Heterogeneous intelligent agent swarm coordination method, system, device and program product

Abstract

The application provides a heterogeneous intelligent agent swarm coordination method, system, device and program product. The method comprises the following steps: a task access layer receives a current to-be-processed task and extracts a task feature to obtain a task demand vector, and then broadcasts the task demand vector to a heterogeneous intelligent agent swarm, so that an intelligent agent receiving the task demand vector calculates a current task matching score based on a current multi-dimensional capability fingerprint vector of the intelligent agent, and decides whether to send a bidding message to the task access layer according to the task matching score; the task access layer collects the bidding messages and arbitrates according to the bidding messages to determine a winning intelligent agent and execute the current to-be-processed task by the winning intelligent agent. The application only generates communication at moments such as broadcasting the task demand vector and sending the bidding message, has adaptive coordination frequency, and reduces communication overhead. In addition, the application also considers the dynamic capability of the intelligent agent, realizes task allocation based on actual capability, and improves task matching accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, system, device and program product for coordinating heterogeneous intelligent agent swarms. Background Technology

[0002] With the development of Large Language Model (LLM) technology, agent systems based on LLM are evolving from single-agent systems to multi-agent collaborative systems. Multi-agent systems, through the collaboration of multiple specialized agents, are able to handle complex tasks that far exceed the capabilities of a single agent.

[0003] The following schemes are currently mainly used for multi-agent coordination: The first approach is a centralized scheduling scheme, which uses a single central node (Supervisor) to handle task decomposition and allocation. However, if the central node fails, all agents will be paralyzed. Furthermore, as the number of agents increases, the central node will experience processing bottlenecks, and the complexity of investigations will increase exponentially with the number of agents.

[0004] The second approach is a static pipeline: agents execute tasks in a predefined linear order. However, if a problem occurs in any link, all agents will stop working. Furthermore, this approach lacks flexibility as it cannot dynamically adjust its route when encountering new types of tasks.

[0005] The third approach is a coordination scheme based on MARL: using multi-agent reinforcement training to improve the cooperative strategy between agents. However, the training cost is extremely high, requiring a large amount of computing resources. The action space of LLM agents is large, making training difficult to converge, and the well-trained strategy is also difficult to adapt to new scenarios.

[0006] The fourth type is the bidding mode, such as the CBAA and CBBA algorithms, which allocate tasks through a bidding mechanism. However, the capabilities of the agents change dynamically with the context and tool availability, and all agents need to participate in each bidding process, resulting in a large amount of communication.

[0007] The fifth method is predefined routing, which enables task transfer between agents through predefined rules. However, this method cannot handle collaborative patterns that are not predefined and lacks self-organization and emergent capabilities.

[0008] The aforementioned existing solutions all have the following shortcomings when dealing with heterogeneous intelligent agent bee swarms driven by large language models: (1) High communication overhead: Agents need to exchange information frequently to maintain coordination. In large-scale bee swarms (tens to hundreds of agents), communication bandwidth becomes the system bottleneck.

[0009] (2) Failure to consider capability heterogeneity: Existing technologies mostly assume that the capabilities of agents are simultaneously or statically assessable, while the actual capabilities of LLM agents change dynamically with the context window occupancy rate, tool configuration changes, and different professional directions.

[0010] (3) Lack of adaptive coordination frequency: Regardless of the urgency of the task or the rate of change of the environment, coordination communication is carried out at a fixed frequency, which wastes bandwidth when idle and may not respond in time when busy.

[0011] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0012] To address the problems in the prior art, the purpose of this application is to provide a heterogeneous intelligent agent swarm coordination method, system, device, and program product, aiming to solve the problems of high communication overhead, failure to consider the heterogeneity of intelligent agent capabilities, and lack of adaptive coordination capabilities in the prior art.

[0013] The first aspect of this application provides a heterogeneous intelligent agent swarm coordination method, comprising the following steps: a task access layer receives a current task to be processed and extracts task features to obtain a task requirement vector; the task access layer broadcasts the task requirement vector to a heterogeneous intelligent agent swarm, so that the agents receiving the task requirement vector can calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer accordingly; the task access layer collects the bidding messages and arbitrates accordingly to determine the winning agent, which then executes the current task to be processed.

[0014] In some embodiments of this application, the formula for calculating the task matching score is as follows: ;in, The task matching score for agent i is the current task. For cosine similarity, Let i be the current multidimensional capability fingerprint vector of agent i. For the task requirement vector, This represents the current load status of agent i.

[0015] In some embodiments of this application, each agent in the heterogeneous agent swarm updates its own multidimensional capability fingerprint vector according to an update rule, which includes: Rule 1: Whenever a task is completed, the quantified value of the corresponding task processing capability in the current multidimensional capability fingerprint vector is updated using an exponential moving average strategy based on the task execution result. ;in, This represents the quantized value of task processing capability j in the current multidimensional capability fingerprint vector. For the result of task execution, For learning rate, 1. The quantized value after dimension j is updated; Rule 2: When the tools that the agent can call change, recalculate the quantized value of the corresponding task processing capability and update the current multidimensional capability fingerprint vector; Rule 3: The quantized value of the task processing capability corresponding to the task that the agent has not executed within the preset time threshold decays exponentially.

[0016] In some embodiments of this application, the agent receiving the task requirement vector decides whether to send a bidding message to the task access layer based on its calculated current task matching score. Specifically, if the agent receiving the task requirement vector calculates its current task matching score, which exceeds the current bidding threshold, then it sends a bidding message to the task access layer; otherwise, it does not send a bidding message to the task access layer.

[0017] In some embodiments of this application, the task access layer collects bidding messages and arbitrates accordingly, specifically including: if the task access layer does not collect bidding messages, the task access layer lowers the bidding threshold and broadcasts the lowered bidding threshold to the heterogeneous agent swarm for re-bidding; if the task access layer collects multiple bidding messages, the task access layer calculates the difference between every two task matching scores based on the task matching scores contained in each of the multiple bidding messages; if all calculated differences are less than the difference threshold, a second round of bidding is initiated; otherwise, the agent with the highest task matching score is determined as the winning agent.

[0018] In some embodiments of this application, the method further includes: periodically broadcasting a global target gradient vector to a heterogeneous agent swarm, so that agents receiving the global target gradient vector can fine-tune their parameters based on the global target gradient vector.

[0019] In some embodiments of this application, the heterogeneous agent swarm communicates according to its internal communication topology. The method further includes: updating the communication topology of the heterogeneous agent swarm according to a topology update rule, wherein the topology update rule includes: increasing the communication link weight between the two agents when two agents in the heterogeneous agent swarm successfully cooperate to complete a task; decreasing the communication link weight between the two agents when the number of competition conflicts between the two agents exceeds a preset threshold; and cutting off the communication link between the two agents when the time without communication between the two agents exceeds a preset communication time threshold.

[0020] A second aspect of this application provides a heterogeneous agent swarm coordination system for implementing the heterogeneous agent swarm coordination method described above. The system includes a task access layer and a heterogeneous agent swarm. The task access layer receives a currently pending task and extracts task features to obtain a task requirement vector. The task access layer broadcasts the task requirement vector to the heterogeneous agent swarm, allowing agents receiving the task requirement vector to calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer. The task access layer collects the bidding messages and arbitrates them to determine the winning agent, which then executes the currently pending task.

[0021] A third aspect of this application also provides a heterogeneous intelligent agent bee colony coordination device, comprising: processor; A memory in which executable instructions of the processor are stored; The processor is configured to execute the steps of the heterogeneous agent bee colony coordination method described above by executing the executable instructions.

[0022] A fourth aspect of this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the heterogeneous intelligent agent bee colony coordination method described above.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application.

[0024] The heterogeneous intelligent agent bee colony coordination method, system, device, and program products of this application have the following beneficial effects: This application only generates communication when broadcasting task demand vectors and sending bidding messages, featuring adaptive coordination frequency while reducing communication overhead. Furthermore, this application considers the dynamic capabilities of the agent, implementing task allocation based on actual capabilities and improving task matching accuracy. Attached Figure Description

[0025] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart of a heterogeneous intelligent agent bee colony coordination method according to an embodiment of this application; Figure 2 This is a schematic diagram illustrating the connection relationship between the task access layer and the heterogeneous intelligent agent swarm according to an embodiment of this application; Figure 3 This is a detailed architecture diagram of the task access layer and heterogeneous intelligent agent swarm according to an embodiment of this application; Figure 4 This is a schematic diagram of three events according to an embodiment of this application; Figure 5 This is a schematic diagram of an emergence guidance layer according to an embodiment of this application; Figure 6 This is a schematic diagram of the communication topology evolution process according to a specific embodiment of this application; Figure 7 This is a schematic diagram of the structure of a heterogeneous intelligent agent bee colony coordination system according to an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a heterogeneous intelligent agent bee colony coordination device according to an embodiment of this application; Figure 9 This is a schematic diagram of a program product according to an embodiment of this application. Detailed Implementation

[0027] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0028] Furthermore, the accompanying drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. Although the terms "first" or "second," etc., are used in this specification to denote certain features, these are only for indicating function and not as a limitation on the number or importance of specific features.

[0029] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined. Therefore, the actual execution order may change depending on the specific circumstances.

[0030] like Figure 1 As shown in the figure, this application provides a method for coordinating heterogeneous intelligent agent bee colonies, including the following steps: Step S101: The task access layer receives the current task to be processed and extracts task features to obtain the task requirement vector.

[0031] In one embodiment, the task requirement vector includes the intensity of requirement for one or more task processing capabilities.

[0032] Step S102: The task access layer broadcasts the task requirement vector to the heterogeneous agent swarm, so that the agents that receive the task requirement vector can calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer.

[0033] In one embodiment, each agent in the heterogeneous agent swarm maintains a dynamic multidimensional capability fingerprint vector, and each agent updates the multidimensional capability fingerprint vector according to an update rule. The multidimensional capability fingerprint vector includes quantified values ​​of various task processing capabilities.

[0034] In one embodiment, the agent that receives the task requirement vector calculates its current task matching score based on its current multidimensional capability fingerprint vector by: the agent that receives the task requirement vector calculates the cosine similarity between its current multidimensional capability fingerprint vector and the task requirement vector, and then multiplies the cosine similarity with the agent's current load status to obtain the agent's current task matching score.

[0035] In one embodiment, the agent that receives the task requirement vector decides whether to send a bidding message to the task access layer based on its calculated current task matching score. Specifically, if the agent's current task matching score exceeds the current bidding threshold, it sends a bidding message to the task access layer; otherwise, it does not send a bidding message to the task access layer.

[0036] Step S103: The task access layer collects bidding messages and arbitrates them accordingly to determine the winning intelligent agent, which then executes the current pending task.

[0037] In one embodiment, the task access layer collects bidding messages and arbitrates accordingly, specifically including: If the task access layer does not collect bidding messages, the task access layer lowers the bidding threshold and broadcasts the lowered bidding threshold to the heterogeneous intelligent agent swarm for re-bidding; If the task access layer collects multiple bidding messages, it calculates the difference between the matching scores of each pair of tasks based on the task matching scores in the multiple bidding messages. If all the calculated differences are less than the difference threshold, a second round of bidding is started; otherwise, the agent with the highest task matching score is determined as the winning agent.

[0038] This application only generates communication when broadcasting task demand vectors and sending bidding messages; no communication occurs at other times. This balances efficiency and real-time performance, features adaptive coordination frequency, and reduces communication overhead. Furthermore, this application considers the dynamic capabilities of the agent, enabling task allocation based on actual capabilities and improving task matching accuracy.

[0039] The application scenarios of this application will be introduced below: The heterogeneous agent swarm coordination method in this application is applied to heterogeneous agent swarms, which are collaborative systems composed of multiple agents of different types and capabilities. Heterogeneous means that the agents have different capabilities and expertise, unlike the assumption of node synchronization in traditional distributed systems. An agent is an AI program unit capable of autonomously executing tasks. Each agent is typically driven by a large language model and configured with specific tools and specializations. For example, an agent for code generation might be equipped with a code interpreter and debugging tools, while an agent for data analysis might be equipped with database query and visualization tools.

[0040] Then, the specific implementation methods of each step will be further introduced.

[0041] Step S101: The task access layer receives the current task to be processed and extracts task features to obtain the task requirement vector.

[0042] In one embodiment, the following two methods are mainly used to send tasks to the task access layer.

[0043] The first method: External distribution. Specifically, tasks are distributed from external third-party systems (such as upstream business platforms, clients, etc.) to the task access layer. For example, you can input "Analyze sales data and write a report" in the client to distribute the task to the task access layer.

[0044] The second approach is that if an agent in a heterogeneous agent swarm discovers the need for subtask assistance during task execution (such as the need to generate visualization code while generating an analysis report), it can form a corresponding task and send it to the task access layer, forming a recursive collaboration. The task will be naturally decomposed into a chain of subtasks.

[0045] In one embodiment, the task access layer may be a task access gateway. The task access layer may be independent of the heterogeneous agent swarm, or it may be located within one agent of the heterogeneous agent swarm.

[0046] In one embodiment, the task requirement vector includes the demand intensity of one or more task processing capabilities. Specifically, the task requirement vector has multiple dimensions, each corresponding to a task processing capability, and the value of that dimension is the demand intensity of that task processing capability. It should be noted that the specific task processing capability corresponding to each dimension can be set according to requirements, and this application does not limit this. The task requirement vector can be represented as follows: k is a positive integer. This indicates the intensity of the demand for task processing capacity k. Task processing capabilities can include coding, data analysis, and collaboration.

[0047] In one embodiment, the task description of the current task to be processed is input into a pre-trained task feature extraction model to extract task features and obtain a task requirement vector. Specifically, the pre-trained task feature extraction model can use a pre-trained lightweight text classifier. The task description of the current task to be processed is input into the pre-trained lightweight text classifier, which automatically outputs the corresponding task requirement vector. For example, if the current task to be processed is "analyze sales data and write a report", the task requirement vector output by the pre-trained lightweight text classifier is... The training process of a lightweight text classifier includes: obtaining task descriptions for various tasks and manually setting task requirement vectors for each task to construct a training set; inputting the training set into a lightweight text classifier (such as the BERT model) for training to obtain a pre-trained lightweight text classifier.

[0048] Step S102: The task access layer broadcasts the task requirement vector to the heterogeneous agent swarm, so that the agents that receive the task requirement vector can calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer.

[0049] In a heterogeneous agent swarm, all or some agents communicate via communication links. The communication topology can be used to characterize the direct communication objects of each node in the heterogeneous agent swarm, as well as the established communication links between nodes. The task requirement vector is transmitted to all reachable agents within the heterogeneous agent swarm according to the communication topology. Specifically, such as... Figure 2 As shown, one or more agents can be selected as message receiving nodes within a heterogeneous agent swarm. After obtaining the task request vector sent by the task access layer, the message receiving node forwards it to the agents with which it has established a communication link. The agents that receive the task request vector then forward it to the agents with which they have established a communication link, and so on, layer by layer. It is worth noting that if an agent has not established a communication link with any other agent in the heterogeneous agent swarm, then that agent will not receive the task request vector.

[0050] It should be noted that, Figure 2 The present invention only illustrates one example of a transmission path from the task access layer to the agent and does not imply any limitation thereof.

[0051] like Figure 3 As shown, the communication topology is not a fixed, fully connected, or predefined structure, but rather dynamically updated based on the history of agent collaboration. Specifically, when registering a heterogeneous agent swarm, an initial communication topology is pre-constructed. This initial topology can be customized or randomly constructed, and then updated according to the topology update rules. It should be understood that swarm registration refers to the process by which a single agent completes its identity level and information reporting to the swarm management node, formally joining the swarm.

[0052] Topology update rules include: (1) When two agents successfully complete a collaborative task, the current link weight of the communication link between the two agents is obtained, and a first preset weight value is added to the current link weight. The added link weight is then updated to the current link weight. It should be noted that the first preset weight value can be set according to requirements, and this application does not limit it. For example, the current link weight is... w a The initial weight value is set to 0.1, and the updated current link weight is... w a +1.

[0053] (2) When the number of times two agents compete for the same task exceeds a preset threshold, the current link weight of the communication link between the two agents is obtained, and a second preset weight value is reduced from the current link weight. The reduced link weight is then updated to the current link weight. It should be noted that both the threshold and the second preset weight value can be set according to requirements, and this application does not limit them. For example, the current link weight is... w b The second weight is set to 0.05, and the updated current link weight is... w b -0.05.

[0054] (3) When the time of no communication between two agents exceeds a preset communication time threshold, the communication link between the two agents is cut off. If necessary, reachability can be restored through multi-hop routing. It should be understood that when the original communication link in the communication topology is broken (lost reachability), a valid communication link path is re-established through mechanisms such as local repair and backup link switching, and connectivity is restored. This whole process is called multi-hop routing to restore reachability.

[0055] (4) Periodically and randomly establish a small number of remote communication links to ensure dense local connections and very few global cross-region forwarding hops (i.e., the small-world characteristics of the topology). The shortest path between any two agents does not exceed O(log n) hops, thus avoiding the bee colony splitting into isolated subgroups that are unreachable to each other.

[0056] It should be understood that communication link weight is a quantitative cost or priority indicator assigned to each communication link, used to support decisions such as routing path selection, traffic allocation, and information scheduling.

[0057] Through the aforementioned communication topology update based on collaboration history, the communication topology of the heterogeneous agent bee colony gradually adapts to the actual agent collaboration mode, reduces invalid communication paths, and enables the heterogeneous agent bee colony to adaptively evolve towards efficient collaboration.

[0058] like Figure 3 As shown, each agent maintains a dynamic multidimensional capability fingerprint vector, which includes quantized values ​​of various task processing capabilities. Specifically, each dimension in the multidimensional capability fingerprint vector corresponds to a type of task processing capability, and the value of that dimension is the quantized value of that task processing capability. The specific task processing capability corresponding to each dimension can be set according to requirements, and this application does not limit this. The number of dimensions of the multidimensional capability fingerprint vector (i.e., the number of task processing capabilities) is determined based on the total number of task types registered in the bee colony. Task processing capabilities can be code, data analysis, writing, etc. When each agent goes online, an initial multidimensional capability fingerprint vector is generated based on its configuration information. The configuration information includes, but is not limited to, model type, list of callable tools, system prompt keywords, etc. One or more quantized features can be set for each task processing capability (for example, the quantized features of analysis processing capability can be analysis tools, model type, etc.), and different quantized features have different quantized values. All quantized features of each task processing capability contained in the configuration information of each agent are counted, and the quantized value of each task processing capability is calculated accordingly. Then, the quantized values ​​of each task processing capability are summarized to obtain the initial multidimensional capability fingerprint vector. Alternatively, configuration information can be input into a pre-trained model for each task processing capability to obtain a quantized value for each capability, which can then be aggregated to obtain an initial multi-dimensional capability fingerprint vector. The model training process for each task processing capability includes: acquiring configuration information for multiple agents and manually setting the quantized value for that task processing capability to construct a training set for that capability; inputting the training set into a language model for training to obtain a pre-trained model for that capability. It should be noted that these are merely two illustrative examples and do not represent a limitation of the present invention.

[0059] Subsequently, each intelligent agent updates its own multidimensional capability fingerprint vector according to the update rules, which include: Rule 1: Whenever a task is completed, the quantified value of the corresponding task processing capability in the current multidimensional capability fingerprint vector is updated using an exponential moving average based on the task execution result. (1) in, This represents the quantized value of task processing capability j in the current multidimensional capability fingerprint vector. For the result of task execution, The learning rate can be set according to requirements. This is the quantified value of the updated task processing capability j.

[0060] Rule 2: When the tools available to the agent change, recalculate the quantized value of the corresponding task processing capability and update the current multidimensional capability fingerprint vector.

[0061] Rule 3: The quantified value of the task processing capability corresponding to a task that the agent has not executed within a preset time threshold decays exponentially.

[0062] The specific execution process of Rule 1 includes: for each agent, upon completion of a task, obtaining the corresponding task execution result. The types of task execution results include success, failure, and partial completion. Each task execution result corresponds to a... The value can be set according to actual needs. The quantized value of the task processing capability corresponding to the task is obtained from the current multi-dimensional capability fingerprint vector and calculated based on a pre-set learning rate. Calculate the quantized value of the new task processing capability according to the above formula (1) and then update it to the current multidimensional capability fingerprint vector.

[0063] The specific execution process of Rule 2 includes: for each agent, when the callable tools change (for example, a tool suddenly disconnects or a new tool is added), the quantization value of the corresponding task processing capability is recalculated using the calculation method described in the above embodiments, and the calculation result is updated to the current multidimensional capability fingerprint vector.

[0064] The specific execution process of Rule 3 includes: when an agent has not performed a certain task within a preset time threshold, it obtains the quantized value of the corresponding task processing capability from the current multi-dimensional capability fingerprint vector. Starting with this quantized value, each time step multiplies the current quantized value by a decay coefficient. For example, the decay coefficient γ≈0.995, the initial value is P, and the value after three time steps is P×γ×γ×γ. It should be understood that a time step refers to the smallest periodic unit in which the agent completes one interaction with the environment.

[0065] By maintaining multidimensional capability fingerprint vectors in real time, the dynamic capability changes of intelligent agents can be reflected more accurately, enabling precise task allocation based on actual capabilities (rather than static labels), and significantly improving task matching accuracy.

[0066] In one embodiment, when the magnitude change of the multidimensional capability fingerprint vector of an agent exceeds a set change threshold, the multidimensional capability fingerprint vector at this time is broadcast to its K-hop agent, but not broadcast to the entire network.

[0067] It is important to note that the update of the multidimensional capability fingerprint vector is performed locally by the agent and does not incur communication overhead. Broadcasting to neighboring agents is only triggered when the magnitude change of the multidimensional capability fingerprint vector exceeds a set threshold, rather than periodically. This means that a stable agent may not communicate for several minutes.

[0068] The following describes the specific implementation of step S102: The multidimensional capability fingerprint vector of agents in a heterogeneous agent bee colony can be represented as: .

[0069] Each AI receiving a task requirement vector calculates its current task matching score based on its own multidimensional capability fingerprint vector: (2) in, The task matching score for agent i is the current task. Cosine similarity, with values ​​ranging from 0 to 1, where 0 indicates complete irrelevance and 1 indicates a perfect match. Let i be the current multidimensional capability fingerprint vector of agent i. For the task requirement vector, This represents the current load status of agent i. The current load status of agent i is calculated as follows: (3) in, Let be the current task queue length of agent i, and represent the number of tasks currently pending processing. is the maximum concurrency of agent i, representing the number of tasks that can be processed simultaneously at one time. The value is 0 when fully loaded and 1 when idle.

[0070] Next, after calculating its current task matching score, it compares it with its current bidding threshold. When its current task matching score is greater than the bidding threshold, it decides to bid and sends a bidding message to the task access layer. The bidding message includes, but is not limited to, the agent name (ID), task matching degree, and estimated completion time. When its current task matching score is less than or equal to the bidding threshold, it does not decide to bid and remains silent without generating any communication.

[0071] Step S103: The task access layer collects bidding messages and arbitrates them accordingly to determine the winning intelligent agent, which then executes the current pending task.

[0072] Specifically, when the task access layer does not receive any bidding messages (starvation conflict), the bidding threshold is lowered, and the lowered bidding threshold and task demand vector are rebroadcast to the heterogeneous agent swarm for re-bidding. The bidding process is the same as that described in the above embodiments, and will not be repeated here.

[0073] When the task layer receives a bidding message, it will determine the winning agent from the agent that sent the bidding message, and the winning agent will execute the current pending task.

[0074] When the task layer receives multiple bidding messages, it calculates the difference between each pair of task matching scores based on the task matching scores contained in each bidding message. If all calculated differences are less than a preset difference threshold, a second round of bidding is initiated. The detailed execution plans of each participating agent are obtained, and the winning agent is determined accordingly. The winning agent then executes the current task to be processed. Otherwise, the agent with the highest task matching score is determined as the winning agent, and the winning agent executes the current task to be processed.

[0075] Combining the above embodiments and Figure 3 It can be seen that this invention defines three types of events that trigger the consensus process: like Figure 4 As shown, event type A (new task event) is triggered when the task access layer receives a task. The task access layer extracts the corresponding task requirement vector, which is then broadcast to all reachable agents in the heterogeneous agent swarm. Agents receiving the task requirement vector calculate the task matching score and send a bidding message if the score exceeds the bidding threshold.

[0076] Event Type B (Capability Mutation Event): Triggered when an agent's multidimensional capability fingerprint vector exceeds a change threshold (e.g., a database tool suddenly disconnects, causing the data analysis capability to drop sharply from 0.9 to 0.2). After triggering, the agent only broadcasts the updated multidimensional capability fingerprint vector to its k-hop agents, and does not broadcast it across the entire network.

[0077] Event type C (conflict detection event): Triggered when two or more agents send bidding messages or no agent sends bidding messages. When two or more agents send bidding messages, the decision on whether to conduct a second round of bidding is based on the task matching score. When no agent sends bidding messages, the bidding threshold is lowered and bidding is restarted.

[0078] The event-driven sparse consensus mechanism employed in this invention only initiates the consensus process at a few necessary moments (when an event is triggered), with no coordination or communication between agents at other times. Compared to the periodic heartbeat / state synchronization method in existing technologies, the communication overhead is reduced from O(n²·T) to O(n·E), where n is the number of agents, T is the number of time steps, and E is the number of events (usually E << T). In typical scenarios, the communication overhead can be significantly reduced, supporting real-time coordination of large-scale swarms (hundreds of agents). There is zero communication overhead during non-event periods, and immediate response is provided during emergency events, balancing efficiency and real-time performance.

[0079] Quantitative comparison example: 10 agents run 1000 time steps, assuming 50 events occur. Traditional solution communication volume ≈ 10 × 10 × 1000 = 100,000 times; this solution communication volume ≈ 10 × 50 = 500 times, a significant reduction.

[0080] To prevent the emergence of uncontrollable problems, such as Figure 5 As shown, this invention introduces an emergent guidance layer. The system-level objective is encoded as a global objective gradient vector in the emergent guidance layer. The global target gradient vector is broadcast to the heterogeneous agent swarm at a low frequency (far lower than the event trigger frequency, such as once per minute). Each agent receiving the global target gradient vector superimposes it with its own state vector, and the resulting vector is used to fine-tune the agent's parameters. For example, when the global target is "prioritize urgent tasks," the agent lowers the bidding threshold for urgent tasks (bidding more aggressively), causing emergent behavior to macroscopically tend towards the desired direction; agents already under high load adjust in the opposite direction to avoid overload. It should be understood that a one-dimensional vector, formed by numerically encoding all current internal and local observation information of a single agent and concatenating them, is used to describe the agent's current state.

[0081] The emergence guidance layer does not make specific task allocation decisions, but only indirectly influences the collective behavior of the swarm by adjusting parameters, thus maintaining the robustness of the decentralized architecture. Even if the emergence guidance layer fails, the swarm can still operate normally (it just loses macro-directional guidance). By using a lightweight emergence guidance layer to achieve macro-guidance of the swarm's collective behavior without compromising decentralized robustness, the problem of uncontrollable emergence in purely decentralized solutions is solved.

[0082] To better illustrate the heterogeneous intelligent agent bee colony coordination method of the present invention, the following embodiments are provided: Example 1: Collaborative processing of composite tasks.

[0083] The heterogeneous agent swarm consists of 10 heterogeneous agents, including 3 code-based agents, 3 data analysis agents, 2 text-writing agents, and 2 general-purpose agents.

[0084] The task access gateway receives the task "Analyze sales data and generate a report". It extracts the task requirement vector R = [Data Analysis: 0.8, Text Writing: 0.6, Code: 0.2, ...]. This task requirement vector is broadcast to a heterogeneous agent swarm. Each agent receiving the task requirement vector calculates its own current task matching degree locally. The data analysis agents have a task matching degree of approximately 0.75, the text writing agents approximately 0.55, the code agents approximately 0.25, and the general agents approximately 0.5. The current bidding threshold is 0.5. Seven agents—three data analysis agents, two text writing agents, and two general agents—send bidding messages, while the three code agents remain silent (do not communicate). The task access gateway selects the data analysis agent with the highest task matching degree as the winning bidder. During execution, if the winning agent discovers that it needs the assistance of a subtask (which requires generating visual code), the winning agent will send the corresponding task to the task access gateway for bidding in the manner described above. Finally, the code-based agent will take over the subtask, forming a recursive collaboration.

[0085] Example 2: Communication topology evolution process.

[0086] like Figure 6 As shown, the heterogeneous agent swarm comprises 10 heterogeneous agents connected by a random graph, with an average degree of 3 (each agent connects to an average of 3 neighbors). After 50 tasks, the link weight between the data analysis agent and the text writing agent gradually increases (compound tasks involving analysis and reporting frequently require collaboration between the two). After 200 tasks: a clear collaborative subgroup is formed, with the data analysis agent and the writing agent forming a tightly connected core circle, while the links between the coding agents partially become dormant (coding tasks typically do not require collaboration between multiple coding agents; each can complete the task independently). A small-world maintenance mechanism maintains long-range reachability across subgroups (any two agents can be reached within a maximum of 3-4 hops).

[0087] Example 3: Multidimensional capability fingerprint vector update.

[0088] When a data analytics agent goes online, its initial multidimensional capability fingerprint vector is C = [Code: 0.3, Data Analysis: 0.85, Writing: 0.2, ...]. The agent completes its first data analysis task (task execution result). Update the quantified values ​​of data analysis in the multidimensional capability fingerprint vector to complete the second data analysis task (task execution result). Then update the quantified values ​​of data analysis in the multidimensional capability fingerprint vector to complete the third data analysis task (task execution result). Then, the quantitative value of data analysis in the multidimensional capability fingerprint vector is updated. Ultimately, the data analysis dimension increases to 0.90.

[0089] When the database connection tool of the data analysis agent suddenly disconnects, the quantification value of the data analysis drops sharply to 0.3, and the change in the multidimensional capability fingerprint vector is 0.6, which is greater than the threshold of 0.2. The agent then broadcasts a new multidimensional capability fingerprint vector to its neighbors.

[0090] If the data analysis agent does not perform the writing task for three days, the writing dimension decays at 0.995 per time step, decreasing from 0.2 to 0.17.

[0091] Example 4: The working mechanism of the emergent guiding layer.

[0092] The system-level goal is to guide a swarm of heterogeneous intelligent agents to prioritize responses to high-priority composite tasks.

[0093] Set the global target gradient vector and broadcast it at a low frequency: G = [priority_weight:+0.3, complexity_bonus:+0.2].

[0094] After receiving the gradient, each agent locally superimposes the global target gradient vector with its own state vector. For high-priority and high-complexity tasks, the bidding threshold is lowered (for a more proactive response), while for agents already under heavy load, the threshold is appropriately raised (to avoid overload). The entire process does not require modification to the core logic of any agent; it only indirectly affects the collective behavior direction through parameter fine-tuning, maintaining the robustness of the decentralized architecture.

[0095] like Figure 7 As shown, the present invention also provides a heterogeneous intelligent agent bee colony coordination system, characterized in that it is used to implement the heterogeneous intelligent agent bee colony coordination method as described above, the system comprising: a task access layer M100 and a heterogeneous intelligent agent bee colony M200; wherein, the task access layer M100 receives the current task to be processed and extracts task features to obtain a task requirement vector. The task access layer M100 broadcasts the task requirement vector to the heterogeneous agent bee colony M200, so that the agents that receive the task requirement vector can calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer M100 accordingly. The task access layer M100 collects bidding messages and arbitrates them accordingly to determine the winning intelligent agent, which then executes the currently pending task.

[0096] In the heterogeneous intelligent agent bee colony coordination system of this application, the functions of each module can be implemented by the specific implementation method of the heterogeneous intelligent agent bee colony coordination method described above, and the technical effects of the heterogeneous intelligent agent bee colony coordination method described above can be obtained, which will not be elaborated here.

[0097] This application embodiment also provides a heterogeneous intelligent agent bee colony coordination device, including a processor; a memory storing executable instructions of the processor; wherein the processor is configured to perform the steps of the heterogeneous intelligent agent bee colony coordination method by executing the executable instructions.

[0098] Those skilled in the art will understand that various aspects of this application can be implemented as systems, methods, or computer program products. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "platform."

[0099] The following reference Figure 8 To describe an electronic device 600 according to this embodiment of the present application. Figure 8 The electronic device 600 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0100] like Figure 5 As shown, the electronic device 600 is presented in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0101] The storage unit stores program code that can be executed by the processing unit 610, causing the processing unit 610 to perform the steps described in the above-described section of the heterogeneous intelligent agent bee colony coordination method according to various exemplary embodiments of this application. For example, the processing unit 610 can perform, as follows: Figure 1 The steps are shown in the figure.

[0102] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 6201 and / or a cache storage unit 6202, and may further include a read-only memory unit (ROM) 6203.

[0103] The storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0104] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0105] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0106] In the heterogeneous intelligent agent bee colony coordination device, when the program in the memory is executed by the processor, it implements the steps of the heterogeneous intelligent agent bee colony coordination method. Therefore, the device can also obtain the technical effects of the above-mentioned heterogeneous intelligent agent bee colony coordination method.

[0107] like Figure 9 As shown, an exemplary embodiment of this application also provides a computer program product 900. The computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described heterogeneous intelligent agent bee colony coordination method.

[0108] In one embodiment, the computer program product can be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. The readable storage medium can be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, including but not limited to: random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory, hard disk drive (HDD), solid-state drive (SSD), etc. Exemplarily, the computer program product can be implemented as a non-volatile storage medium storing a computer program, such as read-only memory, NAND flash memory, etc.

[0109] In one implementation, the computer program product can be an intangible product containing a computer program. For example, the computer program product can be implemented as a virtual digital product, such as an executable file, installation package, or other digital file storing the computer program.

[0110] Computer program code can be written in one or more programming languages. Examples of programming languages ​​include C, Java, C++, and Python. Program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, such as a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via an internet connection provided by a mobile network operator).

[0111] Computer programs can be carried or transmitted via signals such as electricity, magnetism, light, electromagnetic radiation, and infrared radiation. Electronic devices can convert signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program runs on an electronic device, its code is used to cause the electronic device to execute (more specifically, the processor of the electronic device to execute) the method steps of various exemplary embodiments of this application, such as the steps of the heterogeneous intelligent agent bee colony coordination method described above.

[0112] When the computer program is executed by the processor, it implements the steps of the heterogeneous intelligent agent bee colony coordination method described above. Therefore, the computer program product can also obtain the technical effects of the heterogeneous intelligent agent bee colony coordination method described above.

[0113] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for coordinating heterogeneous intelligent agent bee colonies, characterized in that, Includes the following steps: The task access layer receives the current tasks to be processed and extracts task features to obtain the task requirement vector. The task access layer broadcasts the task requirement vector to the heterogeneous agent swarm, so that the agents that receive the task requirement vector can calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer accordingly. The task access layer collects bidding messages and arbitrates them accordingly to determine the winning intelligent agent, which then executes the currently pending task.

2. The heterogeneous intelligent agent bee colony coordination method according to claim 1, characterized in that, The formula for calculating the task matching score is: ； in, The task matching score for agent i is the current task. For cosine similarity, This is the current multidimensional capability fingerprint vector of agent i. For the task requirement vector, This represents the current load status of agent i.

3. The heterogeneous intelligent agent bee colony coordination method according to claim 2, characterized in that, Each agent in the heterogeneous agent swarm updates its own multidimensional capability fingerprint vector according to an update rule, which includes: Rule 1: Whenever a task is completed, based on the task execution result, update the quantified value of the corresponding task processing capability in the current multidimensional capability fingerprint vector using an exponential moving average strategy: ； in, This represents the quantized value of task processing capability j in the current multidimensional capability fingerprint vector. For the result of task execution, For learning rate, The updated quantized value for dimension j; Rule 2: When the tools available to the agent change, recalculate the quantized value of the corresponding task processing capability and update the current multidimensional capability fingerprint vector; Rule 3: The quantified value of the task processing capability corresponding to a task that the agent has not executed within a preset time threshold decays exponentially.

4. The heterogeneous intelligent agent bee colony coordination method according to claim 1, characterized in that, The agent that receives the task requirement vector decides whether to send a bidding message to the task access layer based on its calculated current task matching score. Specifically, this includes: If the agent that receives the task requirement vector calculates its current task matching score, which exceeds the current bidding threshold, it sends a bidding message to the task access layer; otherwise, it does not send a bidding message to the task access layer.

5. The heterogeneous intelligent agent bee colony coordination method according to claim 1, characterized in that, The task access layer collects bidding messages and arbitrates them accordingly, specifically including: If the task access layer does not collect bidding messages, the task access layer lowers the bidding threshold and broadcasts the lowered bidding threshold to the heterogeneous intelligent agent swarm for re-bidding; If the task access layer collects multiple bidding messages, it calculates the difference between the matching scores of each pair of tasks based on the task matching scores contained in each of the multiple bidding messages. If all the calculated differences are less than the difference threshold, a second round of bidding is started; otherwise, the agent with the highest task matching score is determined as the winning agent.

6. The heterogeneous intelligent agent bee colony coordination method according to claim 1, characterized in that, The method further includes: periodically broadcasting a global target gradient vector to a heterogeneous agent swarm, so that agents that receive the global target gradient vector can fine-tune their parameters based on the global target gradient vector.

7. The heterogeneous intelligent agent bee colony coordination method according to claim 1, characterized in that, The heterogeneous agent swarm communicates based on its internal communication topology. The method further includes updating the communication topology of the heterogeneous agent swarm according to topology update rules. The topology update rules include: increasing the communication link weight between the two agents when they successfully cooperate to complete a task; decreasing the communication link weight between the two agents when the number of competition conflicts between them exceeds a preset threshold; and cutting off the communication link between the two agents when the time they do not communicate exceeds a preset communication time threshold.

8. A heterogeneous intelligent agent bee colony coordination system, characterized in that, The system for implementing the heterogeneous agent swarm coordination method according to any one of claims 1 to 7 includes: a task access layer and a heterogeneous agent swarm. The task access layer receives the current task to be processed and extracts task features to obtain a task requirement vector. The task access layer broadcasts the task requirement vector to the heterogeneous agent swarm, so that the agents that receive the task requirement vector can calculate their current task matching score based on their current multidimensional capability fingerprint vector and decide whether to send a bidding message to the task access layer accordingly. The task access layer collects bidding messages and arbitrates them accordingly to determine the winning intelligent agent, which then executes the currently pending task.

9. A heterogeneous intelligent agent bee colony coordination device, characterized in that, include: processor; A memory in which executable instructions of the processor are stored; The processor is configured to perform the steps of the heterogeneous agent bee colony coordination method according to any one of claims 1 to 7 by executing the executable instructions.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the heterogeneous intelligent agent bee colony coordination method according to any one of claims 1 to 7.

Images