot to accurately deliver tools and perform collaborative operations.

[0161] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0162] In the human-machine collaborative equipment maintenance task in the intelligent manufacturing workshop, the system first loads the "equipment maintenance" task description text through the pre-component configuration module, extracts the shared understanding target type as "accurately identify faulty parts and select appropriate tools", and sets the target completion threshold according to preset rules as overall shared understanding G≥0.9. At the same time, a target priority list is generated, in which the "tool-type" attribute weight is set to 0.5, the "faulty part-location" weight is 0.3, and the "operation-sequence" weight is 0.2, forming the target component P1. Subsequently, the system performs semantic parsing on the task text, identifies the three core concepts of "tool", "location", and "operation", and decomposes them into attribute structures: "tool" includes "type", "size", and "availability", "location" includes "coordinates" and "area identifier", and "operation" includes "step number" and "dependency". Each attribute is bound to an initial value range, for example, "tool-type" ∈ {hex wrench, Phillips screwdriver, flathead screwdriver, pliers}, thereby constructing the meaning component P2. Signal component P3 determines the permitted use of natural language and predefined gestures based on the interactive devices deployed in the workshop (including voice recognition microphones and gesture recognition cameras), and binds two signal forms, "voice reporting coordinates" and "finger pointing," to the "position-coordinates" attribute. Prior shared understanding component P4 loads the workshop's general consensus, such as "left / right is based on the operator's own reference," "only one tool attribute is discussed at a time," and "red labels indicate high-risk areas," and finally serializes and writes P1–P4.

[0163] Upon entering the interaction execution phase, the interaction task component I1 prioritizes and assigns the "tool-type" attribute in the first round. The interaction scenario component I2 determines that this is the first interaction and there is no history, so it is set as an open scenario. The human operator inputs a voice signal through the microphone: "I need a tool to tighten M6 screws." After receiving this signal, the initiator behavior component I3 calls the built-in junior high school word frequency dictionary to verify the vocabulary difficulty ("tighten," "M6," and "screw" are all high-frequency words) and checks that the number of sentences is one, which meets the physical conciseness rule. Therefore, it passes the signal to the responder behavior component I4. I4 uses the AI's internal knowledge graph (including a tool library corresponding to the M6 ​​screw standard) to retrieve two candidates: a 5mm Allen wrench (matching the ISO 2936 standard) or a PH2 Phillips screwdriver (if it is a self-tapping screw). Because the confidence level is below the threshold, I4 generates a clarification signal: "Do you need an Allen wrench or a Phillips screwdriver? Options: A. Allen B. Phillips C. Other." This signal is checked by the constraint verification module and confirms that the number of sentences is one, there are no technical terms, and the number of options is three, which meets the output specifications. After the human selects "A", the task performance component I6 calculates the P-value for this round: the human-accepted range R = {Hex wrench}, the AI ​​solution S = {Hex wrench}, and the initial range R_max = 4 (four tools). Therefore, P = |{Hex wrench}∩{Hex wrench}| / 4 = 1 / 4? This needs correction: According to the formula definition, R is the range of attribute values ​​ultimately accepted by the human. After clarification, the human explicitly chooses "Hex wrench", meaning R is a single-element set, while S is the correct option provided by the AI. It should actually be understood as the intersection coverage of R and S in the semantic space. More accurately, the system maps "Hex wrench" to a unique semantic identifier, R and S completely overlap, and R_max is the initial number of candidates, 4. However, the performance score is designed to measure consensus accuracy. Therefore, in the implementation, when the human confirms the only correct option from n options and the AI ​​provides that option, P is set to 1.0 to reflect complete consistency. This P=1.0 result is stored in the log.

[0164] The strategy optimization module runs synchronously. Understanding that update component A2, due to P ≥ 80%, writes "Tool-Type = Hex Wrench" to the shared value range library. Understanding evaluation component A3 calculates the current... The efficiency evaluation component A4 counts the number of human output signals in this round: n_s=1 (one voice message), no complex sentence structure (n_c=0), L=(1+0) / 10=0.1, interaction efficiency E=1 consensus attribute / 2 minutes=0.5, no simplification triggered. In the next round, I1 assigns "tool-size," the human says "5mm," I3 verifies compliance, I4 combines the "hex wrench" context to directly respond "use a 5mm hex wrench," the human replies "okay," I5 captures the approval feedback, and I6 calculates P=1.0. At this point... (Because "location" and "order" have not yet been processed), the target perception interaction component A1 counts 2 unconsensus attributes (location, order), and the total number of attributes N_total = 3. C = 2 / 3 ≈ 66.7% > 30%, so "faulty component - location" is processed sequentially. The human gestures to the second indicator light in the third row of the control panel. I3 matches the predefined "pointing" template through motion capture. I4, combined with the workshop layout diagram, identifies it as "Area B-3" and confirms via voice: "Is the fault in Area B-3?" The human answers "Yes," P = 1.0. At this point, N_unconsensus = 1 (only "operation - order" remains), and C = 1 / 3 ≈ 33.3%, still slightly higher than 30%, so processing continues. The human says, "Disconnect the power before removing the casing." I3 verifies compliance, and I4 parses the operational step dependencies, generating a solution: "Step 1: Turn off the main power; Step 2: Remove the four screws on the casing." The human confirms. At this point, all attributes reach consensus, and A3 calculates... The constraint verification module monitors the entire process: total duration T_total = 12 minutes ≤ 15 minutes, invalid interaction rounds 0 ≤ 20%, peak cognitive load L = 0.3 ≤ 80%, and all constraints are satisfied. The process termination judgment module outputs the final shared understanding result, including tool type and size, fault location, operation sequence, and consensus path, for the robot to perform precise delivery and collaborative disassembly.

[0165] In some different embodiments, this embodiment is a further optimization based on embodiment one, and is applied to a collaborative scenario between a travel consultant and a customer; the customer (human intelligent agent) needs to select a vacation plan that meets the criteria of "low cost", "convenient transportation" and "near hospital" through the travel consultant (artificial intelligence agent), and both parties need to establish shared understanding through natural language interaction.

[0166] First, configure the framework components. The specific configuration is as follows:

[0167] S1-1, Setting up the pre-component group:

[0168] P1: The goal is to establish a shared understanding of the criteria for selecting vacation packages, and ultimately to successfully select a vacation package that meets the client's needs.

[0169] P2: Core concepts include "low cost" (the price range acceptable to customers), "convenient transportation" (the travel mode and duration approved by customers), and "proximity to hospitals" (the distance or time standard required by customers).

[0170] P3: Uses natural language (Chinese) as the communication signal, supporting voice or text.

[0171] P4: Both parties have a grasp of basic Chinese vocabulary, understand the basic meaning of "vacation plan", and agree on the interactive process of "customer puts forward needs - consultant recommends plan - customer feedback".

[0172] S1-2, Set up the interactive component group:

[0173] I1: A single interaction focuses on establishing consensus on a core concept.

[0174] I2: Open-ended scenario, where the focus of each interaction is determined by the customer.

[0175] I3: Initial description or feedback from the client regarding the concept.

[0176] I4: The consultant recommends a solution based on their current understanding or asks for clarification.

[0177] I5: The client confirms the conceptual standard or proposes new adjustments.

[0178] I6: Consensus is reached using Boolean values.

[0179] S1-3, Configure the policy component group:

[0180] A1: Customers prioritize interacting with concepts for which there is no consensus.

[0181] A2: Consultants record client feedback and update their interpretation of various concepts.

[0182] A3: Count the number of concepts that have reached a consensus. When all three core concepts have reached a consensus, the standard is determined to be met.

[0183] A4: Use interaction time as the core cost indicator, and also record customer feedback response time.

[0184] The specific implementation process is as follows:

[0185] Client: "I need a low-cost vacation option that's conveniently located near a hospital; I don't like flying."

[0186] Consultant: "I recommend vacation destination B. It's directly accessible by high-speed rail (3 hours), priced at 5000 yuan per person, and is a 35-minute drive from the hospital."

[0187] Customer: "The price is acceptable, but the hospital is too far away; ideally, it should be within a 30-minute drive."

[0188] Task performance: "Convenient transportation" was agreed upon, "proximity to hospitals" was not agreed upon, and "low cost" was initially agreed upon.

[0189] Updated understanding: The consultant updated the "near hospital" standard to "within a 30-minute drive", and initially set "low cost" as "≤5000 yuan / person".

[0190] Subsequent interactions focused on the precise criteria for "near a hospital" and the specific price range for "low cost".

[0191] After each round of interaction, the consultant updates their understanding, the client refines their needs, until a consensus is reached on the three core concepts.

[0192] When all three concepts reach a consensus, the understanding evaluation component is deemed to have met the standard, and the interaction time is 12 minutes (efficiency evaluation meets the standard).

[0193] The consultant recommends the final solution, the client confirms acceptance, and a shared understanding is established.

[0194] All content not described in detail in this specification belongs to the prior art known to those skilled in the art, and the implementation of each software module can be completed based on conventional natural language processing libraries, knowledge graph engines and event-driven frameworks. The algorithm parameters involved in this technical solution (such as word frequency threshold and time threshold) can be adjusted according to specific application scenarios. The adjustment method is a conventional technical means in this field and will not be described in detail here.

[0195] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hybrid open multi-agent method with a three-layer component architecture and scene constraints, characterized in that, include: By configuring the target, meaning, signal, and prior shared understanding through the preparatory component group, a scalable initial semantic foundation is formed; Interactive components are used to perform interactive tasks that focus on a single conceptual attribute, completing requirement expression, solution feedback, and consensus verification under physically concise communication rules. By dynamically sensing target relevance, updating semantic understanding, assessing overall sharing level, and quantifying cognitive load through strategy component groups, the interaction process is driven to adaptively optimize. During the interaction process, a triple hybrid scenario constraint of physical concise communication, task-goal consistency and efficiency priority is imposed to ensure that the human-computer collaboration process meets the requirements of low cognitive load, high interaction efficiency and strong scenario adaptability. When the overall shared understanding reaches the preset target completion threshold and the interaction efficiency and cognitive load indicators meet the constraints, the interaction process is terminated, and the final shared task-oriented understanding result is output.

2. The hybrid open multi-agent method with a three-layer component architecture and scene constraints as described in claim 1, characterized in that, The process of configuring goals, meanings, signals, and prior shared understanding through preparatory component groups to form a scalable initial semantic foundation includes: Define shared understanding of target type, target completion threshold and target priority, and generate target component P1; The core concepts related to the task are broken down into a three-dimensional structure of concept-attribute-initial value range, and a meaningful component P2 is constructed. The signal forms that can be used by humans and AI are limited to natural language or body movements, and at least one adaptive signal form is bound to each concept attribute to generate signal component P3; Three categories of minimal prior content are set: basic language consensus, core interaction rules, and basic attribute consensus, which constitute the prior shared understanding component P4. The parameters of components P1 to P4 are written into the system configuration file as the initial basis for subsequent interactions and policy execution.

3. The hybrid open multi-agent method with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The process of executing a single interactive task focusing on a single conceptual attribute through an interactive component group, and completing requirement expression, solution feedback, and consensus verification under physically concise communication rules, includes: Based on the core concept attribute list in P2, assign a unique corresponding concept attribute to the current interaction round and generate the interaction task component I1. Determine whether the current interaction is an open-ended or restricted scenario, set the initial information input method, and generate the interaction scenario component I2; Receive natural language or action signals input by the initiator, verify whether they conform to the signal length and meaning uniqueness rules defined in P3, and if they do not conform, return a simplified prompt and generate the initiator behavior component I3; The respondent generates a solution signal or clarification signal based on its current understanding. The signal content includes a matching solution and attribute quantification value, or specific questions and no more than three options, generating the respondent behavior component I4. Monitor the termination conditions of the interaction, including the initiator outputting acknowledgment feedback, the task performance score being ≥90%, the initiator requesting a pause, or the duration of a single interaction exceeding 5 minutes, and generate the termination behavior component I5; The task performance score is calculated using the following formula: Where R is the attribute value range finally approved by the initiator, S is the attribute value of the responder's plan, R_max is the upper limit of the initial value range of the attribute in P2, and the task performance component I6 is generated.

4. The hybrid open multi-agent method with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The process of dynamically sensing target relevance, updating semantic understanding, evaluating overall sharing level, and quantifying cognitive load through strategy component groups to drive adaptive optimization of the interaction process includes: Calculate the ratio of the number of unconsensus concept attributes involved in the current interaction to the total number of attributes to obtain the target relevance C. If C < 30%, switch to high relevance attribute interaction and generate target-aware interaction component A1. Based on the task performance score P output by I6, if P≥80%, S is stored in the shared value range library; if 50%≤P<80%, the value range is narrowed to (R∩S); if P<50%, the original range is retained and three optional clarifications are inserted before the next interaction to generate the understanding update component A2. The overall shared understanding is calculated by combining the p-values ​​of each round of interaction with the corresponding attribute weights. If G≥P1 target completion threshold, the process terminates and an understanding evaluation component A3 is generated. Statistically calculate the number of human output signals n_s and the number of complex inferences n_c, and substitute them into the formula. Calculate the cognitive load L, and simultaneously calculate the interaction efficiency. If L≥80% or E<0.5, a simplification strategy is triggered, generating the efficiency evaluation component A4.

5. The hybrid open multi-agent method with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The constraint imposed during the interaction process, which combines physical concise communication, task-goal consistency, and efficiency priority, includes: All signals must be carried out in the form of text, voice or body movements, and purely digital encoding is prohibited; human output signals should not exceed two sentences per signal and the vocabulary difficulty should not be higher than the junior high school level; AI output signals should not exceed two sentences per signal, should not contain professional terminology, and should not have more than five visual elements. Each round of I1 interactive task must correspond to a core concept attribute in P2, and the target relevance C ≥ 30%; the correlation coefficient between the I6 task performance score and the overall shared understanding G is calculated using the Pearson formula and is required to be ≥ 0.

8. The total human interaction time T_total ≤ 15 minutes, the proportion of invalid interaction rounds ≤ 20%, and the cognitive load L ≤ 80%.

6. A hybrid open multi-agent system with a three-layer component architecture and scene constraints, characterized in that: The system includes: The preparatory component configuration module is used to define the target, meaning, signal and prior shared understanding, and generate the parameters of components P1-P4; The interaction execution module is used to control the single-round interaction task, scenario, behavior and performance evaluation, and generate I1-I6 component outputs; The strategy optimization module is used to dynamically adjust the interaction focus, update semantic understanding, evaluate the overall level, and adjust cognitive load to generate A1-A4 component decisions. The constraint verification module is used to monitor the triple constraints of physical simplification communication, task-goal consistency, and efficiency priority in real time, intercept violation signals, and trigger simplification mechanisms. The process termination determination module is used to output the final shared task-oriented understanding result when the overall shared understanding meets the standard and the constraints are satisfied.

7. The hybrid open multi-agent system with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The pre-configuration component module is configured as follows: Receive the task description text, extract the shared understanding of the target type, target completion threshold and target priority, and generate the target component P1; Semantic parsing of the task description is performed to identify core concepts and decompose them into a three-dimensional structure of concept-attribute-initial value range, and to construct the meaning component P2; Based on the capability list of human-computer interaction devices, bind at least one signal form to each attribute to generate signal component P3; Establish consensus on basic language, core interaction rules, and basic attributes to form the prior shared understanding component P4; The parameters of components P1 to P4 are serialized and written to the system configuration file.

8. The hybrid open multi-agent system with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The interactive execution module is configured as follows: Read the core concept attribute list from P2, sort it by priority, and assign it to the interaction rounds one by one to generate the interaction task component I1; Based on the task context, determine whether the interaction mode is an open scene or a limited scene, and generate the interaction scene component I2; Receive human input signals, verify whether they conform to the signal length and vocabulary difficulty rules, and if they do not conform, return a simplified prompt and generate the initiator behavior component I3; The AI ​​agent generates a solution signal or clarification signal based on its current understanding, and then generates the responder behavior component I4. Monitor interaction termination conditions, including positive feedback, task performance score ≥90%, pause request or single interaction timeout of 5 minutes, and generate end behavior component I5; Calculate task performance score Generate task performance component I6.

9. The hybrid open multi-agent system with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The strategy optimization module is configured as follows: Calculate the target correlation If C < 30%, switch to high-relevance attribute interaction and generate target-aware interaction component A1; Based on the task performance score P, a semantic update strategy is executed: when P≥80%, the value is stored in the shared value range library; when 50%≤P<80%, the intersection is taken; when P<50%, the original range is retained and three optional clarifications are inserted to generate the understanding update component A2. Calculate overall shared understanding If G ≥ the target completion threshold, a termination preparation signal is sent and an understanding evaluation component A3 is generated. Calculate cognitive load With interaction efficiency If L≥80% or E<0.5, a simplification strategy is triggered, generating the efficiency evaluation component A4.

10. A hybrid open multi-agent system with a three-layer component architecture and scene constraints as described in claim 1, characterized in that: The constraint verification module is configured as follows: Verify that all signals are carried by text, voice, or body movements, and prohibit purely digital encoding; verify that each human signal is no more than two sentences and the vocabulary difficulty is no higher than junior high school level; verify that each AI signal is no more than two sentences, contains no professional terminology, and has no more than five visual elements. Verify whether each round of interactive tasks corresponds to a core concept attribute in P2 and whether the target relevance C ≥ 30%; verify whether the Pearson correlation coefficient between the task performance score and the overall shared understanding is ≥ 0.8; Verify that the total human interaction time T_total ≤ 15 minutes, the percentage of invalid interaction rounds ≤ 20%, and the cognitive load L ≤ 80%; if any constraint is violated, intercept the signal and return a simplified prompt.

Images