A multi-agent algorithm network service generation method, device, equipment and medium

Abstract

This application relates to the field of artificial intelligence technology and discloses a method, apparatus, device, and medium for generating multi-agent computing network services. The method includes: determining the local cognitive health of each agent; determining the cognitive uncertainty residuals of each agent based on the local cognitive health and neighbor set, and outputting a cooperative state classification result through causal discrimination; responding to a global cooperative crisis state, correcting the hidden states of each agent used to represent cooperative characteristics using a large model; and generating a target computing network service based on the local decisions executed by each agent during the interaction process of each agent outputting escape actions, based on the re-determined cognitive uncertainty residuals, and assuming the cooperative state classification result has recovered. The technical solution provided by this application enables the evolution of computing network service scheduling from externally heuristic dependencies to endogenous escape capabilities, achieving the technical effect of autonomous closed-loop and robust operation of multi-agent cooperative scheduling.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and medium for generating multi-agent computing network services. Background Technology

[0002] Currently, Multi-Agent Reinforcement Learning (MARL) has become a core technology for adapting to computational network scheduling scenarios with high dynamics and exploding state spaces due to its advantages of low communication overhead and easily decomposed scheduling complexity in a distributed architecture. However, it is limited by the local perspective of the agents, making it difficult to converge the globally optimal policy and unable to handle long-tail rare faults. Therefore, the industry has introduced a hierarchical hybrid decision-making paradigm of "lightweight agent local execution + large language model (LLM) global intervention". When this paradigm is implemented, it still has three core defects of deep coupling: First, the intervention triggering mechanism cannot distinguish between normal environmental fluctuations and system cognitive uncertainty. It will misjudge normal traffic balance as abnormal due to the high policy entropy of the agents, wasting LLM call computing power and scheduling time. It will also cause the agents to overfit the input. First, low-entropy, absurd decision-making can lead to missed real-world faults, triggering a service outage avalanche. Second, text-based intervention introduces uncontrollable delays ranging from hundreds of milliseconds to seconds, violating the real-time requirements of network SLAs. Furthermore, forcibly replacing the action output of local agents disrupts the timing of Markov Decision Processes (MDPs), generating huge temporal differential errors that destroy learned steady-state policies. Additionally, since instructions are bound to specific node IDs and topological coordinates, repeated calls to the LLM are necessary when similar faults recur in other areas, causing call costs to increase linearly or even superlinearly with cluster size. Third, there is a complete lack of internalization mechanisms for intervention knowledge. The LLM intervenes only once, and underlying agents still need to rely on it when encountering similar faults. The system cannot autonomously evolve its strategies, nor can it reduce the burden of large model calls.

[0003] Therefore, in existing multi-agent collaborative decision-making in computing networks, how to solve the problems of false alarms and missed alarms caused by semantic confusion in intervention triggering mechanisms, and how text-based instruction interventions disrupt the real-time performance and policy continuity of the closed loop are urgent technical issues that need to be addressed. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for generating multi-agent computing network services, which solves the technical problems of false alarms and missed alarms caused by semantic confusion in intervention triggering mechanisms, and the disruption of closed-loop real-time performance and policy continuity caused by text-based instruction-based intervention.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a method for generating multi-agent computing network services, the method comprising: The acquired computing network service requests are mapped to multiple preset collaborative stages, and the corresponding intelligent agents in each stage are invoked to perform local decisions in order to determine the local cognitive health of each intelligent agent. The neighbor set of each agent is determined, and the cognitive uncertainty residual of each agent is determined based on the local cognitive health and the neighbor set. The causal discrimination is then performed based on the cognitive uncertainty residual to output the cooperative state classification result. In response to the classification result of the cooperative state of any agent being a global cooperative crisis state, the hidden state of each agent used to characterize the cooperative features is modified using a large model, so that each agent outputs an escape action to avoid cooperative risks based on the modified hidden state. During the interaction process in which each of the intelligent agents outputs the escape action, based on the redefined cognitive uncertainty residual, and if it is determined that the cooperative state classification result has recovered to a non-global cooperative crisis state, a target computing network service is generated based on the local decisions executed by each of the intelligent agents.

[0006] In one implementation, determining the local cognitive health of each of the intelligent agents includes: During the execution of the local decision, the action probability distribution of each agent in the current observation state is obtained, and the random uncertainty measure of each agent is determined based on the maximum probability value in the action probability distribution. A random mask is applied to the hidden layer of the policy network of each agent, a preset number of forward propagations are performed on the current observation state, and the cognitive uncertainty measure of each agent is determined based on the degree of difference between the perturbation policy distribution output by each forward propagation and the average policy distribution of all forward propagations. By product coupling the accidental uncertainty measure and the cognitive uncertainty measure, a joint cognitive crisis index is obtained. The joint cognitive crisis index is normalized to obtain the local cognitive health of each agent.

[0007] In one implementation, determining the cognitive uncertainty residuals of each agent based on the local cognitive health and the neighbor set includes: For any current agent's neighbor set, obtain the semantic control packet broadcast by each neighbor agent in the neighbor set, wherein the semantic control packet contains at least the local cognitive health of each neighbor agent; The average cognitive health of the neighborhood is determined based on the average local cognitive health of all neighbor agents in the neighbor set. Based on the difference between the local cognitive health of the current agent and the average cognitive health of the neighborhood, the cognitive uncertainty residual corresponding to the current agent is determined.

[0008] In one implementation, the step of outputting a cooperative state classification result based on the causal discrimination of the cognitive uncertainty residual includes: When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is less than the second preset threshold, or when the cognitive uncertainty residual is greater than or equal to the third preset threshold, it is determined to be a non-global collaborative crisis state. When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is greater than or equal to the fourth preset threshold, it is determined to be a global collaborative crisis state.

[0009] In one implementation, the step of modifying the hidden states of each agent used to represent cooperative features using a large model includes: Extract the hidden states of the global cooperative crisis state from the graph neural network encoder of each of the agents, and determine the geometric statistics of the hidden states; Based on the geometric statistics, structured cue words are constructed, and the structured cue words are input into the large model to generate source code containing basic tensor operations; The source code is compiled to obtain tensor operators; The tensor operator is injected into the forward propagation path of the policy network of each agent to obtain the corrected hidden state.

[0010] In one implementation, during the interaction process where each of the intelligent agents outputs the escape action, based on the redefined cognitive uncertainty residual, and in the case that the cooperative state classification result is determined to have recovered to a non-global cooperative crisis state, a target computing network service is generated based on the local decisions executed by each of the intelligent agents, including: During the interaction process in which each agent outputs the escape action, the gradient of the policy loss function of each agent is determined, and the gradient is backpropagated to the policy network of each agent via the tensor operator through backpropagation to update the policy parameters of the policy network. Based on the updated policy networks of each agent, the cognitive uncertainty residuals are redefined. Causal discrimination is performed based on the redefined cognitive uncertainty residuals. If it is determined that the cooperative state classification result has been restored to a non-global cooperative crisis state, the tensor operator is removed from the forward propagation path of the policy network. Local decisions are made based on the updated policy network of each agent to generate the target computing network service.

[0011] In one implementation, during the interaction process in which each of the intelligent agents outputs the escape action, the method further includes: The gradient of the policy loss function of each agent is determined, and the gradient is backpropagated to the policy network of each agent via the tensor operator through backpropagation to update the policy parameters of the policy network. The policy entropy and value loss of each agent's policy network are monitored in real time. When the policy entropy recovers to above a preset threshold and the value loss converges, the tensor operator is removed from the forward propagation path of the policy network.

[0012] Secondly, embodiments of this application provide a multi-agent computing network service generation apparatus, the apparatus comprising: The local decision evaluation unit is used to map the acquired computing network service request to multiple preset collaborative stages, and call the corresponding intelligent agents of each stage to perform local decisions in order to determine the local cognitive health of each intelligent agent. The cooperative state discrimination unit is used to determine the neighbor set of each agent, and to determine the cognitive uncertainty residual of each agent based on the local cognitive health and the neighbor set, so as to perform causal discrimination based on the cognitive uncertainty residual and output the cooperative state classification result. The large model correction unit is used to correct the hidden states of each agent that represent the cooperative features in response to the classification result of the cooperative state of any agent being a global cooperative crisis state. This is done by using the large model to correct the hidden states of each agent, so that each agent outputs an escape action to avoid cooperative risks based on the corrected hidden states. The state recovery and service generation unit is used to generate target computing network services based on the local decisions executed by each of the intelligent agents during the interaction process in which each of the intelligent agents outputs the escape action, based on the re-determined cognitive uncertainty residual, and in the case that the cooperative state classification result is determined to be restored to a non-global cooperative crisis state.

[0013] Thirdly, embodiments of this application provide a computer device, including: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to perform the multi-agent computing network service generation method described above.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the multi-agent network service generation method described above.

[0015] The technical solutions provided by one or more embodiments of this application, by mapping network service requests to multiple collaborative stages and calling the corresponding agents in each stage to perform local decisions, can effectively avoid cognitive overload of a single agent when handling complex cross-domain tasks, achieve fine-grained allocation of computing resources, and provide accurate quantitative basis for subsequent collaborative intervention by real-time quantitative output of local cognitive health of each agent. On this basis, by defining the neighbor set of each agent and calculating the cognitive uncertainty residual, it can accurately filter the random uncertainty caused by inherent environmental noise, extract the cognitive uncertainty caused by knowledge blind spots or sudden anomalies, and then output the global collaborative state classification result through causal discrimination, providing an objective decision threshold for whether to trigger macro-level intervention of the large model. When a global collaborative crisis is identified, the hidden states representing collaborative features of each agent are geometrically or semantically modified using a large model. This is equivalent to injecting prior knowledge from a global perspective to break the policy manifold deadlock and guide each agent to output globally oriented escape actions to escape collaborative deadlock or dangerous areas. Finally, after intervention by the large model, the cognitive uncertainty residuals are monitored in real time. When the collaborative state is confirmed to have recovered to a non-global collaborative crisis state, the intervention by the large model is automatically lifted and control is returned. This allows each agent to activate local decisions based on the updated parameters and ultimately generate the target computing network service. This realizes the evolution of computing network service scheduling from external heuristic dependence to endogenous escape capability, achieving autonomous closed-loop and robust operation of multi-agent collaborative scheduling. Attached Figure Description

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

[0017] Figure 1 A flowchart illustrating a method for generating multi-agent computing network services provided in this application embodiment; Figure 2 A flowchart of step S1 provided in the embodiments of this application; Figure 3 A flowchart for determining the cognitive uncertainty residuals of each agent provided in an embodiment of this application; Figure 4 A flowchart for causal determination is provided for embodiments of this application; Figure 5 A flowchart of step S5 provided in an embodiment of this application; Figure 6 A flowchart of step S7 provided in an embodiment of this application; Figure 7 This is a schematic diagram of the method framework provided in the embodiments of this application; Figure 8 Technical schematic diagrams provided for embodiments of this application; Figure 9 A block diagram of a multi-agent computing network service generation device provided in an embodiment of this application; Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] With the rapid iteration of artificial intelligence technology, large-scale multi-agent systems have demonstrated enormous application potential in fields such as drone swarm collaboration, intelligent traffic scheduling, and complex logistics management. Among these, multi-agent reinforcement learning (MARL), as a core paradigm supporting agent collaborative behavior, can learn optimal cooperation strategies through continuous interaction between agents and the environment, and is one of the core research directions in current multi-agent systems. Meanwhile, the in-depth development of the "computing power network" concept has driven the deep integration of computing power resources and network transmission from their traditionally fragmented nature. Modern computing network systems need to address the end-to-end service management lifecycle, covering the entire process of intent resolution, node selection, service deployment, operation monitoring, and service optimization, providing integrated intelligent scheduling capabilities. Therefore, introducing MARL into computing network service scheduling scenarios has become an important technical path to adapt to the complex management and control needs of computing networks.

[0020] However, the computing network environment itself is highly dynamic, heterogeneous, and open. Problems such as sudden fluctuations in data center computing load, instantaneous congestion of backbone links, and even physical downtime at the node level occur frequently, making the computing network service state space exhibit extreme complexity and long-tail distribution characteristics. Traditional scheduling methods based on rules or single reinforcement learning are limited by limited model capacity and prior knowledge, making it difficult to balance real-time performance and generalization in the computing network environment with its exploding state space. In particular, when faced with unseen out-of-distribution states, decision failures are very likely to occur, leading to service level agreement (SLA) defaults. To address this, the industry has introduced a distributed architecture based on multi-agent reinforcement learning. By decomposing the global scheduling complexity and having multiple lightweight agents each responsible for real-time control of local domains, this avoids the problem of soaring network-wide communication overhead under centralized control and alleviates the dimensionality curse caused by the exponential growth of the joint action space with the number of agents. However, solutions relying solely on MARL still have an insurmountable "cognitive gap": agents, limited by their local perspectives, struggle to converge to a globally consistent optimal strategy. This makes it difficult to achieve cross-domain global collaborative decision-making, effectively handle rare faults with long-tail distributions, and adapt to the complex management and control requirements of computing network scenarios.

[0021] Recently, large-scale language models, with their massive knowledge reserves and powerful zero / few-shot generalization reasoning capabilities, have provided a new path to overcome the aforementioned cognitive gap. The industry has therefore proposed a hierarchical hybrid decision-making paradigm of "lightweight agent local execution + LLM global intervention": the LLM plays a macro-level decision-making role, passively waiting for the agent to request assistance, providing high-level guidance only at key decision points or when anomalies occur, while the lightweight agent acts as a micro-level executor responsible for high-frequency closed-loop control. However, this paradigm still suffers from three deeply coupled core defects in practical implementation: First, the intervention triggering mechanism suffers from severe semantic confusion, failing to decouple the accidental uncertainty of the environment from the cognitive uncertainty of the system. This can lead to significant waste of API computing power and scheduling time, even under normal fluctuations such as balanced link bandwidth, where the agent outputs high policy entropy and mistakenly triggers the LLM; and when node hardware failures cause abnormal feature inputs, the local agent overfits and outputs low-entropy but absurd deployment decisions, ultimately causing a service outage avalanche. Secondly, the "text-command-based" intervention mode severely disrupts the real-time performance and policy continuity of the computing network closed loop, and lacks spatial generalization capabilities: text parsing, API calls, and network transmission introduce uncontrollable delays of hundreds of milliseconds or even seconds, violating the stringent real-time requirements of the computing network SLA; forcibly replacing the action output of local agents with text commands abruptly cuts off the Markov Decision Process (MDP) in time, generating huge temporal difference (TD) errors, which can easily destroy the steady-state policy logic already learned by the lightweight agent; in addition, text commands are usually strictly bound to specific node IDs and absolute topological coordinates. When similar outages are reproduced in other availability zones, the LLM needs to be called again to generate new commands, causing the cost of LLM calls to increase linearly or even superlinearly with the task duration and cluster size. Thirdly, the existing framework completely lacks a mechanism for internalizing intervention knowledge. LLM intervention is only a one-time intervention. After the intervention, the underlying lightweight agent is still unaware of similar failures, and will still need to rely on LLM when encountering similar scenarios in the future. The system cannot achieve autonomous policy evolution during operation, nor can it debind and reduce the burden of large model calls.

[0022] In summary, in existing multi-agent collaborative decision-making in computing networks, the urgent technical problems to be solved are how to address the false alarms and missed alarms caused by semantic confusion in intervention triggering mechanisms, and how text-based instruction-based interventions disrupt the real-time performance and policy continuity of the closed loop.

[0023] According to an embodiment of this application, a method for generating multi-agent network services is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] This embodiment provides a method for generating multi-agent computing network services. Figure 1 A flowchart of a multi-agent computing network service generation method provided in this application embodiment is shown below. Figure 1 As shown, the process includes the following steps: Step S1: Map the obtained computing network service request to multiple preset collaborative stages, and call the corresponding intelligent agents of each stage to perform local decisions in order to determine the local cognitive health of each intelligent agent.

[0025] Specifically, a computing network service request refers to an end-to-end service request instruction initiated by a user that spans computing and network resources. For example: "Deploy a video transmission service between location A and location B that meets the requirements of latency <10ms and bandwidth >1Gbps."

[0026] The collaboration phase refers to the process of decoupling the globally complex computing network service generation task into multiple ordered sub-task units according to logical functions to solve the problem of long-link service delivery. In this embodiment, the preset collaboration phases specifically include: intent parsing phase, node selection phase, service deployment phase, operation monitoring phase, and service optimization phase.

[0027] Upon receiving a request for network services, it is mapped to the five collaborative stages mentioned above. Each stage is configured with a pre-built, lightweight agent model (agent) trained on historical data. These agents are responsible for executing local decisions and selecting and calling standardized APIs (such as resource query APIs, configuration distribution APIs, etc.) from a dedicated toolkit, thereby reducing the huge global search space to the local discrete action space of each stage.

[0028] In this embodiment, the agent is based on a policy network trained using reinforcement learning. This agent performs tasks independently under normal conditions, without frequently relying on a large model. Its input is the local observed state. i The output is the action probability distribution. Observation status o i This refers to the local data of the computing network environment perceived by the intelligent agent at the current moment, such as the local network topology, node resource utilization (CPU / memory), link bandwidth, latency, and packet loss rate—dynamic environmental data. Action probability distribution. The agent's policy network in the current observation state o i Below, for the probability set of all available computing network scheduling decisions (action a), the computing network scheduling decisions include selecting a specific link for data flow forwarding, selecting an edge node for computing power task offloading, etc.

[0029] Building upon this foundation, by invoking dedicated agents at each stage to execute local decisions in parallel or sequentially, not only is cognitive overload effectively avoided when a single model handles complex cross-domain tasks, but also fine-grained allocation of computing resources is achieved. More importantly, during decision-making, each agent can assess its own confidence level and state deviation regarding its understanding of the current local environment in real time, thereby quantifying the local cognitive health. This health metric, serving as a core feedback signal reflecting the reliability of agent decisions and the stability of collaborative states, provides precise quantitative evidence for subsequent determination of whether to trigger global collaborative intervention and for dynamically adjusting collaborative strategies. This ensures that the entire computing network service scheduling process is highly autonomous yet controllable and robust.

[0030] Step S3: Determine the neighbor set of each agent, and determine the cognitive uncertainty residual of each agent based on the local cognitive health and the neighbor set, so as to output the cooperative state classification result based on the causal discrimination of the cognitive uncertainty residual.

[0031] Specifically, firstly, the neighbor set of each agent is dynamically defined based on a pre-set distributed communication network (such as the adjacency matrix of a directed graph), thereby clarifying the local boundaries of information interaction. Then, each agent uses a consensus algorithm to fuse its local cognitive health with the cognitive state of its neighboring nodes, deriving a cognitive uncertainty residual. This residual essentially reflects the degree of deviation between individual decisions and group consensus, as well as the potential crisis tension. It effectively filters out accidental uncertainties caused by inherent environmental noise and accurately extracts cognitive uncertainties caused by knowledge blind spots or sudden anomalies. Based on this, the residual is used as a core feature for causal discrimination, tracing the root cause of state deterioration, and outputting a global collaborative state classification result (e.g., determining whether the current state is in a non-global collaborative crisis state or a global collaborative crisis state). This mechanism not only achieves a quantitative assessment of the credibility of multi-agent decisions in a distributed environment but also provides a rigorous and objective decision threshold for whether to trigger macro-level logic injection and operator intervention in the Large Model (LLM).

[0032] Step S5: In response to the classification result of the cooperative state of any agent being a global cooperative crisis state, the hidden state of each agent used to represent the cooperative characteristics is modified using a large model, so that each agent outputs an escape action to avoid cooperative risks based on the modified hidden state.

[0033] Specifically, when the collaborative state classification result of any agent is determined to be a global collaborative crisis, it indicates that conventional local optimization has reached a bottleneck. At this point, the Large Model (LLM) is triggered as a macro-strategic brain to intervene by reducing dimensionality. The LLM does not directly take over the specific execution at the underlying level, but rather applies geometric or semantic-level corrections and perturbations to the hidden states used by each agent to represent collaborative features. This hidden state correction is equivalent to introducing prior knowledge from a global perspective to agents trapped in local optima or cognitive blind spots, breaking the original policy manifold deadlock. Subsequently, each agent performs forward reasoning based on the corrected hidden states, thereby outputting escape actions with global guidance. These escape actions can guide agents to escape the current collaborative deadlock or dangerous area, realizing effective linkage from individual micro-decision-making to group macro-risk avoidance. This lays a crucial intervention foundation for subsequent agents to internalize the logic of the LLM through gradient backpropagation and ultimately restore autonomous collaborative capabilities.

[0034] Step S7: During the interaction process of each agent outputting escape actions, based on the redefined cognitive uncertainty residual, and if the cooperative state classification result is determined to be restored to a non-global cooperative crisis state, the target computing network service is generated based on the local decisions executed by each agent.

[0035] Specifically, after the large model completes the hidden state correction and guides the agents to output escape actions, it enters the dynamic verification phase. By calculating and monitoring the redefined cognitive uncertainty residuals in real time, the degree of absorption and internalization of the escape strategy by the agents is accurately assessed. When it is determined that the cooperative state classification result has substantially recovered to a non-global cooperative crisis state, it means that each agent has successfully overcome the current cognitive blind spot and cooperative deadlock. At this point, the macro-level intervention of the large model is automatically released, and control is completely returned to each agent. Each agent then reactivates and executes efficient local decisions based on the updated weights. After the complete process of distributed cooperative decision-making, cognitive health assessment, and global cooperative crisis intervention, a deployed and continuously running computing network service instance will be output, including the service deployment configuration on each node, real-time resource scheduling decisions, dynamically optimized operating parameters, etc. This mechanism not only ensures that the computing network service can seamlessly connect and continuously and stably generate after crisis intervention, but also marks the successful evolution from externally heuristic dependence to endogenous escape capability, realizing autonomous closed-loop and robust operation of multi-agent cooperative scheduling.

[0036] This embodiment provides a multi-agent computing network service generation method. By mapping computing network service requests to multiple collaborative stages and calling the corresponding agents in each stage to perform local decisions, it can effectively avoid cognitive overload of a single agent when handling complex cross-domain tasks, achieving fine-grained allocation of computing resources. Simultaneously, the real-time quantitative output of local cognitive health by each agent provides accurate quantitative basis for subsequent collaborative intervention. Furthermore, by defining the neighbor set of each agent and calculating the cognitive uncertainty residual, it can accurately filter out accidental uncertainties caused by inherent environmental noise, extract cognitive uncertainties caused by knowledge blind spots or sudden anomalies, and then output global collaborative state classification results through causal discrimination, providing an objective decision threshold for whether to trigger macro-level intervention in a large model. When a global collaborative crisis is identified, the hidden states representing collaborative features of each agent are geometrically or semantically modified using a large model. This is equivalent to injecting prior knowledge from a global perspective to break the policy manifold deadlock and guide each agent to output globally oriented escape actions to escape collaborative deadlock or dangerous areas. Finally, after intervention by the large model, the cognitive uncertainty residuals are monitored in real time. When the collaborative state is confirmed to have recovered to a non-global collaborative crisis state, the intervention by the large model is automatically lifted and control is returned. This allows each agent to activate local decisions based on the updated parameters and ultimately generate the target computing network service. This realizes the evolution of computing network service scheduling from external heuristic dependence to endogenous escape capability, achieving autonomous closed-loop and robust operation of multi-agent collaborative scheduling.

[0037] Figure 2 The flowchart for step S1 provided in the embodiments of this application may include the following steps: Step S11: During the execution of local decision-making, obtain the action probability distribution of each agent under the current observation state, and determine the random uncertainty measure of each agent based on the maximum probability value in the action probability distribution.

[0038] Specifically, the current observation state (o i This refers to the local data of the computing network environment perceived by the intelligent agent at the current moment, such as the local network topology, node resource utilization (CPU / memory), link bandwidth, latency, and packet loss rate, among other dynamic environmental data. Action probability distribution. The agent's policy network in the current observation state o i Below, for the probability set of all available computing network scheduling decisions (action a), the computing network scheduling decisions include selecting a specific link for data flow forwarding, selecting an edge node for computing task offloading, etc. Random uncertainty measurement (U alea This is used to characterize the inherent randomness or noise of the computing network environment. For example, when there are multiple equivalent paths with approximately equal bandwidth and latency, the model may have difficulty making a choice. This is due to the randomness of the environment itself, rather than a failure of model cognition.

[0039] In network scheduling scenarios, agents may face equal probabilities of choosing multiple equivalent paths or minor fluctuations in traffic during normal business operations. Directly calculating the overall distribution entropy in such cases would result in a high entropy value, easily misinterpreted as a cognitive crisis. Therefore, this step extracts the maximum probability value from the action probability distribution for calculation, as shown in the following formula: Among them, U alea (o i () represents the current observation state. i The measure of random uncertainty; when hour, Approaching 0 indicates that the agent has a clear optimal tendency for the current decision; when hour( For example, the preset motion space size =4), If the value approaches 1, it indicates that there are multiple approximately equivalent optimal solutions.

[0040] Step S13: Apply a random mask to the hidden layer of the policy network of each agent, perform a preset number of forward propagations on the current observation state, and determine the cognitive uncertainty measure of each agent based on the degree of difference between the perturbation policy distribution output by each forward propagation and the average policy distribution of all forward propagations.

[0041] Specifically, random mask (M k ) is a binary matrix used to randomly discard or retain some neurons in the hidden layer of the policy network during the k-th forward propagation, thereby simulating small perturbations in the agent's model parameters. The perturbation policy distribution ( (This refers to the situation where a random mask M is applied.) k Then, the policy network applies the same observed state o i The output action probability distribution with random perturbations. Average policy distribution ( This refers to the mean of the distribution of all perturbation strategies after K forward propagations. Cognitive uncertainty measure (U0) epis The cognitive uncertainty metric U is used to characterize the agent's degree of ignorance regarding the current observed state. This metric increases significantly when encountering unknown faults outside the training distribution, such as sudden node crashes or physical link interruptions. The training distribution refers to the set of statistical characteristics of the state-action-reward samples experienced by each agent during offline training or online learning, including but not limited to node load levels, link latency ranges, task arrival patterns, and known fault types. When unknown abnormal events (such as physical node crashes or link interruptions) occur in the actual operating environment beyond this training distribution, the policy network, lacking relevant experience, experiences a significant increase in the output variance of multiple forward propagations, thus increasing the cognitive uncertainty metric U. episIncrease.

[0042] Specifically, a random mask M is applied to the hidden layer of the policy network. k Secondly, maintain the current observation state of the input. i Without changing the policy distribution, perform K random forward propagations to obtain K perturbed policy distributions; then, calculate the average policy distribution: The distribution of the perturbation policy can be obtained using existing methods such as Monte Carlo Dropout (MC Dropout). In the Dropout layer of the hidden layer of the policy network, a binary random mask M is generated with a preset drop probability p. k Each mask is sampled independently for the same observation state o i Keeping the input constant, perform K forward propagations, applying a different random mask M each time. k We obtain K perturbation strategy distributions.

[0043] Calculate the KL divergence between each perturbation policy distribution and the average policy distribution, and take the mean as a measure of cognitive uncertainty. When o i When the random mask is located within the training distribution, it has little impact on the output, and the KL divergence remains low; when o i When outside the training distribution (OOD state), random perturbations cause drastic fluctuations in the output, U epis (o i () significantly increased.

[0044] Step S15: The accidental uncertainty measure and the cognitive uncertainty measure are productized and coupled to obtain the joint cognitive crisis index.

[0045] Specifically, the Joint Cognitive Crisis Indicator (UCI) raw This is used to comprehensively assess the degree of real crisis currently faced by the agent. Through product logic, it ensures that a real crisis is only determined when the agent simultaneously exhibits hesitation in choosing an action and a lack of awareness of its state.

[0046] Logically couple the two metrics mentioned above, and the calculation formula is as follows: This mechanism achieves precise semantic-level filtering: if the random uncertainty is high but the cognitive uncertainty is low (such as normal exploration of equivalent paths), or if the cognitive uncertainty is high but the random uncertainty is low (such as false confidence caused by model overfitting), the joint indicators remain low and do not trigger intervention; only when both are high at the same time will the joint indicators rise significantly and be marked as a real cognitive crisis, fundamentally preventing large model mis-triggering caused by normal environmental fluctuations.

[0047] Step S17: Normalize the joint cognitive crisis index to obtain the local cognitive health of each agent.

[0048] Specifically, local cognitive health ( In essence, local cognitive crisis scalar represents a local cognitive health score. Local cognitive health compresses cognitive anomalies in a complex, high-dimensional state space into a unified, low-dimensional semantic scalar, with values ​​ranging from (0,1). A higher value indicates lower cognitive health (more severe crisis). Normalization is performed using a fractional compression function, and the calculation formula is as follows: Here, 'c' is a preset sensitivity adjustment parameter (constant). A larger 'c' value indicates stronger compression, greater conservatism, and a lower likelihood of triggering a crisis; a smaller 'c' value indicates greater sensitivity. Through this step, a stable and comparable scalar local cognitive health score is ultimately output. This provides a reliable data foundation for subsequent collaborative crisis identification based on neighborhood interaction.

[0049] In this embodiment, during the execution of local decision-making, by obtaining the action probability distribution of each agent under the current observation state and calculating the random uncertainty measure based on the maximum probability value, it can effectively eliminate decision hesitation caused by inherent randomness such as multiple equivalent path selection and traffic micro-jitter in the computing network environment, and avoid misjudging the difficult choice characteristics of the environment itself as a cognitive crisis. On this basis, by applying a random mask to the hidden layer of the policy network to perform multiple forward propagations, and calculating the cognitive uncertainty measure based on the KL divergence between the perturbation policy distribution and the average policy distribution, it can accurately characterize the degree of ignorance of the agent regarding unknown faults outside the training distribution, such as sudden node downtime and physical link interruption, thereby distinguishing between agent cognitive failure and random fluctuations in the environment. Furthermore, by productizing and coupling the measures of accidental uncertainty and cognitive uncertainty, a joint cognitive crisis index is obtained, achieving semantic-level precise filtering. This ensures that a real crisis is only identified when the agent simultaneously exhibits hesitation in action selection and a lack of state cognition, fundamentally eliminating false triggering of large models due to normal environmental fluctuations or false confidence caused by model overfitting. Finally, the joint cognitive crisis index is normalized to obtain local cognitive health, compressing cognitive anomalies in the complex high-dimensional state space into a unified low-dimensional semantic scalar. This provides a stable and comparable reliable data foundation for subsequent collaborative crisis discrimination based on neighborhood interaction, thereby significantly improving the accuracy of cognitive crisis detection and the precision of intervention triggering in computing network service scheduling.

[0050] Figure 3 The flowchart for determining the cognitive uncertainty residuals of each agent provided in the embodiments of this application may include the following steps: Step S311: For any current agent's neighbor set, obtain the semantic control packet broadcast by each neighbor agent in the neighbor set. The semantic control packet contains at least the local cognitive health of each neighbor agent.

[0051] Specifically, the neighbor set (N) i A semantic control package (m) refers to the set of agents in the network topology that have a direct communication connection or topological adjacency with the current agent i. j () refers to the lightweight data packet broadcast by neighboring agent j to current agent i. In this embodiment, this data packet abandons the transmission of high-dimensional state features in traditional graph neural networks and only contains low-dimensional cognitive state identifiers, including at least the local cognitive health of the neighboring agents.

[0052] While preserving its local high-dimensional state embedding for internal feature modeling, agent i receives the neighbor set N through the communication link. i The semantic control packets broadcast by each neighbor node j: in, Local cognitive health level (value range 0 to 1). k j k is the action index set, obtained by topological encoding of the policy output probability vector. j It is not a simple action number, but a low-dimensional but structure-preserving decision topology descriptor used to characterize the local geometry and decision topology relationships of the current agent's decision space. For example, when multiple actions have similar probabilities, their ranking structure will exhibit obvious "multi-peak competition" characteristics; when the probability of a certain action is much higher than that of other actions, it corresponds to a stable single-peak decision state.

[0053] Specifically, we take the Top-K action index set of the policy distribution of the neighboring nodes (for example, if agent i's action space is 100 and agent j's action space is 50, we take their Top-K action index set respectively (when K=4, k...). i =[0.85,0.10,0.03,0.02],k j =[0.34,0.32,0.25,0.09]), and then calculate the confidence difference between adjacent actions {Δp1,Δp2,...,Δp K This is an example (action space 50, 100, etc., K=4) to explain the algorithm flow above: k iIt is a compressed representation further extracted from the average policy distribution; the action space size (100 / 50) is a preset hyperparameter, i.e., the dimension of the policy network output layer. Agent i has 100 candidate actions, and the network outputs a 100-dimensional probability vector (normalized by Softmax, with a sum of 1).

[0054] Sort the average policy distribution in descending order of probability, and calculate k by taking the indices and probability values ​​of the first K (e.g., K=4) actions. i Example: In the 100-dimensional vector of agent i, the four actions with the highest probabilities are action 37 (0.85), action 12 (0.10), action 71 (0.03), and action 5 (0.02). Then k i The encoding is [0.85, 0.10, 0.03, 0.02].

[0055] Confidence difference: After sorting the Top-K probability values ​​in descending order, subtract adjacent terms, for example: intelligent agent i(k) i =[0.85,0.10,0.03,0.02]): Δp1 = 0.85 - 0.10 = 0.75 (maximum); Δp2 = 0.10 - 0.03 = 0.07; Δp3 = 0.03 - 0.02 = 0.01; Difference sequence: [0.75, 0.07, 0.01] → Single peak dominance (first action probability is absolutely leading).

[0056] Agent j(k) j =[0.34,0.32,0.25,0.09]): Δp1 = 0.34 - 0.32 = 0.02 (minimal) Δp2 = 0.32 - 0.25 = 0.07 Δp3 = 0.25 - 0.09 = 0.16 Differential sequence: [0.02, 0.07, 0.16] → Multi-peak competition (the probabilities of the first three actions are similar, with no obvious dominance).

[0057] If a certain confidence difference is large, it indicates the existence of an absolutely dominant action, and the policy distribution exhibits a sharp unimodal geometry. If all confidence differences are small, it indicates that multiple actions have similar probabilities, and the policy distribution exhibits a flat multimodal competitive geometry. Compared to directly transmitting the complete action probability vector, the action confidence ranking index has lower communication overhead and can maintain cross-agent consistency and comparability across different action dimensions and task scales, providing crucial action space geometric priors for constructing crisis contexts in subsequent large-scale models.

[0058] It should be noted that traditional graph attention networks achieve collaborative modeling by broadcasting high-dimensional state features. However, in out-of-distribution (OOD) training states, anomalous features generated by locally failing nodes can easily lead to the rapid spread of erroneous information throughout the topology. Therefore, this embodiment redefines the attention weight calculation formula: Where, α ij Let be the attention weights of agent i to neighbor node j; W is the shared linear mapping matrix, and are the learnable parameters of the policy network, obtained through gradient descent optimization during the training phase, used to map the local high-dimensional states x of agent i and neighbor j. i ,x j Extracted as low-dimensional robust topological features; a is a learnable attention vector; | is the concatenation operation; LeakyReLU is a non-linear activation function; This refers to the local cognitive health level received from neighboring nodes; This represents the uncertainty penalty coefficient.

[0059] Using this formula, when the uncertainty of neighbor node j... When the concentration increases, the molecules in the molecules This term causes the exponential value to decrease drastically, thus exponentially suppressing the attention weight of that neighbor. Even if an abnormal node broadcasts disordered features, it will be automatically muted due to its high uncertainty value, preventing erroneous information from polluting the entire topology.

[0060] The anti-interference attention weight α is calculated. ij Then, agent i first uses the shared linear mapping matrix W to map the local features x of neighbor node j. j Perform a linear transformation to obtain the robust topological feature h. j =Wx j Subsequently, with α ij The weighted summation of the robust topological features of the neighbors is performed using the following formula: Through this aggregation process, agent i can dynamically adjust its dependence on neighbor features based on the uncertainty of its neighbors, thereby enhancing the robustness of local features.

[0061] In subsequent operations, k i Serving as a semantic bridge connecting local nodes and the larger model, it constitutes the core input of the topology-level crisis response mechanism. This is achieved by aggregating the k-values ​​broadcast by neighbors. i By monitoring the overall situation, when multiple nodes consistently exhibit a low-confidence, multi-peak competitive ranking pattern, a topology-level cooperative stagnation phenomenon can be identified. Subsequently, these aggregated k... iAs a geometric prior in the action space, it is input into the large model to construct the crisis context. Based on this, the large model perceives the predicament of being trapped in flat local optima, generates an implicit tensor operator, and distributes it to the stagnant nodes. This operator applies a small geometric bias to the local feature space, disrupting the original multimodal equilibrium and prompting the agent to make deterministic decisions, thereby driving the system to successfully escape the cooperative stagnation predicament.

[0062] Step S313: Determine the average cognitive health of the neighborhood based on the average local cognitive health of all neighbor agents in the neighbor set.

[0063] Specifically, the neighborhood average cognitive health refers to the average level of local cognitive health of all neighboring agents within the local topological region where the current agent is located, and is used to characterize the overall cognitive state of the current local cooperative environment.

[0064] The current agent i extracts the data from each of its neighboring agents j∈N. i Local cognitive health And calculate their arithmetic mean, using the following formula: in, The average cognitive health of the neighborhood; This represents the total number of nodes in the neighbor set. This average value reflects the overall cognitive risk level of the agent's current surrounding environment.

[0065] Step S315: Determine the cognitive uncertainty residual corresponding to the current agent based on the difference between the local cognitive health of the current agent and the average cognitive health of the neighborhood.

[0066] Specifically, cognitive uncertainty residuals (Δu) i This refers to the difference between the current local cognitive health of the agent and the average cognitive health of its neighborhood. Essentially, this residual is a causal discrimination based on topological cross-validation, used to determine whether the current cognitive anomaly is an isolated local fault or has evolved into a systemic crisis across nodes.

[0067] The current local cognitive health of agent i Compared with the average cognitive health of the neighborhood The formula for finding the difference is as follows: like This indicates that the current agent exhibits significant cognitive abnormalities, while surrounding nodes maintain normal cognitive abilities (local noise); if and The continuous increase indicates that the cognitive blind spot has formed a synchronous resonance in the topological network, that is, a cascading failure or cooperative deadlock has occurred across multiple stages.

[0068] For the neighbor set of any current agent, this embodiment acquires the semantic control packets containing the local cognitive health degree and the action confidence ranking index broadcast by each neighbor agent, abandons the transmission of high-dimensional state features in the traditional graph neural network, and realizes lightweight communication. The neighborhood average cognitive health degree is obtained by calculating the average value of the local cognitive health degrees of all neighbor agents in the neighbor set, and the cognitive uncertainty residual is obtained by subtracting the neighborhood average value from the local cognitive health degree of the current agent, so as to realize causal discrimination based on topology cross-validation, which can accurately distinguish isolated local faults from cross-node systematic crises. When the cognitive uncertainty residual is significantly greater than zero, it indicates that the current agent has local noise; when the cognitive uncertainty residual approaches zero and the local cognitive health degree continues to increase, it indicates that the cognitive blind area has formed a cascading failure or collaborative deadlock of synchronous resonance in the topology network, thereby providing a rigorous and objective decision threshold for whether to trigger the macro intervention of the large model subsequently, and enabling the large model to perceive the collaborative stagnation dilemma through the aggregated semantic control packet and generate implicit tensor operators to break the multimodal balance and drive the system to escape from collaborative stagnation, finally realizing the accurate quantitative evaluation of the decision credibility of multi-agents in a distributed environment and topology-level crisis response.

[0069] Figure 4 is a flowchart for performing causal discrimination provided by the embodiments of the present application, and the process may include the following steps: In step S331, when the cognitive uncertainty residual is less than a first preset threshold and the local cognitive health degree is less than a second preset threshold, or when the cognitive uncertainty residual is greater than or equal to a third preset threshold, a non-global collaborative crisis state is determined.

[0070] Specifically, the non-global collaborative crisis state refers to a state where no topology-level systematic deadlock occurs currently and no global intervention by the large model is required. This state includes a "stable state" and a "local noise state".

[0071] According to Δu i and the joint state to perform causal discrimination: (1) Stable state (Class0): When Δu i <T1 (the first preset threshold, such as 0.06) and <T2 (the second preset threshold, such as 0.1), it indicates that both the current node and the neighbor nodes maintain low uncertainty, are in a stable collaboration stage, are determined to be in a non-global collaborative crisis state, and no external intervention is required.

[0072] (2) Local noise state (Class1): When Δu i≥T3 (the third preset threshold, e.g., 0.7), it indicates that the current agent has a significant cognitive anomaly, but surrounding nodes are normal. At this point, the fault is strictly confined within the single node, which is handled through a local security policy (such as rollback, retry or fuse mechanism), it is determined as a non-global collaborative crisis state, triggering of large model calls is explicitly prohibited, so as to avoid resource waste.

[0073] Step S333: when the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health degree is greater than or equal to the fourth preset threshold, a global collaborative crisis state is determined.

[0074] Specifically, when Δu i <T1 (the first preset threshold, e.g., 0.06) and ≥T4 (the fourth preset threshold, e.g., 0.9), it indicates that the current node and its neighboring nodes are all in a state of high cognitive confusion. At this point, the originally significantly large Δu i has converged to close to zero, while the overall keeps increasing. This indicates that the local cognitive anomaly has evolved from an isolated fault to cross-node collaborative instability, and cognitive blind spots have formed synchronous resonance in the topological network. Accordingly, it is determined as a global collaborative crisis state, and a large model intervention request is formally generated, authorizing the large model to intervene to break the collaborative deadlock.

[0075] This embodiment achieves accurate classification and graded response to collaborative crisis states by jointly determining the state causality of cognitive uncertainty residuals and local cognitive health. When the cognitive uncertainty residual is less than a first preset threshold and the local cognitive health is less than a second preset threshold, it is determined to be a stable state, indicating that the current node and its neighboring nodes maintain low uncertainty and are in a stable collaborative stage, requiring no external intervention. When the cognitive uncertainty residual is greater than or equal to a third preset threshold, it is determined to be a local noise state, indicating that the current agent has a significant cognitive abnormality but the surrounding nodes are normal. In this case, the fault is strictly limited to a single node, handled through a local security policy, and the triggering of large model calls is explicitly prohibited, effectively avoiding the waste of large model resources caused by isolated local faults. When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is greater than or equal to the fourth preset threshold, it is judged as a global collaborative crisis state. This indicates that the local cognitive anomaly has evolved from an isolated failure to cross-node collaborative instability, and the cognitive blind spot has formed a synchronous resonance in the topology network. Based on this, a large model intervention request is formally generated and the large model is authorized to intervene to break the collaborative deadlock. This mechanism achieves accurate differentiation between stable state, local noise state and global collaborative crisis state through multi-threshold joint discrimination. It not only prevents the large model from being falsely triggered by normal environmental fluctuations or local failures, but also ensures that the real topology-level systemic crisis can be identified in time and trigger macro-intervention. Thus, while ensuring the autonomy of computing network service scheduling, it achieves accurate allocation of intervention resources and a significant improvement in collaborative robustness.

[0076] Figure 5 The flowchart for step S5 provided in the embodiments of this application may include the following steps: Step S51: Extract the hidden states of the global collaborative crisis state from the graph neural network encoder of each agent, and determine the geometric statistics of the hidden states.

[0077] Specifically, the graph neural network encoder refers to the front-end module in the agent policy network responsible for feature communication and topology aggregation among multiple agents. Internally, it stacks graph attention layers with uncertainty penalty terms. The input is the local observations of the local agent (such as topology state, resource utilization, cognitive uncertainty values, etc.), and the output is a node-level hidden representation. The hidden state refers to the encoding result of the current multi-agent cooperative state in the latent feature space. In this embodiment, node-level hidden representation... After the readout functions are converged, the hidden state z of the graph level is obtained. crisis Its shape is N×d (N is the number of agents participating in the collaboration, and d is the feature dimension). Geometric statistics refer to the manifold structure feature data used to diagnose the current crisis state of the policy network. Specifically, it includes the variance σ of each dimension. 2The policy network (Actor network) includes gradient saturation rate, feature symmetry, and local Jacobian spectrum distribution.

[0078] Upon detecting a global collaborative crisis state, the hidden state z under the current crisis state is extracted from the graph neural network encoder. crisis Here, the original hidden state z is not directly expressed. crisis Instead of inputting a large model, it further computes the hidden state z. crisis Geometric statistics. For example, when multiple actions have highly symmetrical probability distributions (i.e., high feature symmetry), it often means falling into "multimodal equivalence," corresponding to saddle points or cooperative deadlock regions in reinforcement learning. These geometric statistics will serve as structured context, driving large models to generate accurate interferometric inference metrics.

[0079] Step S53: Construct structured cue words based on geometric statistics, and input the structured cue words into the large model to generate source code containing basic tensor operations.

[0080] Specifically, structured prompts are constructed, such as: "Given a tensor manifold exhibiting symmetric stagnation statistics [statistics dictionary], write a strictly differentiable PyTorch function f(z) using only basic tensor operations. This function must be injected with a minimal, bounded asymmetric perturbation to break gradient deadlock." The large model then generates executable source code S based on this. code (Using prompts to allow LLM to generate code is feasible and widely used.) Step S55: Compile the source code to obtain tensor operators.

[0081] Specifically, the source code S returned by the large model is processed using just-in-time compilation tools such as torch.jit.script. code Compile it to transform it into a differentiable tensor operator f. LLM ( ). This tensor operator applies continuous, smooth, and local geometric perturbations only to the latent feature manifold, thereby eliminating the hard switching and policy breakage problems caused by traditional text commands and achieving local tensor computation intervention with zero inference latency.

[0082] Step S57: Inject tensor operators into the forward propagation path of the policy network of each agent to obtain the corrected hidden state.

[0083] Specifically, the tensor operator f obtained from the compilation LLM ( ) Dynamically injected policy is used to form the corrected hidden state in the forward propagation path of the network: Where, Θcode This represents the fixed numerical parameters (such as bias vectors, perturbation coefficients, etc.) generated by the large model. The corrected hidden state z. modified Input to the original policy network The final action is output autonomously by the intelligent agent, where 'a' represents the action.

[0084] This embodiment, upon detecting a global collaborative crisis state, extracts the hidden states from the graph neural network encoders of each agent and calculates their geometric statistics, enabling accurate diagnosis of the crisis type currently trapped in the policy network. Based on these geometric statistics, a structured prompt word input model is constructed, allowing the model to generate executable source code containing basic tensor operations, achieving a direct conversion from natural language instructions to executable intervention logic. Furthermore, the source code is compiled using a just-in-time (JIT) compiler to obtain a differentiable tensor operator. This tensor operator applies continuous, smooth, and local geometric perturbations only to the latent feature manifold, eliminating the traditional... The hard switching and policy breakage caused by traditional text commands were addressed by implementing local tensor computation intervention with zero inference latency. Finally, tensor operators were dynamically injected into the forward propagation path of the policy network to obtain the corrected hidden state. The agent autonomously outputs the final action based on the corrected hidden state. This approach not only preserves the global perspective and semantic understanding of the large model, but also ensures the differentiability and policy continuity of the intervention process. This allows the agent to smoothly escape cooperative deadlock without interrupting operation, realizing the evolution from externally heuristic dependence to endogenous escape capability. This significantly improves the intervention accuracy and system robustness of the computing network service scheduling in crisis situations.

[0085] Figure 6 The flowchart for step S7 provided in the embodiments of this application may include the following steps: Step S71: During the interaction process of each agent outputting escape actions, determine the gradient of the policy loss function of each agent, and backpropagate the gradient back to the policy network of each agent via tensor operators to update the policy parameters of the policy network.

[0086] Specifically, due to the injected tensor operator f LLM ( ) is strictly differentiable and can be natively integrated with underlying reinforcement learning optimization frameworks. In the continuous interaction phase after tensor operator injection, the policy gradient can be backpropagated via this implicit operator. According to the chain rule, the policy loss function L policy The gradient formula for the policy parameter θ of the policy network is: in, The partial derivative of the policy loss function with respect to the corrected hidden state (upstream gradient); θ represents the partial derivative of the original hidden state with respect to the policy parameter θ, reflecting the influence of the policy parameter on the hidden state; I is the identity matrix, representing the original gradient path. For tensor operator f LLM The Jacobian matrix of ( ) (i.e. This reflects the local linear transformation and perturbation intensity of the tensor operator on the input. This indicates the linear modulation of the gradient as it passes through large model operators.

[0087] Under this mechanism, the positive temporal difference (TD) error constitutes the gradient verification of the effectiveness of the escape direction. The agent can synchronously update the local policy parameters within the T-step interaction window of crisis intervention without relying on offline experience replay and full parameter update.

[0088] Step S73: Based on the policy network of each agent after parameter update, redetermine the cognitive uncertainty residual.

[0089] Specifically, as the policy parameters θ are updated online, the agent gradually internalizes the escape logic. The new action probability distribution output by the policy network after the parameter update is acquired in real time, and the cognitive uncertainty residual is recalculated based on this new distribution. This residual value will serve as a quantitative indicator for subsequent judgments on whether the agent has truly escaped the crisis state.

[0090] Step S75: Based on the redefined cognitive uncertainty residual, perform causal discrimination. If the classification result of the cooperative state is determined to be restored to a non-global cooperative crisis state, remove the tensor operator from the forward propagation path of the policy network.

[0091] Specifically, to ensure long-term policy optimality, for the tensor operator f LLM ( ) is equipped with an automatic pruning trigger. Causal discrimination is performed on the redefined cognitive uncertainty residuals. Once a non-global collaborative crisis state is determined, indicating that external heuristics are no longer needed, the tensor operator f is pruned via the automatic pruning trigger. LLM ( ) is removed from the forward propagation path of the policy network.

[0092] Step S77: Perform local decisions based on the updated policy network of each agent to generate the target computing network service.

[0093] Specifically, after the tensor operators are successfully pruned and removed, each agent relies entirely on the updated policy network (Actor network) for independent forward propagation and local decision-making. Since the policy network has internalized the escape logic and fixed the topological defects of the policy manifold, it can continuously and stably generate the target computing network service with zero external inference latency and no policy breakage.

[0094] In this embodiment, during the interaction process of each agent outputting escape actions, the gradient of the policy loss function is determined and backpropagated to the policy network via a strictly differentiable tensor operator to update the policy parameters, thus realizing an online learning paradigm that unifies inference trajectory and parameter adaptation. Based on this, the cognitive uncertainty residual is redefined based on the updated policy network. As the policy parameters are updated online, the agents gradually internalize the escape logic, and this cognitive uncertainty residual serves as a quantitative indicator for determining whether the agent has truly escaped the crisis state. Furthermore, causal discrimination is performed based on the redefined cognitive uncertainty residual to determine the cooperative state. When the class result reverts to a non-global collaborative crisis state, the tensor operator is removed from the forward propagation path of the policy network through an automatic pruning trigger, ensuring long-term policy optimality and avoiding continuous dependence on external heuristics. Ultimately, each agent relies entirely on the policy network after parameter updates for independent forward propagation and local decision-making. Since the policy network has internalized the escape logic and repaired topological defects, it can continuously and stably generate target computing network services with zero external inference latency and no policy breakage. This completes the evolution from external heuristic dependence to endogenous escape capability, realizing autonomous closed-loop and robust operation of multi-agent collaborative scheduling.

[0095] In a preferred embodiment, the method further includes: determining the gradient of the policy loss function of each agent, and backpropagating the gradient back to the policy network of each agent via tensor operators to update the policy parameters of the policy network; monitoring the policy entropy and value loss of the updated policy network of each agent in real time, and removing the tensor operators from the forward propagation path of the policy network when the policy entropy recovers to above a preset threshold and the value loss converges.

[0096] Specifically, policy entropy is an indicator that measures the degree of uncertainty in the output distribution of a policy network, and it is calculated based on the action output distribution of the policy network. The formula for calculating policy entropy can be: Where A is the discrete action space; For the policy network in state o i The probability of choosing action 'a'.

[0097] During the crisis injection phase, due to the tensor operator f LLMWhen an asymmetric perturbation is applied to the hidden state, the output distribution of the policy network changes, and the entropy fluctuates accordingly. As the policy parameter θ is continuously updated through gradient backpropagation, the agent gradually internalizes the escape logic, and the policy entropy will show a downward trend. When the entropy is high, it indicates that the output distribution of the policy network is relatively dispersed and the decision is uncertain; when the entropy falls back to above the preset threshold (or stabilizes within a reasonable range), it indicates that the decision becomes certain again and has escaped multimodal equivalence stagnation. Value loss refers to the training loss during the optimization of the policy function. Convergence of value loss indicates that the policy network's evaluation of the current state has become stable.

[0098] Specifically, the determination of recovery relies on a multi-dimensional recovery condition matrix, which generally needs to satisfy several custom conditions, such as entropy reduction, reward recovery, achievement of consensus, mitigation of gradient saturation, and crisis duration constraints. Once it is determined that the entropy of the local policy has recovered and the value loss has converged, this indicates that the policy parameter θ has learned to independently generate the correct escape trajectory and no longer requires external inspiration. At this point, the tensor operator f is... LLM ( ) is removed from the forward propagation path of the policy network, thereby realizing the closed-loop self-evolution process of "injection-learning-pruning", enabling the agent to gradually evolve from external inspiration dependence to endogenous escape capability.

[0099] This embodiment achieves an online learning paradigm that unifies inference trajectory and parameter adaptation by determining the gradient of the policy loss function of each agent and backpropagating the gradient to the policy network via a strictly differentiable tensor operator to update the policy parameters. This allows agents to synchronously update local policy parameters within the crisis intervention interaction window, gradually internalizing escape logic. Based on this, a multidimensional recovery judgment condition matrix is ​​constructed by real-time monitoring of the policy entropy and value loss of the updated policy network. The policy entropy is used to measure the uncertainty of the policy network's output distribution. When the entropy falls from a high level to above a preset threshold, it indicates that the policy network has escaped multimodal equivalence stagnation and the decision has become deterministic again, while the value loss... Convergence indicates that the policy network's evaluation of the current state has stabilized, and both convergence and convergence serve as quantitative indicators for determining whether an agent has truly escaped the crisis state. When the policy entropy value recovers to above the preset threshold and the value loss converges, the tensor operator is removed from the forward propagation path of the policy network, realizing a closed-loop self-evolutionary process of "injection-learning-pruning". This allows the agent to gradually evolve from external heuristic dependence to endogenous escape capability, thereby ensuring the accuracy of intervention while avoiding continuous dependence on external large models. Ultimately, this enables each agent to continuously and stably generate target computing network services in a manner with zero external inference latency and no policy breakage, significantly improving the autonomous closed-loop capability and long-term robustness of multi-agent collaborative scheduling.

[0100] The specific implementation of the present invention will now be described with reference to the schematic diagram. See also Figures 7-8The core process of this solution is divided into five stages, forming a closed loop from task planning to continuous optimization: Step 1001: Task Division and Model Design: First, the complex computing network service tasks (i.e., computing network service requests) are decoupled into multiple collaborative stages (such as intent parsing, node selection, service deployment, etc.), and a dedicated agent model (intelligent agent) is designed for each collaborative stage. These agent models share a dedicated toolkit (standard API interface).

[0101] Step 1002: Decoupling Cognitive Uncertainty from Two Perspectives: During operation, the state of the agent model is monitored in real time. The "random uncertainty measure" (i.e., environmental noise) is evaluated by calculating the action distribution entropy, and the "cognitive uncertainty measure" (i.e., model knowledge blind spot) is evaluated by using prediction variance or MCDropout.

[0102] Step 1003: Uncertainty-Aware Graph Routing: Each agent node encapsulates its local cognitive health into a semantic control packet and propagates it throughout the network topology. By calculating the "cognitive uncertainty residual," it can distinguish between local anomalies and global crises. If it is a non-global collaborative crisis state, local processing continues; if it is a global collaborative crisis state, large-scale model intervention is triggered.

[0103] Step 1004: Prompt-driven implicit operator injection: Once intervention is triggered, the current hidden state (such as GNN encoded output) is extracted, and a structured Prompt input is constructed for the large model. The large model generates a differentiable code, which is compiled into a tensor operator and dynamically injected into the surrogate model to modify the original policy network, thereby breaking the deadlock.

[0104] Step 1005: Online Optimization and Automatic Operator Pruning: The injected tensor operator is strictly differentiable, so environmental feedback (rewards, TD errors) can update the agent model's parameters via backpropagation through this tensor operator (online implicit knowledge distillation). Simultaneously, the policy entropy and value loss of this operator are monitored. If the operator is no longer necessary or causes performance degradation, it is automatically pruned and removed, restoring the agent model's independent reasoning ability. Alternatively, causal discrimination is performed using the redefined cognitive uncertainty residuals. If the cooperative state classification result reverts to a non-global cooperative crisis state, the tensor operator is removed from the policy network's forward propagation path.

[0105] Accordingly, please refer to Figure 9 A block diagram of a multi-agent computing network service generation device provided in this application embodiment, the device comprising: The local decision evaluation unit 101 is used to map the acquired computing network service request to multiple preset collaborative stages, and call the corresponding intelligent agents of each stage to perform local decisions in order to determine the local cognitive health of each intelligent agent. The cooperative state discrimination unit 103 is used to determine the neighbor set of each agent and determine the cognitive uncertainty residual of each agent based on the local cognitive health and the neighbor set, so as to output the cooperative state classification result based on the causal discrimination of the cognitive uncertainty residual. The large model correction unit 105 is used to correct the hidden states of each agent that represent the cooperative characteristics in response to the classification result of the cooperative state of any agent being a global cooperative crisis state. This is done by using the large model to correct the hidden states of each agent to avoid cooperative risks, so that each agent can output an escape action based on the corrected hidden state. The state recovery and service generation unit 107 is used to generate target computing network services based on the local decisions executed by each agent during the interaction process of each agent outputting escape actions, based on the re-determined cognitive uncertainty residual, and when the cooperative state classification result is determined to be restored to a non-global cooperative crisis state.

[0106] In some optional implementations, the local decision evaluation unit 101 includes: During the execution of local decisions, the action probability distribution of each agent in the current observation state is obtained, and the random uncertainty measure of each agent is determined based on the maximum probability value in the action probability distribution. A random mask is applied to the hidden layer of the policy network of each agent, and a preset number of forward propagations are performed on the current observation state. The cognitive uncertainty measure of each agent is determined based on the degree of difference between the perturbation policy distribution output by each forward propagation and the average policy distribution of all forward propagations. By product coupling the measures of accidental uncertainty and cognitive uncertainty, a joint cognitive crisis index is obtained. The joint cognitive crisis index is normalized to obtain the local cognitive health of each agent.

[0107] In some optional implementations, the cooperative state determination unit 103 includes: For any current agent's neighbor set, obtain the semantic control packet broadcast by each neighbor agent in the neighbor set. The semantic control packet contains at least the local cognitive health of each neighbor agent. The average cognitive health of the neighborhood is determined based on the average local cognitive health of all neighbor agents in the neighborhood set. Based on the difference between the current agent's local cognitive health and the neighborhood average cognitive health, the cognitive uncertainty residual corresponding to the current agent is determined.

[0108] In some optional implementations, the cooperative state determination unit 103 includes: When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is less than the second preset threshold, or when the cognitive uncertainty residual is greater than or equal to the third preset threshold, it is determined to be a non-global collaborative crisis state. When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is greater than or equal to the fourth preset threshold, it is determined to be a global collaborative crisis state.

[0109] In some alternative implementations, the large model correction unit 105 includes: Extract the hidden states of the global cooperative crisis state from the graph neural network encoder of each agent, and determine the geometric statistics of the hidden states. Structured cue words are constructed based on geometric statistics. These structured cue words are then input into a large model to generate source code containing basic tensor operations. The source code is compiled to obtain tensor operators; Tensor operators are injected into the forward propagation path of the policy network of each agent to obtain the corrected hidden state.

[0110] In some optional implementations, the state recovery and service generation unit 107 includes: During the interaction process of each agent outputting escape actions, the gradient of the policy loss function of each agent is determined, and the gradient is backpropagated to the policy network of each agent via tensor operators through backpropagation to update the policy parameters of the policy network. Based on the updated policy network of each agent, the cognitive uncertainty residual is redefined. Causal discrimination is performed based on the redefined cognitive uncertainty residuals. If the classification result of the cooperative state is determined to be restored to a non-global cooperative crisis state, the tensor operator is removed from the forward propagation path of the policy network. Local decisions are made based on the updated policy network of each agent to generate the target computing network service.

[0111] In some alternative embodiments, the apparatus is also used for: The gradient of the policy loss function of each agent is determined, and the gradient is backpropagated to the policy network of each agent via tensor operators through backpropagation to update the policy parameters of the policy network. The policy entropy and value loss of each agent's policy network are monitored in real time. When the policy entropy recovers to above the preset threshold and the value loss converges, the tensor operator is removed from the forward propagation path of the policy network.

[0112] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0113] In this embodiment, a multi-agent computing network service generation device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0114] Please see Figure 10 , Figure 10 This application provides a schematic diagram of the structure of a computer device, as shown in the embodiment of the present application. Figure 10 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 10 Take a processor 10 as an example.

[0115] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0116] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0117] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0118] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0119] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0120] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0121] The apparatus and units described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0122] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0123] Those skilled in the art will understand that the embodiments of this application can be provided as methods or apparatus. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, and devices according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0128] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0129] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0130] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for generating multi-agent computing network services, characterized in that, The method includes: The acquired computing network service requests are mapped to multiple preset collaborative stages, and the corresponding intelligent agents in each stage are invoked to perform local decisions in order to determine the local cognitive health of each intelligent agent. The neighbor set of each agent is determined, and the cognitive uncertainty residual of each agent is determined based on the local cognitive health and the neighbor set. The causal discrimination is then performed based on the cognitive uncertainty residual to output the cooperative state classification result. In response to the classification result of the cooperative state of any agent being a global cooperative crisis state, the hidden state of each agent used to characterize the cooperative features is modified using a large model, so that each agent outputs an escape action to avoid cooperative risks based on the modified hidden state. During the interaction process in which each of the intelligent agents outputs the escape action, based on the redefined cognitive uncertainty residual, and if it is determined that the cooperative state classification result has recovered to a non-global cooperative crisis state, a target computing network service is generated based on the local decisions executed by each of the intelligent agents.

2. The method according to claim 1, characterized in that, Determining the local cognitive health of each of the aforementioned agents includes: During the execution of the local decision, the action probability distribution of each agent in the current observation state is obtained, and the random uncertainty measure of each agent is determined based on the maximum probability value in the action probability distribution. A random mask is applied to the hidden layer of the policy network of each agent, a preset number of forward propagations are performed on the current observation state, and the cognitive uncertainty measure of each agent is determined based on the degree of difference between the perturbation policy distribution output by each forward propagation and the average policy distribution of all forward propagations. By product coupling the accidental uncertainty measure and the cognitive uncertainty measure, a joint cognitive crisis index is obtained. The joint cognitive crisis index is normalized to obtain the local cognitive health of each agent.

3. The method according to claim 1, characterized in that, The step of determining the cognitive uncertainty residuals of each agent based on the local cognitive health and the neighbor set includes: For any current agent's neighbor set, obtain the semantic control packet broadcast by each neighbor agent in the neighbor set, wherein the semantic control packet contains at least the local cognitive health of each neighbor agent; The average cognitive health of the neighborhood is determined based on the average local cognitive health of all neighbor agents in the neighbor set. Based on the difference between the local cognitive health of the current agent and the average cognitive health of the neighborhood, the cognitive uncertainty residual corresponding to the current agent is determined.

4. The method according to claim 1, characterized in that, The causal discrimination output of the cooperative state classification result based on the cognitive uncertainty residual includes: When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is less than the second preset threshold, or when the cognitive uncertainty residual is greater than or equal to the third preset threshold, it is determined to be a non-global collaborative crisis state. When the cognitive uncertainty residual is less than the first preset threshold and the local cognitive health is greater than or equal to the fourth preset threshold, it is determined to be a global collaborative crisis state.

5. The method according to claim 1, characterized in that, The step of modifying the hidden states of each agent used to represent cooperative features using a large model includes: Extract the hidden states of the global cooperative crisis state from the graph neural network encoder of each of the agents, and determine the geometric statistics of the hidden states; Based on the geometric statistics, structured cue words are constructed, and the structured cue words are input into the large model to generate source code containing basic tensor operations; The source code is compiled to obtain tensor operators; The tensor operator is injected into the forward propagation path of the policy network of each agent to obtain the corrected hidden state.

6. The method according to claim 5, characterized in that, During the interaction process in which each of the intelligent agents outputs the escape action, based on the redefined cognitive uncertainty residual, and in the case that the cooperative state classification result is determined to have recovered to a non-global cooperative crisis state, a target computing network service is generated based on the local decisions executed by each of the intelligent agents, including: During the interaction process in which each agent outputs the escape action, the gradient of the policy loss function of each agent is determined, and the gradient is backpropagated to the policy network of each agent via the tensor operator through backpropagation to update the policy parameters of the policy network. Based on the updated policy networks of each agent, the cognitive uncertainty residuals are redefined. Causal discrimination is performed based on the redefined cognitive uncertainty residuals. If it is determined that the cooperative state classification result has been restored to a non-global cooperative crisis state, the tensor operator is removed from the forward propagation path of the policy network. Local decisions are made based on the updated policy network of each agent to generate the target computing network service.

7. The method according to claim 5, characterized in that, During the interaction process in which each of the intelligent agents outputs the escape action, the method further includes: The gradient of the policy loss function of each agent is determined, and the gradient is backpropagated to the policy network of each agent via the tensor operator through backpropagation to update the policy parameters of the policy network. The policy entropy and value loss of each agent's policy network are monitored in real time. When the policy entropy recovers to above a preset threshold and the value loss converges, the tensor operator is removed from the forward propagation path of the policy network.

8. A multi-agent computing network service generation device, characterized in that, The device includes: The local decision evaluation unit is used to map the acquired computing network service request to multiple preset collaborative stages, and call the corresponding intelligent agents of each stage to perform local decisions in order to determine the local cognitive health of each intelligent agent. The cooperative state discrimination unit is used to determine the neighbor set of each agent, and to determine the cognitive uncertainty residual of each agent based on the local cognitive health and the neighbor set, so as to perform causal discrimination based on the cognitive uncertainty residual and output the cooperative state classification result. The large model correction unit is used to correct the hidden states of each agent that represent the cooperative features in response to the classification result of the cooperative state of any agent being a global cooperative crisis state. This is done by using the large model to correct the hidden states of each agent, so that each agent outputs an escape action to avoid cooperative risks based on the corrected hidden states. The state recovery and service generation unit is used to generate target computing network services based on the local decisions executed by each of the intelligent agents during the interaction process in which each of the intelligent agents outputs the escape action, based on the re-determined cognitive uncertainty residual, and in the case that the cooperative state classification result is determined to be restored to a non-global cooperative crisis state.

9. A computer device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the multi-agent computing network service generation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the multi-agent network service generation method according to any one of claims 1 to 7.

Images