Multi-agent confrontation cooperation and iterative generation method and device

Abstract

The application provides a kind of multi-agent confrontation cooperation and iterative generation method and device, it is related to artificial intelligence technical field, wherein the method comprises: based on the task demand information of target task, determine the heterogeneous intelligent agent cluster of executing target task;Trigger the multi-round confrontation cooperation iteration of heterogeneous intelligent agent cluster to target task in shared semantic environment, and in the case where confrontation cooperation iteration reaches preset convergence condition, based on the record information in shared semantic environment, determine the output result of target task.Through triggering heterogeneous intelligent agent cluster to carry out multi-round confrontation cooperation iteration in shared semantic environment, the structured confrontation and complementary cooperation between different role agents are utilized, and self-determination high-quality output in complex innovation task is realized.Based on the record information in shared semantic environment, the output result is synthesized, to ensure that the final output fully integrates multi-dimensional reasoning results, and the reasoning quality of complex task is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for multi-agent adversarial cooperation and iterative generation. Background Technology

[0002] With the rapid development of artificial intelligence technology, complex task reasoning and content generation based on large language models has become one of the core research directions in this field. In practical applications, complex reasoning tasks such as designing technological innovation solutions and formulating business strategies often involve balancing multi-dimensional constraints, stimulating creative thinking, and rigorous logical reasoning, which places extremely high demands on the comprehensive ability of task reasoning.

[0003] Existing technologies for reasoning in complex tasks typically involve forming clusters of multiple agents to generate viewpoints or solutions for the same task. The outputs of each agent are then integrated through voting, fusion, or other methods. Because different agents produce different results, a collaborative adversarial approach can be used to obtain the reasoning outcome. However, in existing multi-agent collaborative schemes, agents often employ the same or similar cognitive strategies and decision-making logic, lacking effective structured adversarial and differentiated cognitive driving mechanisms. This leads to highly convergent semantic viewpoints among agents, resulting in significant homogenization of output content and ultimately, low overall quality of the output. Summary of the Invention

[0004] This invention provides a method and apparatus for multi-agent adversarial cooperation and iterative generation, which can improve the output quality of the reasoning process in complex tasks.

[0005] This invention provides a method for adversarial cooperation and iterative generation of multiple agents, comprising the following steps: Based on the task requirements information of the target task, a heterogeneous intelligent agent cluster is determined to execute the target task. The heterogeneous intelligent agent cluster includes multiple intelligent agents with different role types. Each intelligent agent is configured with cognitive bias parameters and decision parameters corresponding to its role type. The cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model. The decision parameters are the maximum inference steps parameter, search width parameter, and confidence threshold parameter in the agent's large language model. The heterogeneous intelligent agent cluster is triggered to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment. When the adversarial cooperative iteration reaches a preset convergence condition, the output result of the target task is determined based on the recorded information in the shared semantic environment. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0006] According to a multi-agent adversarial cooperation and iterative generation method provided by the present invention, the shared semantic environment further includes an information subset of each agent in the heterogeneous agent cluster, wherein the information subset records the role description information of the agents, and each round of adversarial cooperation iteration includes: Based on the role type of each agent in the heterogeneous agent cluster, the target agent to be discussed first is determined from the heterogeneous agent cluster; The information subset is loaded into the target agent, and the initial viewpoint of the target agent with the loaded information subset on the target task is obtained. If the quality assessment result of the initial viewpoint meets the preset quality conditions, the initial viewpoint is written into the shared semantic environment to obtain the updated shared semantic environment. Obtain response viewpoints and write them into the updated shared semantic environment. The response viewpoints are viewpoints generated by other agents using adversarial strategies against the initial viewpoints. The other agents are agents in the heterogeneous agent cluster other than the target agent.

[0007] According to the present invention, a multi-agent adversarial cooperation and iterative generation method is provided, wherein the adversarial strategy includes at least one of an alternative generation strategy, a hypothesis challenge strategy, a constraint relaxation strategy, and a perspective switching strategy. The alternative generation strategy is used to generate at least one alternative for the initial viewpoint; The hypothesis challenge strategy is used to identify implicit hypothesis information in the initial viewpoint and to raise questions about the implicit hypothesis information; The constraint relaxation strategy is used to reduce the constraints of the target task and generate new candidate viewpoints based on the reduced constraints. The perspective shifting strategy is used to re-evaluate the initial viewpoint and generate evaluation opinions based on agents in the heterogeneous agent cluster whose role categories are different from those of the target agent.

[0008] According to the multi-agent adversarial cooperation and iterative generation method provided by the present invention, determining that the adversarial cooperation iteration reaches a preset convergence condition includes: After each iteration, the current state data in the shared semantic environment is extracted; Based on the current state data, convergence index information in multiple different dimensions is determined; The convergence index information of the multiple different dimensions is weighted and fused to obtain the convergence score of the current state data; If the convergence score is greater than a preset score threshold, it is determined that the adversarial cooperative iteration has reached the preset convergence condition.

[0009] According to the multi-agent adversarial cooperation and iterative generation method provided by the present invention, the convergence index information of the multiple different dimensions includes the viewpoint consistency index, the innovation stagnation index, the constraint satisfaction index, and the discussion efficiency index. The opinion consistency index is determined based on the variance of the quality assessment results of the opinions generated by each agent in the current round; the innovation stagnation index is determined based on the change in the number of new innovative opinions between the current round and the previous round; the constraint satisfaction index is determined based on the proportion of opinions in the current shared semantic environment that satisfy the constraints of the target task; and the discussion efficiency index is determined based on the proportion of effective opinion rounds to the total number of iteration rounds.

[0010] According to the multi-agent adversarial cooperation and iterative generation method provided by the present invention, the step of determining the output result of the target task based on the recorded information in the shared semantic environment includes: Traverse each node in the graph structure of the shared semantic environment and determine the node that satisfies the preset consensus condition as the target viewpoint node. The edges of the target viewpoint node in the graph structure are determined, and the information of the target viewpoint node is concatenated according to the logical relationship of the edges to obtain the output result of the target task.

[0011] The present invention also provides a multi-agent adversarial cooperation and iterative generation device, comprising the following modules: The agent cluster construction module is used to determine a heterogeneous agent cluster to execute the target task based on the task requirement information of the target task. The heterogeneous agent cluster includes multiple agents with different role types. Each agent is configured with cognitive bias parameters and decision parameters corresponding to its role type. The cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model. The decision parameters are the maximum inference steps parameter, search width parameter, and confidence threshold parameter in the agent's large language model. The iterative execution module is used to trigger the heterogeneous intelligent agent cluster to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment, and to determine the output result of the target task based on the recorded information in the shared semantic environment when the adversarial cooperative iteration reaches a preset convergence condition. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-agent adversarial cooperation and iterative generation method as described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-agent adversarial cooperation and iterative generation method as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-agent adversarial cooperation and iterative generation method as described above.

[0015] The multi-agent adversarial cooperation and iterative generation method and apparatus provided by this invention dynamically determines a heterogeneous agent cluster based on the task requirements information of the target task. Each agent is configured with differentiated cognitive bias parameters and decision parameters corresponding to its role type, enabling multiple agents to analyze the same task from different cognitive dimensions. This effectively avoids the problem of a single model getting trapped in local optima due to path dependence. By triggering the heterogeneous agent cluster to conduct multiple rounds of adversarial cooperative iteration in a shared semantic environment, and utilizing structured adversarial and complementary cooperation between agents with different roles, autonomous high-quality output is achieved in complex and innovative tasks. The output results are synthesized based on the recorded information in the shared semantic environment, ensuring that the final output fully integrates multi-dimensional reasoning results and significantly improving the reasoning quality of complex tasks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the multi-agent adversarial cooperation and iterative generation method provided by the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of the multi-agent adversarial cooperation and iterative generation device provided by the present invention.

[0019] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] This invention provides a method for adversarial cooperation and iterative generation of multiple agents. Figure 1 This is a flowchart illustrating the multi-agent adversarial cooperation and iterative generation method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 110: Based on the task requirement information of the target task, determine the heterogeneous intelligent agent cluster to execute the target task; the heterogeneous intelligent agent cluster includes multiple intelligent agents with different role types, and each intelligent agent is configured with cognitive bias parameters and decision parameters corresponding to its role type; the cognitive bias parameters are the temperature parameter, kernel sampling parameter, candidate lexical number parameter, frequency penalty parameter, existence penalty parameter, and lexical-level log probability bias parameter in the intelligent agent's large language model; the decision parameters are the maximum inference step parameter, search width parameter, and confidence threshold parameter in the intelligent agent's large language model; Step 120: Trigger the heterogeneous intelligent agent cluster to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment, and determine the output result of the target task based on the recorded information in the shared semantic environment when the adversarial cooperative iteration reaches the preset convergence condition. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0022] The execution subject of the multi-agent adversarial cooperation and iterative generation method provided by this invention can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, a mobile electronic device can be a mobile phone, tablet computer, laptop computer, PDA, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc., while a non-mobile electronic device can be a server, network attached storage (NAS), or personal computer (PC), etc. This invention does not impose specific limitations.

[0023] The following section uses the example of a computer executing the multi-agent adversarial cooperation and iterative generation method provided by this invention to illustrate the technical solution of this invention in detail.

[0024] In step 110, based on the task requirement information of the target task, a heterogeneous intelligent agent cluster to execute the target task is determined.

[0025] The target task can be any task requiring innovative thinking, complex logical reasoning, or multi-dimensional trade-offs, such as designing technological innovation solutions, formulating business strategies, or creating creative content. Task requirement information refers to the collection of information extracted from the target task, describing its type characteristics, domain attributes, complexity level, constraints, etc. For example, when the target task is "designing a battery management system for next-generation electric vehicles," the task requirement information could include the technical field to which the task belongs, the constraints of the task, and the professional perspective required for the task.

[0026] In multi-agent systems, agent clusters can be classified into homogeneous agent clusters and heterogeneous agent clusters based on the homogeneity and heterogeneity of their constituent members. A homogeneous agent cluster refers to a set of agents in the cluster that have the same role positioning, the same capability configuration, and the same behavioral strategy. Its characteristics are that the agents are functionally equivalent and substitutable.

[0027] The heterogeneous agent cluster in this invention refers to a collection of agents with different role types, different cognitive ability configurations, and different decision-making tendencies. The agents in the heterogeneous agent cluster have significant differences in their functional roles. This heterogeneity simulates the composition of human professional teams with different professional backgrounds, thinking styles, and functional roles.

[0028] Specifically, the types of roles in a heterogeneous intelligent agent cluster can include, but are not limited to: Innovator role: an exploratory thinking-oriented role that focuses on concept expansion, analogical reasoning, and breakthrough association; Critics role: a conservative thinking-oriented role that focuses on logical testing, vulnerability discovery, and constraint verification; Coordinator role: a systematic thinking-oriented role that focuses on conflict mediation, consensus refinement, and progress management; Domain expert role: a professional thinking-oriented role that focuses on the application of professional knowledge and standard compliance checks.

[0029] The core value of heterogeneous intelligent agent clusters lies in the fact that intelligent agents of different roles can analyze and reason about the same task from different cognitive dimensions, thereby forming multi-dimensional cognitive coverage at the cluster level and avoiding the limitations of a single perspective.

[0030] Optionally, the determination process for the heterogeneous agent cluster can be dynamically generated using a role factory module. The role factory module can automatically parse the required role types and quantities based on task requirements, and configure differentiated cognitive bias parameters and decision parameters for each agent through a capability configurator.

[0031] In this invention, to achieve differentiated behavior among agents in a heterogeneous agent cluster, it is necessary to configure cognitive bias parameters and decision parameters corresponding to the role type of each agent. The specific settings of the cognitive bias parameters and decision parameters are achieved by adjusting the relevant parameters of the agent's underlying large language model.

[0032] It should be noted that intelligent agents are all based on large language models as their core capability components. Each intelligent agent contains a large language model for task processing. Cognitive bias parameters are parameters in the intelligent agent's large language model used to control the model's output and information processing tendencies, while decision parameters are parameters in the intelligent agent's large language model used to control the model's decision-making.

[0033] Specifically, the cognitive bias parameters may include temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters.

[0034] Specifically, the temperature parameter is a scaling factor used by large language models to adjust the smoothness of the output probability distribution when performing softmax sampling. For example, in a heterogeneous agent cluster, agents playing the innovator role are configured with a higher temperature parameter to promote divergent thinking and breakthrough associations, while agents playing the critic role are configured with a lower temperature parameter to ensure the rigor of logical reasoning and the determinism of the output.

[0035] The kernel sampling parameter (Top-p) is a cumulative probability threshold used by large language models during the sampling phase to control the range of candidate lexical units. A larger kernel sampling parameter allows the model to sample from a wider range of candidate lexical units, resulting in greater output diversity. Conversely, a smaller kernel sampling parameter allows the model to sample only from the few candidate lexical units with the highest probabilities, leading to a more focused and predictable output. For example, in a heterogeneous agent cluster, agents with innovator roles are configured with larger kernel sampling parameters to expand the exploration space, while agents with domain expert roles are configured with smaller kernel sampling parameters to ensure that the output is focused within their domain knowledge.

[0036] The Top-k parameter is used by large language models during the sampling phase to limit the upper limit of the number of candidate lexical units participating in the sampling. It can control the diversity and concentration of the model's output. A larger Top-k value allows the model to consider more low-probability but potentially innovative candidate lexical units, while a smaller Top-k value allows the model to focus on high-probability deterministic outputs.

[0037] The frequency penalty parameter is a penalty coefficient used by large language models during the generation process to reduce the probability of a previously encountered word being selected again. A higher frequency penalty parameter indicates stronger suppression of repetitive content by the model, making it more inclined to generate content containing novel expressions and diverse concepts. For example, in a heterogeneous agent cluster, the innovator agent is configured with a higher frequency penalty parameter to avoid generating repetitive content and promote the introduction of new concepts, while the coordinator agent is configured with a moderate frequency penalty parameter to achieve a balance between content novelty and expression coherence.

[0038] The existence penalty parameter is a bias coefficient used by large language models during the generation process to encourage or inhibit the introduction of new topics. A larger existence penalty parameter value indicates that the model is more inclined to introduce new topics and concepts that have not yet been discussed into the output; a smaller existence penalty parameter value indicates that the model is more inclined to delve deeper into existing topics. For example, in a heterogeneous agent cluster, agents playing the innovator role are configured with a higher existence penalty parameter to stimulate cross-domain associations and the introduction of new concepts, while agents playing the critic role are configured with a lower existence penalty parameter to ensure in-depth logical analysis of existing viewpoints rather than frequent topic jumps.

[0039] Lexical-level log-probability bias parameters are parameters that directly apply a bias to the log-probability value of a specific lexical during the generation process in a large language model. By setting positive or negative bias values ​​for key lexical terms related to specific role functions, the agent can be guided to tend to or avoid using specific types of expressions during the generation process. For example, setting positive bias values ​​for lexical terms related to questioning, rebutting, and risk for an agent with a critic role can enhance its ability to express critical analysis.

[0040] Specifically, the decision parameters may include the maximum number of inference steps, the search width, and the confidence threshold.

[0041] Specifically, the maximum inference steps parameter controls the upper limit of the chain-like reasoning steps performed by the large language model when generating a single viewpoint. A larger value for the maximum inference steps parameter results in a longer and more in-depth reasoning chain in the large language model; a smaller value results in a simpler reasoning process. For example, in a heterogeneous agent cluster, the critic agent is configured with a larger maximum inference steps parameter to ensure in-depth logical deconstruction and vulnerability detection of the initial viewpoint, while the innovator agent is configured with a moderate maximum inference steps parameter to achieve a balance between reasoning depth and divergent thinking.

[0042] The search width parameter is used by large language models to control the number of candidate sequences retained in parallel when employing a bundle search decoding strategy. A larger search width parameter allows the large language model to consider more candidate solutions simultaneously during generation, increasing the likelihood of finding the globally optimal expression; a smaller value results in higher generation efficiency but may miss better candidates. For example, in a heterogeneous agent cluster, an innovator agent can be configured with a larger search width parameter to search for innovative expressions in a larger candidate space, while a critic agent can be configured with a smaller search width parameter to quickly generate focused critical analyses.

[0043] The confidence threshold parameter controls the minimum confidence level required by a large language model to accept or reject a viewpoint. When the large language model's internal confidence level in a viewpoint or fact falls below the threshold, it will challenge the viewpoint or request supplementary evidence; when the confidence level is above the threshold, it will accept the viewpoint. For example, in a heterogeneous agent cluster, the critic agent might configure a higher confidence threshold parameter to increase its standard for requiring sufficient evidence, while the innovator agent might configure a lower confidence threshold parameter to be more tolerant of unconventional viewpoints and innovative hypotheses that are not fully validated.

[0044] Optionally, the required combination of role types can be determined based on the task type and domain attributes in the task requirement information; and the number of agents required for each role type can be determined based on the complexity level in the task requirement information.

[0045] In step 120, the heterogeneous intelligent agent cluster is triggered to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment. When the adversarial cooperative iteration reaches the preset convergence condition, the output result of the target task is determined based on the recorded information in the shared semantic environment.

[0046] A shared semantic environment is pre-constructed. A shared semantic environment refers to a globally visible information space used to record the interaction information of various agents in a heterogeneous agent cluster during the adversarial and collaborative iterative process. The shared semantic environment is equivalent to a dynamically evolving shared cognitive space. All agents in the heterogeneous agent cluster can read and write information within this shared semantic environment, thereby achieving information sharing and interaction.

[0047] The shared semantic environment is specifically an information space constructed based on a graph structure. Nodes in the graph structure represent conceptual information, opinion information, or factual information. Specifically, conceptual information refers to abstract concepts or terms related to the target task; opinion information refers to the analytical opinions, suggested solutions, or evaluation conclusions proposed by each agent regarding the target task; and factual information refers to objective data or verified knowledge related to the target task.

[0048] In a graph structure, edges represent the logical relationships between nodes. A logical relationship refers to the type of semantic connection between the information carried by a node. Through these logical relationships, a shared semantic environment can structurally organize and associate the interactive information generated by various agents, forming a semantic network.

[0049] Optionally, nodes in the graph structure can also include node attributes, which can be determined based on confidence (representing the reliability of the node's information), novelty (representing the originality of the node's information), and importance weight (representing the importance of the node's information to the target task). Node attributes can be dynamically updated based on the evaluations and feedback from each agent during the adversarial cooperative iteration process.

[0050] The state of a shared semantic environment can be represented as S(t) = {G(t), M(t), C(t)}, where G(t) is the graph structure at the current moment; M(t) is metadata used to record discussion process information; and C(t) is the consensus set used to record the core viewpoints that have reached a consensus. The state of the shared semantic environment dynamically evolves as the adversarial collaboration iterates, and its evolution function can be represented as S(t+1) = f(S(t), A_i(t), R), where A_i(t) is the action of agent i at time t, and R is the environment evolution rule.

[0051] In a heterogeneous intelligent agent cluster, the agents engage in adversarial and collaborative iteration through multiple rounds of viewpoint generation, critical analysis, rebuttal argumentation, and iterative correction within a shared semantic environment, gradually advancing the analysis and solution of the target task.

[0052] In each round of adversarial collaborative iteration, multiple agents generate new viewpoints or evaluate and revise existing viewpoints based on information in the current shared semantic environment, and write the results into the shared semantic environment, thereby updating the state of the shared semantic environment.

[0053] High-quality results can only be gradually achieved through multiple rounds of repeated interaction. Each iteration is based on in-depth analysis and improvement of the discussion results of the previous rounds, gradually refining the solution.

[0054] When the adversarial collaborative iteration reaches a preset convergence condition, the output of the target task is determined based on the information recorded in the shared semantic environment. The preset convergence condition can be set based on various factors, such as a threshold based on the number of iteration rounds, an evaluation based on the quality of discussion, or a judgment based on the degree of convergence.

[0055] The output result refers to the final output scheme for the target task obtained after integrating and refining all the interaction information recorded in the shared semantic environment, based on the preset convergence conditions of the adversarial collaboration iteration.

[0056] Understandably, a shared semantic environment built on a graph structure can efficiently complete the sorting process when sorting out the output results.

[0057] The multi-agent adversarial collaboration and iterative generation method provided by this invention dynamically determines a heterogeneous agent cluster based on the task requirements information of the target task. Each agent is configured with differentiated cognitive bias parameters and decision parameters corresponding to its role type, enabling multiple agents to analyze the same task from different cognitive dimensions. This effectively avoids the problem of a single model getting trapped in local optima due to path dependence. By triggering multiple rounds of adversarial collaborative iterations within a shared semantic environment, and utilizing structured adversarial and complementary collaboration between agents with different roles, autonomous high-quality output is achieved in complex and innovative tasks. The output results are synthesized based on the recorded information in the shared semantic environment, ensuring that the final output fully integrates multi-dimensional reasoning results and significantly improving the reasoning quality for complex tasks.

[0058] In one embodiment, the shared semantic environment further includes a subset of information about each agent in the heterogeneous agent cluster, the subset of information recording the role description information of the agents, and each round of adversarial-cooperative iteration process includes: Based on the role type of each agent in the heterogeneous agent cluster, the target agent to be discussed first is determined from the heterogeneous agent cluster; The information subset is loaded into the target agent, and the initial viewpoint of the target agent with the loaded information subset on the target task is obtained. If the quality assessment result of the initial viewpoint meets the preset quality conditions, the initial viewpoint is written into the shared semantic environment to obtain the updated shared semantic environment. Obtain response viewpoints and write them into the updated shared semantic environment. The response viewpoints are viewpoints generated by other agents using adversarial strategies against the initial viewpoints. The other agents are agents in the heterogeneous agent cluster other than the target agent.

[0059] The shared semantic environment contains subsets of information about each agent in the heterogeneous agent cluster. These subsets include role descriptions of the agents. Role descriptions are instructional texts written in natural language that define an agent's identity, behavioral guidelines, and task response methods. These descriptions are pre-written into the corresponding subset when an agent is created and joined to the heterogeneous agent cluster. During the adversarial and collaborative iteration process, each time an agent reads information from the shared semantic environment, it loads its own role description from its subset to ensure that its subsequent viewpoint generation and decision-making actions remain consistent with its role positioning.

[0060] For example, for an agent playing the role of a critic, the role description information recorded in its information subset could be: "Your role is that of a rigorous critical reviewer. When you receive a task, you should first examine whether the logical chain of existing viewpoints in the current shared semantic environment is complete, identify the implicit assumptions and potential loopholes, and raise questions in a well-supported manner; you should not easily agree with viewpoints that have not been fully verified, and your core responsibility is to improve the rigor and reliability of the solution through structured critical analysis." For an agent playing the role of an innovator, the role description information recorded in its information subset could be: "Your role is that of a creative solution explorer. When you receive a task, you should prioritize directions that have not yet been explored in the existing discussions and unconventional associative paths, try to break through the existing framework of thought for solutions, and generate novel and groundbreaking candidate viewpoints; you should actively expand the boundaries of the solution space, and even if some viewpoints seem impractical, they are worth proposing as seeds of thought to stimulate further discussion."

[0061] By recording role description information in the information subset, each agent can maintain consistency and stability of its role behavior throughout the entire process of adversarial collaboration and iteration, preventing agents from deviating from their preset role positioning during multiple rounds of interaction. This ensures that the differentiated cognitive advantages of the heterogeneous agent cluster can continue to play a role throughout the entire iteration process.

[0062] The target agent can be the first agent to express its opinion on the target task in the current round of adversarial collaboration iteration. The target agent is determined based on the agent's role type, and in the specific discussion, the first agent to speak can be dynamically configured according to the actual discussion needs.

[0063] The target agent generates an initial viewpoint for the target task based on information in the current shared semantic environment, combined with the cognitive bias parameters and decision parameters configured according to its own role type.

[0064] The specific process of generating the initial viewpoint involves the target agent first reading a subset of information relevant to its role from the shared semantic environment. For example, an innovator role might focus more on information related to new concepts and breakthrough solutions, while a critic role might focus more on information related to logical flaws and risk factors in existing solutions. Then, the target agent processes and analyzes the read information based on its cognitive bias parameters and decision parameters to generate the initial viewpoint.

[0065] If the quality assessment results of the initial viewpoint meet the preset quality conditions, the initial viewpoint is written into the shared semantic environment to obtain the updated shared semantic environment.

[0066] Quality assessment can be conducted from multiple dimensions, such as innovation, logic, and practicality. The quality assessment result can be obtained by weighting the scores of each dimension. For example, the score of the target intelligent agent P_i can be determined as: Q(P_i) = w1 × innovation + w2 × logic + w3 × practicality, where w1, w2, and w3 are the weights of each dimension.

[0067] Preset quality conditions refer to pre-defined threshold conditions used to determine whether an initial viewpoint is of sufficient quality to be included in the shared semantic environment. When the quality assessment result of an initial viewpoint meets the preset quality conditions, it indicates that the initial viewpoint has sufficient discussion value and can be included in the shared semantic environment. This filtering mechanism using preset quality conditions prevents low-quality viewpoints from entering the shared semantic environment and affecting discussion efficiency.

[0068] The process of writing the initial viewpoint into the shared semantic environment can specifically involve creating new nodes in the graph structure of the shared semantic environment to represent the information of the initial viewpoint, and establishing edges between the new node and existing nodes in the graph structure to represent the logical relationship between the initial viewpoint and existing information, thereby obtaining the updated shared semantic environment.

[0069] Obtain the response viewpoint and write the response viewpoint into the updated shared semantic environment.

[0070] A response viewpoint refers to a viewpoint generated by other agents using adversarial strategies in response to an initial viewpoint. Other agents refer to agents in a heterogeneous agent cluster other than the target agent. After the target agent's initial viewpoint is written into the shared semantic environment, other agents can read this initial viewpoint and, based on their respective role-type configured cognitive bias parameters and decision parameters, respond to the initial viewpoint using adversarial strategies, generating a response viewpoint.

[0071] Adversarial strategies refer to strategies and methods used to stimulate structured conflict and creative thinking among agents. Adversarial strategies are not simply about negation, but rather about conducting in-depth analysis of initial viewpoints in a purposeful and structured manner to promote deeper discussion and optimization of solutions.

[0072] The response viewpoints are written into the updated shared semantic environment, which further enriches and updates the graph structure of the shared semantic environment, providing a more comprehensive information foundation for the next round of adversarial collaboration iterations.

[0073] In one embodiment, the adversarial strategy includes at least one of an alternative generation strategy, a hypothesis challenge strategy, a constraint relaxation strategy, and a perspective shift strategy; The alternative generation strategy is used to generate at least one alternative for the initial viewpoint; The hypothesis challenge strategy is used to identify implicit hypothesis information in the initial viewpoint and to raise questions about the implicit hypothesis information; The constraint relaxation strategy is used to reduce the constraints of the target task and generate new candidate viewpoints based on the reduced constraints. The perspective shifting strategy is used to re-evaluate the initial viewpoint and generate evaluation opinions based on agents in the heterogeneous agent cluster whose role categories are different from those of the target agent.

[0074] The alternative solution generation strategy is used to generate at least one alternative to the initial viewpoint. Instead of directly rejecting the initial viewpoint, the strategy provides feasible alternatives that differ from the initial viewpoint for comparison and reference.

[0075] Hypothesis-challenging strategies are used to identify implicit assumptions in an initial viewpoint and to raise questions about these assumptions. Implicit assumptions refer to presuppositions in the initial viewpoint that are not explicitly stated but exist as premises.

[0076] Constraint relaxation strategies are used to reduce the constraints on a target task and generate new candidate ideas based on the reduced constraints. The core idea of ​​constraint relaxation strategies is to temporarily relax some constraints on the target task to explore possible solutions under more relaxed conditions, thereby expanding the space for innovative thinking.

[0077] Perspective shifting strategies are used to re-evaluate initial viewpoints and generate evaluation opinions based on an agent with a different role category than the target agent that proposed the initial viewpoint. The core idea of ​​perspective shifting strategies is to re-examine the same initial viewpoint from different role perspectives and stakeholder perspectives in order to discover problems and opportunities that may be overlooked from a single perspective.

[0078] It should be noted that the four adversarial strategies described above can be used individually or in combination. In actual adversarial collaborative iterations, other agents can select one or more of these strategies to generate response viewpoints based on the current discussion state and their own role type.

[0079] In one embodiment, determining that the adversarial cooperative iteration has reached a preset convergence condition includes: After each iteration, the current state data in the shared semantic environment is extracted; Based on the current state data, convergence index information in multiple different dimensions is determined; The convergence index information of the multiple different dimensions is weighted and fused to obtain the convergence score of the current state data; If the convergence score is greater than a preset score threshold, it is determined that the adversarial cooperative iteration has reached the preset convergence condition.

[0080] Current state data refers to all the information contained in the shared semantic environment after the completion of the current round of adversarial collaboration iteration. Current state data can include all node information, all edge information, metadata, and consensus set in the graph structure of the shared semantic environment.

[0081] The process of extracting current state data can be performed by a dedicated environment state tracker module. The environment state tracker module monitors changes in the shared semantic environment in real time and extracts current state data after each iteration, providing a data basis for subsequent convergence judgments.

[0082] Based on the current state data, multiple convergence metrics are determined across different dimensions. These metrics reflect the progress of the discussion from various perspectives, making the convergence assessment more accurate.

[0083] The convergence score of the current state data is obtained by weighted fusion of convergence metrics from multiple dimensions. The weights of each dimension can be pre-set according to the type and requirements of the target task, or can be adaptively adjusted through machine learning methods.

[0084] If the convergence score is greater than a preset score threshold, it is determined that the adversarial cooperative iteration has reached the preset convergence condition.

[0085] Optionally, convergence criteria can also be combined with an upper limit threshold for the number of iterations. Even if the convergence score has not yet reached the preset score threshold, the iteration process can be terminated if the number of iterations has reached the preset maximum limit, in order to avoid endless and invalid discussions.

[0086] Optionally, the convergence discrimination process can also be trained using reinforcement learning. In the reinforcement learning framework, the state is the current state S(t) of the shared semantic environment, the actions are {continue iteration, stop iteration}, and the reward function is = final solution quality - λ × discussion cost (where λ is the cost trade-off coefficient). Through reinforcement learning training, the convergence discrimination can learn a more accurate convergence judgment strategy, minimizing discussion cost while ensuring output quality.

[0087] In one embodiment, the convergence index information of the multiple different dimensions includes the consensus index, the innovation stagnation index, the constraint satisfaction index, and the discussion efficiency index. The opinion consistency index is determined based on the variance of the quality assessment results of the opinions generated by each agent in the current round; the innovation stagnation index is determined based on the change in the number of new innovative opinions between the current round and the previous round; the constraint satisfaction index is determined based on the proportion of opinions in the current shared semantic environment that satisfy the constraints of the target task; and the discussion efficiency index is determined based on the proportion of effective opinion rounds to the total number of iteration rounds.

[0088] The viewpoint consistency index reflects whether the understanding of the target task among agents in a heterogeneous agent cluster tends to be consistent. Specifically, in the current round, the quality of the viewpoints generated by each agent is evaluated, and the quality evaluation results of each agent's viewpoints are obtained. Then, the variance of these quality evaluation results is calculated. The viewpoint consistency index can be represented by Var(Q(P_i)), where Q(P_i) is the score of viewpoint P_i generated by agent i, and Var is the variance function. When Var(Q(P_i)) < ε1 (ε1 is a preset variance threshold), it indicates that the viewpoint consistency index has reached the convergence requirement.

[0089] The innovation stagnation index is determined based on the change in the number of new innovative ideas between the current round and the previous round. It reflects whether the adversarial-collaborative iterative process is still generating substantial new ideas and perspectives. The innovation stagnation index can be represented by ΔInnovation(t, t-Δ), where t is the current round and Δ is the round interval for comparison. When ΔInnovation(t, t-Δ) < ε2 (ε2 is a preset innovation stagnation threshold), it indicates that the innovation stagnation index has reached convergence.

[0090] The constraint satisfaction index is determined based on the proportion of constraints met by the viewpoints within the current shared semantic environment for the target task. It reflects the extent to which the currently discussed solutions satisfy the various constraints set for the target task. The constraint satisfaction index can be expressed as ConstraintSatisfaction = Number of satisfied constraints / Total number of constraints. When ConstraintSatisfaction > η (where η is a preset constraint satisfaction threshold), it indicates that the constraint satisfaction index has reached convergence.

[0091] The discussion efficiency metric is determined based on the proportion of effective viewpoint rounds out of the total number of iteration rounds. The discussion efficiency metric reflects the effectiveness of the adversarial collaborative iterative process. Specifically, it counts the number of rounds that generated effective viewpoints within the total number of completed iteration rounds and calculates their proportion of the total iteration rounds. The discussion efficiency metric can be expressed as: Effective viewpoint rounds (UsefulTurns) / TotalTurns (TotalTurns). When UsefulTurns / TotalTurns > ρ (where ρ is a preset discussion efficiency threshold), it indicates that the discussion efficiency metric has met the requirements. It should be noted that a continuous decline in discussion efficiency can also serve as an auxiliary signal to trigger convergence.

[0092] In one embodiment, determining the output of the target task based on the recorded information in the shared semantic environment includes: Traverse each node in the graph structure of the shared semantic environment and determine the node that satisfies the preset consensus condition as the target viewpoint node. The edges of the target viewpoint node in the graph structure are determined, and the information of the target viewpoint node is concatenated according to the logical relationship of the edges to obtain the output result of the target task.

[0093] Preset consensus conditions refer to the criteria used to select core viewpoint nodes that reach consensus from a graph structure in a shared semantic environment. Preset consensus conditions can be based on node attribute information. For example, a node's confidence level exceeding a preset confidence threshold can be used as part of the preset consensus conditions. Alternatively, a node being recognized by multiple agents with different roles can be used as part of the preset consensus conditions. Furthermore, a node not being effectively refuted during adversarial collaborative iterations can also be used as part of the preset consensus conditions.

[0094] A target viewpoint node refers to the set of nodes in a graph structure within a shared semantic environment that satisfy predefined consensus conditions. Target viewpoint nodes carry the core viewpoint information that has reached a consensus after multiple rounds of adversarial and collaborative iterations, and serve as the information foundation for the synthesized output.

[0095] This can be implemented using graph traversal algorithms such as breadth-first search or depth-first search. For each node visited, determine whether it meets the preset consensus conditions. If it does, mark the node as the target viewpoint node.

[0096] After obtaining the set of target viewpoint nodes, the edges of these nodes in the graph structure are further obtained. The logical relationships between the edges reflect the semantic connections and logical structure between the target viewpoint nodes. Concatenating the information of the target viewpoint nodes according to the logical relationships between the edges means organizing the viewpoint information carried by each node into a coherent and complete output content in a logical order based on the logical relationships between the target viewpoint nodes.

[0097] Optionally, the following processing can also be performed during the assembly process: consensus extraction, extracting the core viewpoints that reached a consensus from the historical record of adversarial collaborative iterations; innovation point identification, identifying breakthrough innovative viewpoints generated during the discussion; constraint satisfaction verification, verifying the degree to which the final solution satisfies the various constraints of the target task; and structured organization, organizing the scattered viewpoints into a logically coherent and complete solution. The above processing can be implemented through the result synthesizer module, whose synthesis algorithm can be represented as: Result=Organize(Filter(Rank(C∪I, Relevance), Constraints)), where C is the set of consensus viewpoints, I is the set of innovation points, Result is the final result of the synthesis algorithm, Constraints are the constraints, Relevance is the relevance, Rank is the ranking operation based on relevance, Filter is the filtering operation based on constraints, and Organize is the structured organization operation.

[0098] Optionally, the implementation process of the present invention can be implemented using a layered modular architecture, including an environment management layer, an intelligent agent cluster layer, an innovation engine layer, and a convergence control layer.

[0099] The environment management layer is responsible for the management and maintenance of the shared semantic environment, including the shared semantic environment, the environment state tracker (which records and visualizes the evolution of thought in real time), and the consensus anchor manager (which identifies and solidifies the core viewpoints for reaching a consensus).

[0100] The agent cluster layer is responsible for managing heterogeneous agent clusters, including the role factory (which dynamically generates specialized agent roles based on task type), the capability configurator (which configures differentiated cognitive and decision-making capabilities for each agent), and the interaction protocol engine (which manages communication rules and timing control between agents).

[0101] The innovation engine layer is responsible for driving innovation stimulation in the adversarial and collaborative iteration process, including the adversarial stimulator (designing structured conflict scenarios to stimulate innovative thinking), the constraint integrator (transforming multi-dimensional constraints into decision-making influencing factors), and the solution evolver (supporting incremental improvement and breakthrough innovation of solutions).

[0102] The convergence control layer is responsible for the convergence control of the collaborative iterative process, including a quality evaluator (evaluating the quality and innovation of the discussion from multiple dimensions), a convergence discriminator (predicting the optimal stopping point), and an outcome synthesizer (integrating scattered viewpoints into a structured final solution).

[0103] The adversarial cooperation and iterative generation device of the present invention is described below. The adversarial cooperation and iterative generation device of the present invention described below can be referred to in correspondence with the adversarial cooperation and iterative generation method of the present invention described above.

[0104] like Figure 2 A schematic diagram of the multi-agent adversarial cooperation and iterative generation device provided by the present invention is shown. The device includes: The agent cluster construction module 210 is used to determine a heterogeneous agent cluster to execute the target task based on the task requirement information of the target task. The heterogeneous agent cluster includes multiple agents with different role types. Each agent is configured with cognitive bias parameters and decision parameters corresponding to its role type. The cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model. The decision parameters are the maximum inference step parameters, search width parameters, and confidence threshold parameters in the agent's large language model. The iterative execution module 220 is used to trigger the heterogeneous intelligent agent cluster to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment, and to determine the output result of the target task based on the recorded information in the shared semantic environment when the adversarial cooperative iteration reaches a preset convergence condition. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0105] In one embodiment, the iterative execution module 220 is specifically used for: The shared semantic environment also includes a subset of information about each agent in the heterogeneous agent cluster. This subset records the agent's role description information. Each round of adversarial-cooperative iteration includes: Based on the role type of each agent in the heterogeneous agent cluster, the target agent to be discussed first is determined from the heterogeneous agent cluster; The information subset is loaded into the target agent, and the initial viewpoint of the target agent with the loaded information subset on the target task is obtained. If the quality assessment result of the initial viewpoint meets the preset quality conditions, the initial viewpoint is written into the shared semantic environment to obtain the updated shared semantic environment. Obtain response viewpoints and write them into the updated shared semantic environment. The response viewpoints are viewpoints generated by other agents using adversarial strategies against the initial viewpoints. The other agents are agents in the heterogeneous agent cluster other than the target agent.

[0106] In one embodiment, the iterative execution module 220 is further configured to: The adversarial strategy is determined to include at least one of the following: alternative generation strategy, hypothesis challenge strategy, constraint relaxation strategy, and perspective shift strategy; The alternative generation strategy is used to generate at least one alternative for the initial viewpoint; The hypothesis challenge strategy is used to identify implicit hypothesis information in the initial viewpoint and to raise questions about the implicit hypothesis information; The constraint relaxation strategy is used to reduce the constraints of the target task and generate new candidate viewpoints based on the reduced constraints. The perspective shifting strategy is used to re-evaluate the initial viewpoint and generate evaluation opinions based on agents in the heterogeneous agent cluster whose role categories are different from those of the target agent.

[0107] In one embodiment, the iterative execution module 220 is further configured to: Determine if the adversarial cooperative iteration reaches the preset convergence condition, including: After each iteration, the current state data in the shared semantic environment is extracted; Based on the current state data, convergence index information in multiple different dimensions is determined; The convergence index information of the multiple different dimensions is weighted and fused to obtain the convergence score of the current state data; If the convergence score is greater than a preset score threshold, it is determined that the adversarial cooperative iteration has reached the preset convergence condition.

[0108] In one embodiment, the iterative execution module 220 is further configured to: The convergence index information of the multiple different dimensions includes the consensus index, innovation stagnation index, constraint satisfaction index, and discussion efficiency index; The opinion consistency index is determined based on the variance of the quality assessment results of the opinions generated by each agent in the current round; the innovation stagnation index is determined based on the change in the number of new innovative opinions between the current round and the previous round; the constraint satisfaction index is determined based on the proportion of opinions in the current shared semantic environment that satisfy the constraints of the target task; and the discussion efficiency index is determined based on the proportion of effective opinion rounds to the total number of iteration rounds.

[0109] In one embodiment, the iterative execution module 220 is further configured to: Determining the output result of the target task based on the recorded information in the shared semantic environment includes: Traverse each node in the graph structure of the shared semantic environment and determine the node that satisfies the preset consensus condition as the target viewpoint node. The edges of the target viewpoint node in the graph structure are determined, and the information of the target viewpoint node is concatenated according to the logical relationship of the edges to obtain the output result of the target task.

[0110] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a multi-agent adversarial cooperation and iterative generation method. This method includes: determining a heterogeneous agent cluster to execute the target task based on the task requirement information of the target task; the heterogeneous agent cluster includes multiple agents with different role types, each agent being configured with cognitive bias parameters and decision parameters corresponding to its role type; the cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model; the decision parameters are the maximum inference step parameters, search width parameters, and confidence threshold parameters in the agent's large language model. The heterogeneous intelligent agent cluster is triggered to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment. When the adversarial cooperative iteration reaches a preset convergence condition, the output result of the target task is determined based on the recorded information in the shared semantic environment. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0111] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-agent adversarial cooperation and iterative generation method provided by the above methods. The method includes: determining a heterogeneous agent cluster to execute the target task based on the task requirement information of the target task; the heterogeneous agent cluster includes multiple agents with different role types, and each agent is configured with cognitive bias parameters and decision parameters corresponding to its role type; the cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model; the decision parameters are the maximum inference step parameters, search width parameters, and confidence threshold parameters in the agent's large language model. The heterogeneous intelligent agent cluster is triggered to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment. When the adversarial cooperative iteration reaches a preset convergence condition, the output result of the target task is determined based on the recorded information in the shared semantic environment. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0113] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the multi-agent adversarial cooperation and iterative generation method provided by the above methods. The method includes: determining a heterogeneous agent cluster to perform the target task based on the task requirement information of the target task; the heterogeneous agent cluster includes multiple agents of different role types, each agent being configured with cognitive bias parameters and decision parameters corresponding to its role type; the cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model; the decision parameters are the maximum inference step parameters, search width parameters, and confidence threshold parameters in the agent's large language model. The heterogeneous intelligent agent cluster is triggered to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment. When the adversarial cooperative iteration reaches a preset convergence condition, the output result of the target task is determined based on the recorded information in the shared semantic environment. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

[0114] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adversarial cooperation and iterative generation of multi-agent systems, characterized in that, include: Based on the task requirements information of the target task, determine the heterogeneous intelligent agent cluster to execute the target task; The heterogeneous agent cluster includes multiple agents with different role types. Each agent is configured with cognitive bias parameters and decision parameters corresponding to its role type. The cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model. The decision parameters are the maximum inference steps, search width, and confidence threshold parameters in the agent's large language model. The heterogeneous intelligent agent cluster is triggered to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment. When the adversarial cooperative iteration reaches a preset convergence condition, the output result of the target task is determined based on the recorded information in the shared semantic environment. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

2. The multi-agent adversarial cooperation and iterative generation method according to claim 1, characterized in that, The shared semantic environment also includes a subset of information about each agent in the heterogeneous agent cluster. This subset records the agent's role description information. Each round of adversarial-cooperative iteration includes: Based on the role type of each agent in the heterogeneous agent cluster, the target agent to be discussed first is determined from the heterogeneous agent cluster; The information subset is loaded into the target agent, and the initial viewpoint of the target agent with the loaded information subset on the target task is obtained. If the quality assessment result of the initial viewpoint meets the preset quality conditions, the initial viewpoint is written into the shared semantic environment to obtain the updated shared semantic environment. Obtain response viewpoints and write them into the updated shared semantic environment. The response viewpoints are viewpoints generated by other agents using adversarial strategies against the initial viewpoints. The other agents are agents in the heterogeneous agent cluster other than the target agent.

3. The multi-agent adversarial cooperation and iterative generation method according to claim 2, characterized in that, The adversarial strategy includes at least one of the following: alternative generation strategy, hypothesis challenge strategy, constraint relaxation strategy, and perspective shifting strategy; The alternative generation strategy is used to generate at least one alternative for the initial viewpoint; The hypothesis challenge strategy is used to identify implicit hypothesis information in the initial viewpoint and to raise questions about the implicit hypothesis information; The constraint relaxation strategy is used to reduce the constraints of the target task and generate new candidate viewpoints based on the reduced constraints. The perspective shifting strategy is used to re-evaluate the initial viewpoint and generate evaluation opinions based on agents in the heterogeneous agent cluster whose role categories are different from those of the target agent.

4. The multi-agent adversarial cooperation and iterative generation method according to claim 1, characterized in that, Determine if the adversarial cooperative iteration reaches the preset convergence condition, including: After each iteration, the current state data in the shared semantic environment is extracted; Based on the current state data, convergence index information in multiple different dimensions is determined; The convergence index information of the multiple different dimensions is weighted and fused to obtain the convergence score of the current state data; If the convergence score is greater than a preset score threshold, it is determined that the adversarial cooperative iteration has reached the preset convergence condition.

5. The multi-agent adversarial cooperation and iterative generation method according to claim 4, characterized in that, The convergence metrics information from multiple dimensions includes the consensus metric, innovation stagnation metric, constraint satisfaction metric, and discussion efficiency metric. The opinion consistency index is determined based on the variance of the quality assessment results of the opinions generated by each agent in the current round; the innovation stagnation index is determined based on the change in the number of new innovative opinions between the current round and the previous round; the constraint satisfaction index is determined based on the proportion of opinions in the current shared semantic environment that satisfy the constraints of the target task; and the discussion efficiency index is determined based on the proportion of effective opinion rounds to the total number of iteration rounds.

6. The multi-agent adversarial cooperation and iterative generation method according to claim 1, characterized in that, Determining the output result of the target task based on the recorded information in the shared semantic environment includes: Traverse each node in the graph structure of the shared semantic environment and determine the node that satisfies the preset consensus condition as the target viewpoint node. The edges of the target viewpoint node in the graph structure are determined, and the information of the target viewpoint node is concatenated according to the logical relationship of the edges to obtain the output result of the target task.

7. A multi-agent adversarial cooperation and iterative generation device, characterized in that, include: The intelligent agent cluster construction module is used to determine the heterogeneous intelligent agent cluster that will execute the target task based on the task requirement information of the target task. The heterogeneous agent cluster includes multiple agents with different role types. Each agent is configured with cognitive bias parameters and decision parameters corresponding to its role type. The cognitive bias parameters are temperature parameters, kernel sampling parameters, candidate lexical number parameters, frequency penalty parameters, existence penalty parameters, and lexical-level log probability bias parameters in the agent's large language model. The decision parameters are the maximum inference steps, search width, and confidence threshold parameters in the agent's large language model. The iterative execution module is used to trigger the heterogeneous intelligent agent cluster to perform multiple rounds of adversarial cooperative iteration on the target task in a shared semantic environment, and to determine the output result of the target task based on the recorded information in the shared semantic environment when the adversarial cooperative iteration reaches a preset convergence condition. The shared semantic environment is used to record the interaction information of each agent in the heterogeneous agent cluster during the adversarial collaboration iteration process. The shared semantic environment is an information space constructed based on a graph structure, in which nodes are used to represent conceptual information, opinion information or factual information, and edges are used to represent the logical relationships between nodes.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent adversarial cooperation and iterative generation method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-agent adversarial cooperation and iterative generation method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-agent adversarial cooperation and iterative generation method as described in any one of claims 1 to 6.

Images