AI native networking multi-agent spu collaborative circulation system

Abstract

The application discloses an AI native e-commerce networking multi-agent SPU collaborative circulation system and belongs to the technical field of AI native e-commerce networking multi-agent collaboration. The system comprises a knowledge core layer, a collaborative network layer and an autonomous evolution layer. The knowledge core layer constructs an SPU standardized knowledge expression unit and performs semantic topology mapping to generate a globally consistent SPU semantic topology graph. The collaborative network layer calls the topology graph to arrange point-to-point interaction tasks for multiple peer agents to generate a centerless collaborative flow for SPU circulation tasks. The autonomous evolution layer performs track tracking on the circulation tasks in the collaborative flow to generate a task track tensor and adaptively adjusts the structural parameters of the knowledge core layer and the collaborative network layer according to the tensor. The system solves the problems that the centralized bottleneck and the static structure are difficult to adapt to dynamic circulation through decentralized topology-driven collaboration and autonomous evolution based on task tracks.

Description

Technical Field

[0001] This invention relates to the field of AI-native business network multi-agent collaborative technology, specifically to an AI-native business network multi-agent SPU collaborative circulation system. Background Technology

[0002] In commodity circulation and transaction scenarios, the multi-agent SPU collaborative circulation system of the Internet of Things (IoT) requires multiple independent agents to collaborate in discovering, negotiating, trading, and fulfilling standard product units. Existing IoT collaborative solutions generally adopt a centralized scheduling architecture, with a central server or orchestration node uniformly managing task allocation and message routing among agents. This centralized architecture faces single-point performance bottlenecks and insufficient elasticity. When the concurrency of circulation tasks increases, message congestion at the central node slows down overall collaboration efficiency, and a failure of the central node directly leads to collaboration interruption. Simultaneously, each agent represents commodity objects differently, lacking a consistent semantic model. Information such as commodity attributes and transaction conditions exchanged between agents frequently leads to misunderstandings due to semantic conflicts, affecting the success rate of collaboration. Some solutions introduce fixed knowledge graphs to attempt to unify semantics, but such graph structures are statically preset and cannot adjust automatically with changes in circulation patterns. Existing technologies have significant shortcomings in both decentralized collaboration and adaptive evolution. Regarding decentralized collaboration, while existing distributed peer-to-peer networks can eliminate central nodes, they lack topology guidance mechanisms oriented towards product semantics. Interactions between agents are mostly based on simple broadcasting or random exploration, and task transmission lacks directionality, easily introducing a large amount of ineffective communication. In terms of adaptive evolution, information such as actual task flow paths, frequencies, and congestion levels generated during system operation is not systematically collected and utilized. Knowledge structures and collaborative relationships cannot be optimized in a closed-loop manner based on online circulation performance, making it difficult for the system to cope with actual changes such as changes in product types and the dynamic addition and withdrawal of agents. To solve these problems, a new collaborative circulation mechanism is needed that can provide semantically driven interaction guidance for agents under decentralized conditions and can also provide feedback on the system structure for self-adjustment based on task execution trajectories. Summary of the Invention

[0003] This invention provides an AI-native business network multi-agent SPU collaborative circulation system, which aims to eliminate the bottleneck of centralized collaboration and the limitations of static knowledge structure, realize multi-point peer collaboration driven by semantic topology in a decentralized mode, and drive the continuous evolution of knowledge representation and collaborative relationship based on tensor analysis of circulation task trajectory data.

[0004] To achieve the above objectives, this invention provides the following technical solution: This invention provides an AI-native e-commerce network multi-agent SPU collaborative circulation system, comprising a knowledge core layer, a collaborative network layer, and an autonomous evolution layer. The knowledge core layer is used to construct standardized SPU knowledge representation units and perform semantic topology mapping processing on these units to generate a globally consistent SPU semantic topology graph. The collaborative network layer is used to invoke the globally consistent SPU semantic topology graph to orchestrate point-to-point interactive tasks among multiple peer agents, generating a decentralized collaborative flow for SPU circulation tasks. The autonomous evolution layer is used to perform task trajectory tracking on the circulation tasks in the decentralized collaborative flow, generating task trajectory tensors, and adaptively adjusting the structural parameters of the knowledge core layer and the collaborative network layer based on these task trajectory tensors. Through the closed-loop collaboration of this three-layer architecture, consistent expression of SPU knowledge, autonomous negotiation among multiple agents, and continuous system evolution can be achieved in a decentralized environment, significantly improving the efficiency of commodity circulation and task collaboration capabilities in the e-commerce network.

[0005] As a technical solution of this invention, the knowledge core layer performs semantic topology mapping processing on the SPU standardized knowledge representation units to generate a globally consistent SPU semantic topology graph. Specifically, it extracts the attribute dimension and relation dimension from the SPU standardized knowledge representation units to construct a set of attribute-relation pairs. It then performs topology embedding processing on the attribute-relation pair set to generate SPU semantic nodes and SPU semantic edges. Finally, it connects multiple SPU semantic nodes according to the SPU semantic edges to generate a globally consistent SPU semantic topology graph. This solution transforms discrete SPU attributes and relations into a global semantic graph with a topological structure, providing a unified and scalable knowledge interaction foundation for multi-agent systems.

[0006] Preferably, the process by which the collaborative network layer invokes a globally consistent SPU semantic topology graph to orchestrate point-to-point interaction tasks among multiple peer agents, generating a decentralized collaborative flow for SPU circulation tasks, includes: deploying a local view copy of the globally consistent SPU semantic topology graph in each peer agent; calculating the negotiation offset of SPU circulation tasks between adjacent peer agents based on the local view copy; and adjusting the interaction sequence of adjacent peer agents based on the negotiation offset to generate a decentralized collaborative flow. Agents do not need to rely on a central node; they can form a globally coordinated circulation task flow solely through local views and local negotiation, improving the system's robustness and response speed.

[0007] Furthermore, the autonomous evolution layer performs task trajectory tracking for tasks flowing through the decentralized collaborative flow. When generating the task trajectory tensor, it binds a unique task identifier to each SPU flowing task. In the decentralized collaborative flow, each peer agent traversed by the SPU flowing task is timestamped according to the unique task identifier, generating a task flow path. Multiple task flow paths are stacked according to the time dimension to generate a task trajectory tensor. This trajectory tensor fully records the spatiotemporal distribution of tasks in the agent network, providing a high-fidelity decision basis for subsequent structural optimization.

[0008] An optimal approach for the autonomous evolution layer to adaptively adjust the structural parameters of the knowledge core layer and the collaborative network layer based on the task trajectory tensor is as follows: Perform sparsity analysis on the task trajectory tensor to generate a sparsity distribution graph; perform graph matching between the sparsity distribution graph and the globally consistent SPU semantic topology graph to determine the subgraph to be optimized; and adjust the SPU semantic nodes and SPU semantic edges in the globally consistent SPU semantic topology graph based on the subgraph to be optimized. In this way, the system can automatically identify inefficient or redundant connections in the knowledge graph and autonomously complete the evolution of the knowledge structure, ensuring that the SPU semantic topology graph continuously aligns with the actual flow pattern.

[0009] As another technical solution of the present invention, the cooperative network layer also includes an interaction context fusion module, which is used to fuse the SPU circulation context information exchanged between adjacent peer agents in a decentralized cooperative flow, generate a fused context feature vector, and inject the fused context feature vector into the decentralized cooperative flow. Through context fusion, the information exchange between agents is no longer limited to discrete instructions, but contains rich semantic features, thereby improving the accuracy of cooperative decision-making.

[0010] Specifically, during the fusion process, the interaction context fusion module extracts the SPU attribute metadata and transaction parameter metadata exchanged between neighboring peer agents; it performs spatial topology alignment processing on the SPU attribute metadata to obtain a spatial alignment vector; it performs transaction intent weighting processing on the transaction parameter metadata to obtain a weighted intent vector; and it concatenates the spatial alignment vector and the weighted intent vector to generate a fusion context feature vector. This method jointly encodes product attributes and transaction intent, ensuring that the fused features maintain both spatial topology consistency and reflect dynamic transaction tendencies.

[0011] In one embodiment, the knowledge core layer further includes a semantic consistency arbitration module for detecting version conflicts in the globally consistent SPU semantic topology graph. When a version conflict is detected, an arbitration request is generated and broadcast to multiple peer agents in the cooperative network layer. The module receives version voting results from multiple peer agents and determines the authoritative version topology graph based on these results. In a decentralized environment, multiple agents may modify the knowledge graph in parallel, leading to version conflicts. This arbitration mechanism ensures the consistency of global knowledge.

[0012] Preferably, when the semantic consistency arbitration module determines the authoritative version topology graph based on the version voting results, it counts the number of supporting votes obtained by each candidate version topology graph in the version voting results; the candidate version topology graph with the most supporting votes is selected as the authoritative version topology graph; and the authoritative version topology graph is called to overwrite and update the globally consistent SPU semantic topology graph. Through the majority voting mechanism, the system quickly converges to the authoritative version in a low-overhead manner, ensuring the global reliability and timeliness of knowledge representation.

[0013] The collaborative network layer further includes a collaborative result attribution module, used to backtrack the execution path of the final result of the SPU circulation task in the decentralized collaborative flow, generating an execution path backtracking chain. Based on the execution path backtracking chain, the contribution weight of each participating peer agent is calculated, and the task rewards in the decentralized collaborative flow are distributed according to the contribution weight. Based on traceable execution paths and contribution quantification, a fair and transparent multi-agent incentive mechanism can be achieved, promoting active participation of agents in collaborative circulation tasks.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A globally consistent SPU semantic topology graph is generated through the knowledge core layer, and a local view copy of this topology graph is deployed for each peer agent in the collaborative network layer. Neighboring agents calculate the negotiation offset of SPU circulation tasks based on the local view, and then adjust the interaction sequence to generate a decentralized collaborative flow. This scheme replaces the central coordinator with a semantic topology graph, allowing agents to autonomously determine the transmission direction and negotiation object of circulation tasks locally based on the guidance of product semantic nodes and associated edges in the graph, eliminating the performance bottleneck and single point of failure risk caused by the central node. The point-to-point interaction between agents is based on a consistent semantic topology, and the communication link is aligned with the product circulation path, which can significantly reduce the invalid communication overhead caused by blind broadcasting. The circulation tasks in the decentralized collaborative flow are tracked through an autonomous evolution layer. The task flow path is generated using a unique task identifier and timestamp, and multiple paths are stacked according to the time dimension to form a task trajectory tensor. The sparsity distribution map of the task trajectory tensor is obtained by performing sparsity analysis on the task trajectory tensor. This map is then matched with the globally consistent SPU semantic topology graph to locate the subgraph of the topology to be optimized, and the SPU semantic nodes and semantic edges are added, deleted, and adjusted accordingly. This closed-loop evolution mechanism transforms implicit knowledge such as task frequency and path patterns accumulated in actual circulation into explicit topology adjustment instructions, enabling the semantic expression of the knowledge core layer and the connection relationship of the collaborative network layer to be automatically reconstructed as the business network environment changes dynamically, maintaining the circulation efficiency and response flexibility of the system in high-concurrency and dynamic scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a schematic diagram of the structure of an AI-native e-commerce network multi-agent SPU collaborative circulation system; Figure 2 This is the processing flowchart of the interaction context fusion module; Figure 3 This is the curve showing the impact of the connection threshold on edge connection precision, recall, and F1 score. Figure 4 This is the curve showing the impact of the negotiation tolerance threshold on the global task completion time and the average number of negotiation rounds; Figure 5 It is the curve showing the impact of the fusion context feature injection coefficient on the average task achievement rate and the average negotiation offset. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 This invention provides an AI-native business network multi-agent SPU collaborative circulation system, comprising a knowledge core layer, a collaborative network layer, and an autonomous evolution layer. The knowledge core layer constructs standardized knowledge representation units for SPUs and performs semantic topology mapping on these units to generate a globally consistent SPU semantic topology graph. The collaborative network layer invokes the globally consistent SPU semantic topology graph to orchestrate point-to-point interactive tasks among multiple peer agents, generating a decentralized collaborative flow for SPU circulation tasks. The autonomous evolution layer tracks the circulation tasks within the decentralized collaborative flow, generates task trajectory tensors, and adaptively adjusts the structural parameters of the knowledge core layer and the collaborative network layer based on these task trajectory tensors.

[0019] Example 1:

[0020] In practical implementation, the knowledge core layer performs semantic topology mapping processing on the SPU standardized knowledge representation units to generate a globally consistent SPU semantic topology graph. Each SPU standardized knowledge representation unit consists of a pre-defined set of SPU attributes and a set of SPU relationships. Each SPU attribute in the SPU attribute set is defined as an attribute dimension, and each SPU relationship in the SPU relationship set is defined as a relationship dimension. The knowledge core layer extracts the attribute dimensions and relationship dimensions from the SPU standardized knowledge representation units to construct a set of attribute-relation pairs. An attribute-relation pair consists of an attribute dimension vector and a relationship dimension vector. The attribute dimension vector is represented as... ,in This is the index number of the attribute dimension. The value range is from 1 to M, where M is the total number of attribute dimensions. The relation dimension vector is represented as... ,in, This is the index number for the relation dimension. The value range is from 1 to N, where N is the total number of relation dimensions. For any attribute dimension vector... And a relational dimension vector that has a semantic connection with it. The core knowledge layer will and Perform vector concatenation to obtain attribute-relation pairs. ,symbol This represents a vector concatenation operation. All attribute-relation pairs constitute the attribute-relation pair set.

[0021] The knowledge core layer performs topological embedding processing on the attribute-relation pair set to generate SPU semantic nodes and SPU semantic edges. The topological embedding processing is implemented using a semantic topological embedding model. This model includes an input layer, an attention encoding layer, and a topological decoding layer. The input layer receives the attribute-relation pair set. The attention encoding layer processes each attribute-relation pair... Perform multi-head self-attention encoding to generate the corresponding embedding representation. The attention encoding layer contains H attention heads, where H is set to 8. The output of the h-th attention head is... ,in, , , Based on attribute-relation pairs The generated query matrix, key matrix, and value matrix The dimension of the key vector is set to 64. The attention encoding layer concatenates the outputs of the H attention heads and then performs a linear transformation to obtain the embedding representation. .

[0022] The topology decoding layer is based on all embedded representations Generate SPU semantic nodes and SPU semantic edges. The topology decoding layer contains a node generation sublayer and an edge generation sublayer. The node generation sublayer generates each attribute dimension vector. Aggregate all containing The embedding representation corresponding to the attribute-relation tuple is calculated using the following formula for the SPU semantic node. :

[0023] in, Representation and attribute dimension vector A set of indices for relational dimension vectors that have semantic relationships. for The number of elements in the middle. Embed weight matrices for trainable nodes. Embed bias vectors for trainable nodes. The Tanh activation function is used. The weight matrix is... The dimension is , For embedded representation Dimensions This is the final dimension of the SPU semantic node. The value is set to 128. The edge generation sublayer pairs SPU semantic nodes that have semantic association. and Based on the relationship dimension vector of co-occurrence of the two Calculate SPU semantic edges SPU semantic edge The calculation method is as follows: ,in, For a trainable bilinear transformation matrix, The dimension is , Relational dimension vector The semantic topology embedding model employs a negative sampling contrastive loss function for parameter optimization during training. The training data consists of attribute-relation pairs extracted from historical SPU circulation records. Positive samples are attribute-relation pairs co-occurring in actual circulation, while negative samples are generated by randomly replacing the relation dimension vector. The training process uses the Adam optimizer with a learning rate of 0.001, a batch size of 256, and 200 training iterations.

[0024] The knowledge core layer connects multiple SPU semantic nodes according to the SPU semantic edges to generate the globally consistent SPU semantic topology graph. Specifically, for any SPU semantic edge output by the edge generation sublayer... ,like The value is greater than the preset connection threshold. Then the core knowledge layer is in the SPU semantic node. and Establish an undirected edge between them, and... The value is used as the weight of the undirected edge. Connection threshold. The value is set to 0.5, based on the threshold that maximizes the F1 score of edge connections on the validation set of the semantic topology embedding model. All undirected edges that meet the condition, along with all SPU semantic nodes, constitute a globally consistent SPU semantic topology graph. The globally consistent SPU semantic topology graph is stored in the knowledge core layer as an adjacency matrix for use by the collaborative network layer.

[0025] See Figure 3 In this diagram, corresponding to Example 1, when the knowledge core layer performs topological embedding processing on the attribute-relationship binary set to generate a globally consistent SPU semantic topology graph, the threshold for edge connection in the topology decoding layer is... The curve showing the relationship between edge connection precision, edge connection recall, and edge connection F1 score. The horizontal axis represents the connection threshold. The values ​​range from 0 to 1, and the vertical axis represents the scores of each indicator, also ranging from 0 to 1. The curves consist of three lines: the edge-connected precision curve (solid line, blue), the edge-connected recall curve (dashed line, orange), and the edge-connected F1 score curve (dotted line, green).

[0026] With connection threshold As the threshold gradually increases from 0 to 1, the edge connection precision curve shows a clear monotonically increasing trend, starting at a low value of approximately 0.05 and eventually approaching a high value of 0.95, indicating that a higher threshold can effectively improve connection accuracy and reduce the generation of erroneous edges. The edge connection recall curve, however, shows a monotonically decreasing trend, initially as high as approximately 0.95, and rapidly decreasing to near 0 as the threshold increases, indicating that an excessively high threshold can filter out actual edges, reducing recall. The edge connection F1 score curve first rises rapidly from a low value, and then... It peaks at around 0.56 in the range of approximately 0.35 to 0.5, and then slowly declines, reflecting the trade-off between precision and recall.

[0027] The thresholds are marked with dashed vertical lines in the figure. The threshold corresponds to the edge connecting to the peak of the F1 score, confirming the threshold in Example 1. Choosing 0.5 is reasonable, as it represents the optimal balance point, ensuring that the accuracy and coverage of edges in the generated globally consistent SPU semantic topology graph are both taken into account.

[0028] Example 2:

[0029] In practice, the collaborative network layer calls the globally consistent SPU semantic topology graph generated by the knowledge core layer to orchestrate point-to-point interactive tasks for multiple peer agents, generating a decentralized collaborative flow for SPU circulation tasks.

[0030] A local view copy of the globally consistent SPU semantic topology graph is deployed in each peer agent. A peer agent corresponds to a physical or logical node in the cooperative network layer. The cooperative network layer stores the complete adjacency matrix of the globally consistent SPU semantic topology graph. For any peer agent, the cooperative network layer determines the set of neighboring peer agents with whom it has direct communication relationships. Using the SPU semantic node corresponding to the peer agent as the central node, the cooperative network layer performs K-hop neighborhood sampling on the globally consistent SPU semantic topology graph, where K is set to 2. The reason for setting K to 2 is to achieve a balance between the communication overhead and negotiation accuracy of the cooperative network layer; a 2-hop neighborhood can cover both directly connected and indirectly connected nodes of the central node, ensuring the integrity of the local topology information. The sampled SPU semantic nodes and the SPU semantic edges between them constitute a local view subgraph. The cooperative network layer serializes the adjacency matrix and node feature matrix of the local view subgraph and deploys them to the local storage of the peer agent, forming a local view copy. Each peer agent's local view copy contains only the portion of the globally consistent SPU semantic topology graph that is related to the peer agent's neighborhood, and there is topological overlap between local view copies of different peer agents.

[0031] Based on the local view replicas, the cooperative network layer calculates the negotiated offset for SPU circulation tasks between adjacent peer agents. For an SPU circulation task to be processed, the cooperative network layer generates a task requirement vector. The task requirement vector is composed of the starting SPU semantic node identifier, the target SPU semantic node identifier, and constraint parameters involved in the SPU circulation task. The starting and target SPU semantic node identifiers are embedded representations of their respective SPU semantic nodes. The constraint parameters include time window limits and circulation cost caps, which are normalized and then concatenated after the embedded representations. Adjacent peer agents exchange their respective task requirement vectors. The first peer agent receives the second task requirement vector sent by the second peer agent, and the second peer agent receives the first task requirement vector sent by the first peer agent. The first peer agent performs path reasoning on its own local view replica using the first task requirement vector to generate a first path preference distribution; the second peer agent performs path reasoning on its own local view replica using the second task requirement vector to generate a second path preference distribution. Path reasoning is accomplished through a path preference reasoning model deployed on each peer agent. The path preference reasoning model comprises a graph convolutional encoder and a fully connected predictor. The graph convolutional encoder takes as input the adjacency matrix and node feature matrix of the local view subgraph, and the task requirement vector, and outputs a path selection preference score for each SPU semantic node. The graph convolutional encoder consists of two graph convolutional layers, each followed by ReLU activation. The fully connected predictor converts the node preference scores output by the graph convolutional encoder into a probability distribution pointing to the next SPU semantic node. The first peer agent aggregates the node probability distributions on all possible paths from the starting node to the target node in its own local view replica to obtain a first path preference distribution. Similarly, the second peer agent generates a second path preference distribution. The first and second peer agents exchange their respective first and second path preference distributions.

[0032] The cooperative network layer calculates the negotiation offset for SPU flow tasks between adjacent peer agents based on the first and second path preference distributions of the exchange. The negotiation offset is calculated using the following formula:

[0033] in, This represents the negotiation offset between the first peer agent p and the second peer agent q for the same SPU circulation task. This represents the intersection of SPU semantic nodes between the local view copies of the first peer agent and the local view copies of the second peer agent. for One of the SPU semantic nodes. In the first path preference distribution generated by the first peer agent p, the SPU semantic node represents... The probability value of being included in the task path. The value range is [0,1]. In the second path preference distribution generated by the second peer agent q, the SPU semantic node represents... The probability value of being included in the task path. The value range is [0,1]. It is a smoothing constant. The value is set to This is used to prevent numerical overflow when taking the logarithm with zero probability. For SPU semantic nodes Negotiation weight, The value is an SPU semantic node. The betweenness centrality values ​​in the globally consistent SPU semantic topology graph are pre-calculated and stored as node features in the globally consistent SPU semantic topology graph. The value range is [0,1].

[0034] The interaction sequences of neighboring peer agents are adjusted based on negotiated offsets to generate a decentralized cooperative flow. The cooperative network layer will... With the preset negotiation tolerance threshold Comparison, The value is set to 0.5, based on the parameter scanning during the development phase, which minimizes the global task completion time of the decentralized collaborative flow. Less than or equal to At this time, the cooperative network layer maintains the current SPU flow task interaction sequence between the first peer agent and the second peer agent unchanged. Greater than At this time, the cooperative network layer adjusts the interaction sequence. The adjustment method is as follows: the cooperative network layer determines the cause... Exceeding The main difference node, the main difference node is that makes The top L SPU semantic nodes with the largest values ​​are selected, with L set to 3. The rationale for setting L to 3 is that intervening only in the three nodes with the largest differences effectively reduces the negotiation offset and avoids over-adjustment. The cooperative network layer inserts a direct negotiation step regarding the main differing nodes into the interaction sequence between the first and second peer agents. In this direct negotiation step, the first and second peer agents exchange finer-grained neighborhood information from their respective local view copies for the main differing nodes, and determine whether to include or exclude the main differing nodes from the task path based on majority voting principles. After direct negotiation, the first and second peer agents update their respective first and second path preference distributions and recalculate the negotiation offset until it falls below the negotiation tolerance threshold. Or it may reach the maximum number of negotiation rounds. The maximum number of negotiation rounds is set to 5, based on limiting communication overhead while ensuring negotiation convergence.

[0035] After adjusting the interaction sequences of all neighboring peer agents, the cooperative network layer splices together the local path segments determined by all peer agents to form a complete decentralized cooperative flow for SPU circulation tasks. The decentralized cooperative flow is output in the form of a task execution sequence, which sequentially records the peer agent identifiers executing the SPU circulation tasks and the corresponding SPU circulation operation instructions.

[0036] See Figure 4 The horizontal axis in the figure represents the negotiated tolerance threshold. The value ranges from 0 to 1. The left vertical axis corresponds to the global task completion time, represented by the solid red line, in seconds; the right vertical axis corresponds to the average number of negotiation rounds, represented by the dashed purple line, in the number of rounds. The legend clearly distinguishes the indicators represented by the two curves.

[0037] With the negotiation tolerance threshold As the threshold gradually increases from 0 to approximately 0.5, the global task completion time (represented by the solid red line) shows a clear downward trend, decreasing from approximately 29 seconds to a minimum of about 15 seconds. This stage indicates that appropriately relaxing the negotiation tolerance threshold can effectively reduce task completion time and improve circulation efficiency. (Exceeding...) Afterwards, the global task completion time curve reversed and increased. This leads to an increase in task completion time, indicating that when the negotiation tolerance threshold is too high, the task coordination effect decreases and the circulation efficiency is reduced.

[0038] The average number of negotiation rounds (purple dashed line) varies. The increase continues to decline. From It started at about 9 times and quickly decreased to The number of negotiation rounds initially decreased to approximately 2, then continued to decline slowly to near 1. This trend suggests that increasing the negotiation tolerance threshold can reduce the number of negotiation rounds between neighboring peer agents, thereby reducing communication and computational overhead.

[0039] The vertical gray dashed line in the figure indicates The key threshold position. This threshold corresponds to the preset negotiation tolerance threshold in Example 2. After parameter scanning, it is determined to be the optimal point with the shortest global task completion time, reflecting the reasonable setting of the negotiation tolerance threshold by the cooperative network layer.

[0040] Example 3:

[0041] In practical implementation, the autonomous evolution layer performs task trajectory tracking on the circulation tasks in the decentralized collaborative flow, generating task trajectory tensors. When generating the decentralized collaborative flow for SPU circulation tasks, the collaborative network layer binds a unique task identifier to each newly created SPU circulation task. This unique task identifier uses a 128-bit universally unique identification code, generated by the task allocation module of the collaborative network layer during the initial task creation phase, and injected into the decentralized collaborative flow along with the task requirement vector. Each collaborative step in the decentralized collaborative flow includes the peer agent identifier of the currently executing step, the SPU circulation operation instruction, and a global timestamp generated by the global logical clock of the collaborative network layer, with a precision of milliseconds.

[0042] For a unique task identifier In the SPU circulation task, during the execution of the decentralized collaborative flow, each time the task passes through a peer agent, the peer agent associates its own peer agent identifier with the current global timestamp and adds a unique task identifier. The autonomous evolution layer continuously collects records from the start of the SPU circulation task until the SPU circulation task status changes to completion or termination. The autonomous evolution layer arranges the collected records in ascending order by global timestamp, forming an ordered sequence. Each element in the sequence is a tuple consisting of the peer agent's identifier and the global timestamp. This ordered sequence is the task flow path.

[0043] Multiple task flow paths are stacked along the time dimension to generate a task trajectory tensor. The autonomous evolution layer defines an observation window, which is composed of a start timestamp. and window length Sure, The value is set to 1800 seconds, based on the lifecycle of a typical SPU circulation task. The self-evolutionary layer divides the observation window evenly into T time slices along the time axis, with each time slice having a length of... The value is 5 seconds. The value of T is... The total number of peer agents participating in task collaboration in the cooperative network layer is N. The autonomous evolution layer constructs a three-dimensional tensor. tensor The dimension is , where U is the total number of SPU circulation tasks processed within the observation window.

[0044] Index of any SPU circulation task Its value ranges from 1 to U; any time slice index Its value ranges from 1 to T; index of any peer agent Its value ranges from 1 to N. Task trajectory tensor The elements in are determined by the following formula:

[0045] in, Represents the task trajectory tensor in the task Time slice peer-to-peer intelligent agents The element value at that location. The autonomous evolution layer performs assignment operations by traversing each record in the task flow path. For tasks... A record in the corresponding task flow path, assuming the peer agent is identified as... The global timestamp is Then calculate the time slice index. The corresponding position in the task trajectory tensor The value is set to 1. The tensor elements of peer agents and time slice locations that do not appear in any records remain 0. The generated task trajectory tensor T is stored in the tensor storage module of the autonomous evolution layer for subsequent sparsity analysis.

[0046] The self-evolutionary layer adaptively adjusts the structural parameters of the knowledge core layer and the collaborative network layer based on the task trajectory tensor. In specific implementation, the self-evolutionary layer adjusts the task trajectory tensor... Sparsity analysis was performed to generate a sparsity distribution map. The sparsity analysis was conducted using the CP tensor decomposition method. The task trajectory tensor was then analyzed. The CP decomposition approximates it as the sum of R rank-1 components:

[0047] Where R is the rank of the CP decomposition, the value of R is determined by the explained variance ratio, specifically starting from 1 and increasing R, stopping when the explained variance ratio exceeds 90%, and finally the value of R is generally between 5 and 15. Let r be the weight coefficient of the r-th component. For task factor vectors, A time factor vector, For peer agent factor vectors, the symbol is... Let represent the outer product. The CP decomposition is solved using the alternating least squares method, with a maximum number of iterations set to 100 and a convergence threshold of . After decomposition, the autonomous evolution layer does not directly use the factor vectors, but instead reconstructs an approximate tensor of the task trajectory tensor based on the decomposition results. The smooth sparsity of each peer agent in each time slice is calculated using approximate tensors. In time slice sparsity on Defined as:

[0048] in, The element value at the corresponding position in the approximate tensor takes a continuous value between 0 and 1, representing the task. In time slice By peer intelligent agents The probability intensity of execution. The sparsity values ​​of all time slices and all peer agents constitute a sparsity distribution map, which is stored in matrix form, with the matrix row index corresponding to the time slice. Column indexes correspond to peer-to-peer intelligent agents .

[0049] The self-evolutionary layer performs graph matching between the sparsity distribution map and the globally consistent SPU semantic topology graph to determine the subgraph to be optimized. The graph matching process is executed based on node mapping rules. Each peer agent in the cooperative network layer... It holds a local view copy containing a central SPU semantic node and its K-hop neighborhood SPU semantic nodes, denoted as . The autonomous evolution layer is established from peer-to-peer intelligent agents. To its central SPU semantic node The direct mapping relationship, and the peer agents The sparsity values ​​of each time slice in the sparsity distribution map are assigned to the central SPU semantic node in the globally consistent SPU semantic topology graph. The temporal activity of the SPU semantic node is calculated as T. For other non-central SPU semantic nodes in the globally consistent SPU semantic topology graph, if the SPU semantic node appears simultaneously in the local view replicas of multiple peer agents, its temporal activity is taken as the maximum sparsity of all peer agents containing it in the same time slice. Thus, each SPU semantic node in the globally consistent SPU semantic topology graph obtains an activity sequence of length T. The autonomous evolution layer calculates the comprehensive activity of each SPU semantic node, which is defined as the average of the temporal activity sequences. For any SPU semantic edge, its edge activity is defined as the average of the comprehensive activities of the two SPU semantic nodes connected by the edge.

[0050] The self-evolutionary layer extracts the subgraph of the topology to be optimized based on overall activity and edge activity. A preset node activity threshold is used. The edge activity threshold is 0.2. The threshold value is 0.15. The node activity threshold and edge activity threshold are set based on the following: in the activity distribution curves across multiple historical observation windows, the 20th percentile is used as the low-activity boundary. Graph elements below this boundary are considered valuable for optimization because they have not participated in SPU circulation tasks for a long time. The self-evolutionary layer traverses all SPU semantic nodes in the globally consistent SPU semantic topology graph, and nodes with a comprehensive activity level below a certain threshold are selected. SPU semantic nodes are marked as low-activity nodes; all SPU semantic edges are traversed, and edges with activity levels lower than a certain threshold are marked as low-activity nodes. SPU semantic edges are marked as low-activity edges. All low-activity nodes, low-activity edges, and the connections between these nodes constitute a subgraph, which is the subgraph of the topology to be optimized.

[0051] Based on the subgraph to be optimized, SPU semantic nodes and SPU semantic edges in the globally consistent SPU semantic topology graph are added, deleted, or adjusted. This includes deletion and addition operations. Deletion operations target graph elements in the subgraph to be optimized. For each inactive SPU semantic node in the subgraph to be optimized, if the overall activity of the SPU semantic node is less than a preset deletion threshold for three consecutive observation windows... If the self-evolutionary layer removes the SPU semantic node from the globally consistent SPU semantic topology graph, it also removes all SPU semantic edges connected to that SPU semantic node. Deletion threshold. The value is set to 0.1, based on the assumption that if the overall activity level is below 0.1 for three consecutive observation windows, it indicates that the flow function of the SPU semantic node has been ineffective for a long time. For each inactive SPU semantic edge in the subgraph to be optimized, if its edge activity level is less than the edge deletion threshold... If the edge is deleted, the SPU semantic edge is removed by the self-evolutionary layer. Edge deletion threshold. Set to 0.1. New operations are based on strong transition relationships revealed in the task trajectory tensor. The autonomous evolution layer computes any two different peer agents. and The task transfer frequency between peer agents. The task transfer frequency is defined as: in the task trajectory tensor, the frequency of task transfers between peer agents... In time slice Execute the task And peer-to-peer intelligent agents In time slice Perform the same task The frequency value obtained by dividing the number of tasks by the total number of tasks U is denoted as . .when Exceeding the preset threshold for new frequency At that time, the autonomous evolution layer determines the peer intelligent agents. Central SPU semantic node and peer intelligent agents Central SPU semantic node And check in the globally consistent SPU semantic topology graph and Check if an SPU semantic edge already exists between them. If not, add a new connection. and The SPU semantic edge, whose initial weight is set to the transition frequency. Add a new frequency threshold. The value is set to 0.12, based on filtering out strong transition relationships that exceed the average transition frequency by two standard deviations, to ensure the necessity of adding new edges.

[0052] After completing the above additions, deletions, and adjustments, the autonomous evolution layer persists the updated globally consistent SPU semantic topology graph to the knowledge core layer and notifies the collaborative network layer. Upon receiving the update notification, the collaborative network layer regenerates a local view copy for each peer agent and replaces the original copy in the peer agent's local storage, thereby completing the adaptive adjustment of the structural parameters of the knowledge core layer and the collaborative network layer.

[0053] Example 4:

[0054] In specific implementation, please refer to Figure 2 The collaborative network layer includes an interaction context fusion module. This module is used to fuse the SPU circulation context information exchanged between neighboring peer agents in a decentralized collaborative flow, generate a fused context feature vector, and inject this fused context feature vector into the decentralized collaborative flow.

[0055] It is understandable that in the decentralized collaborative flow generated by the collaborative network layer for SPU circulation tasks, two adjacent peer agents will exchange context description data related to the SPU circulation task when performing point-to-point interactive task orchestration. This context description data constitutes the SPU circulation context information. The SPU circulation context information consists of two parts: SPU attribute metadata and transaction parameter metadata. The SPU attribute metadata records the specification attribute description of the SPU to be circulated, and the transaction parameter metadata records the transaction constraints directly associated with the SPU circulation task. When the first peer agent and the second peer agent in the collaborative network layer establish a point-to-point collaborative link, the first peer agent generates the first SPU circulation context information, and the second peer agent generates the second SPU circulation context information. The first and second SPU circulation context information are exchanged on the point-to-point collaborative link in the form of structured data packets.

[0056] The interaction context fusion module extracts SPU attribute metadata and transaction parameter metadata exchanged between adjacent peer agents. An instance of the interaction context fusion module is deployed within the first peer agent. Upon receiving the second SPU circulation context information sent by the second peer agent, this instance parses the second SPU circulation context information into second SPU attribute metadata and second transaction parameter metadata. Simultaneously, this instance reads the first SPU circulation context information generated by the first peer agent itself and extracts the first SPU attribute metadata and first transaction parameter metadata from it. Both the first and second SPU attribute metadata are represented using multi-dimensional feature vectors, with each dimension corresponding to an attribute field of the SPU. Attribute fields include brand identifier, category code, specification parameter group, quality grade, and place of origin code. Both the first and second transaction parameter metadata are represented using structured vectors, with the transaction parameter metadata including a price range vector, quantity demand scalar, time window vector, and settlement method code.

[0057] The interactive context fusion module performs spatial topology alignment processing on the SPU attribute metadata to obtain a spatial alignment vector. The input to the spatial topology alignment processing is the first SPU attribute metadata and the second SPU attribute metadata. The interactive context fusion module maps the first and second SPU attribute metadata to the same SPU semantic embedding space. The mapping operation is performed using a pre-trained SPU attribute encoder in the knowledge core layer. The SPU attribute encoder is a fully connected neural network containing two hidden layers: the first hidden layer has a dimension of 256, the second hidden layer has a dimension of 128, and the output layer has a dimension of 128. The input to the SPU attribute encoder is the original feature vector of the SPU attribute metadata, and the output is the SPU attribute semantic embedding vector. Inputting the first SPU attribute metadata into the SPU attribute encoder yields the first SPU attribute semantic embedding vector; inputting the second SPU attribute metadata into the same SPU attribute encoder yields the second SPU attribute semantic embedding vector.

[0058] The interaction context fusion module constructs a topological alignment transformation between the first SPU attribute semantic embedding vector and the second SPU attribute semantic embedding vector in the SPU semantic embedding space. This topological alignment transformation is implemented based on shared SPU semantic nodes in the respective local view replicas of the two peer agents. The interaction context fusion module reads the intersection of SPU semantic nodes in the local view replicas held by the first and second peer agents and extracts the relative position coordinates of these intersection nodes in the globally consistent SPU semantic topology graph. The position coordinates use the Laplacian feature map coordinates of the SPU semantic nodes in the globally consistent SPU semantic topology graph. The dimension of the Laplacian feature map coordinates is set to 64. For each SPU semantic node in the intersection nodes, the interaction context fusion module calculates the first cosine similarity between the first SPU attribute semantic embedding vector and the node features of that SPU semantic node, and the second cosine similarity between the second SPU attribute semantic embedding vector and the node features of that SPU semantic node. Using the first and second cosine similarities as distribution weights, the Laplacian feature map coordinates corresponding to the intersection nodes are weighted and summed to obtain the first and second spatial centroid vectors. The first and second spatial centroid vectors are subtracted to obtain the topological offset vector. The interactive context fusion module subtracts the topological offset vector from the second SPU attribute semantic embedding vector to obtain the aligned second SPU attribute semantic embedding vector. The aligned second SPU attribute semantic embedding vector is concatenated with the first SPU attribute semantic embedding vector along the vector dimension to obtain a 256-dimensional concatenated vector, which is the spatial alignment vector.

[0059] The interaction context fusion module performs transaction intent weighting processing on the transaction parameter metadata to obtain a weighted intent vector. The inputs to the transaction intent weighting processing are the first and second transaction parameter metadata. The interaction context fusion module constructs a transaction intent encoder, which is a two-input fully connected network. The first input branch of the transaction intent encoder receives the first transaction parameter metadata, and the second input branch receives the second transaction parameter metadata. Each of the first and second input branches contains a 64-dimensional embedding layer, which maps the original transaction parameter metadata to a 64-dimensional embedding representation. The two 64-dimensional embedding representations are concatenated to form a 128-dimensional concatenated representation. The concatenated representation is fed into a 128-dimensional hidden layer, and then passes through a sigmoid activation function to output a scalar value, which is the intent coordination factor. The value range of the intent coordination factor is [0,1].

[0060] The network parameters of the transaction intent encoder are obtained through pre-training. The training data consists of successful and failed collaboration records of neighboring peer agents in historical SPU circulation tasks. In successful collaboration records, the transaction intentions represented by the first and second transaction parameter metadata match, and the training label is set to 1; in failed collaboration records, the transaction intentions represented by the first and second transaction parameter metadata conflict, and the training label is set to 0. Training uses a binary cross-entropy loss function, with Adam as the optimizer, a learning rate of 0.0005, a batch size of 128, a maximum training epoch of 50, and an early stopping strategy, terminating training when the validation set loss does not decrease for 5 consecutive epochs.

[0061] The interaction context fusion module uses a transaction intent encoder to process the first and second transaction parameter metadata to obtain an intent coordination factor. The module multiplies the first transaction parameter metadata by the intent coordination factor to obtain weighted first transaction parameter metadata; it then multiplies the second transaction parameter metadata by the difference between the value 1 and the intent coordination factor to obtain weighted second transaction parameter metadata. Finally, the weighted first and second transaction parameter metadata are concatenated as vectors to obtain a weighted intent vector. The dimension of the weighted intent vector is twice the dimension of the first transaction parameter metadata.

[0062] The interaction context fusion module concatenates the spatial alignment vector and the weighted intent vector to generate a fused context feature vector. Before concatenation, the spatial alignment vector undergoes a linear transformation layer, compressing its dimension from 256 to 128. The weighted intent vector undergoes another linear transformation layer, unifying its dimension to 128. The concatenation of these two 128-dimensional vectors yields a 256-dimensional vector, which is the fused context feature vector.

[0063] It can be understood that the fused context feature vector simultaneously implies the semantic alignment relationship between the first and second peer agents at the SPU attribute level and the intent coordination relationship at the transaction parameter level. The interaction context fusion module injects the fused context feature vector into the decentralized collaborative flow. The injection method is as follows: the interaction context fusion module replaces the original task requirement vector used by the first peer agent when calculating the negotiation offset of the SPU flow task between adjacent peer agents with the fused context feature vector. Specifically, before the first and second peer agents exchange task requirement vectors, the interaction context fusion module weights and superimposes the original task requirement vector and the fused context feature vector to obtain the enhanced task requirement vector. The weighting formula is:

[0064] in, This represents the vector of enhanced task requirements. Represents the original task requirement vector. This represents the fusion context feature vector. This represents the retention factor for the original task requirements. This represents the fusion context feature injection coefficient. The value is set to 0.7. The value was set to 0.3, based on a balance between preserving the original task intent and introducing collaborative context information. This coefficient combination was obtained through grid search during the offline optimization phase, using the average task completion rate of the decentralized collaborative flow as the evaluation metric. The search step size is 0.05, and the search space ranges from 0.05 to 0.95. After the enhanced task requirement vector replaces the original task requirement vector, it is transmitted along with the normal interaction process between adjacent peer agents in the cooperative network layer, participating in the calculation of path preference distribution and negotiation offset, thereby realizing information enhancement of the decentralized cooperative flow.

[0065] See Figure 5 In the figure, the horizontal axis represents the fusion context feature injection coefficient β, ranging from 0.05 to 0.95, indicating the weight of the fusion context feature vector when the interaction context fusion module weights and superimposes the enhanced task requirement vector in Example 4. The left vertical axis represents the average task achievement rate, ranging from 0 to 1, used to measure the completion effect of SPU circulation tasks in the decentralized collaborative flow; the right vertical axis represents the average negotiation offset, ranging from approximately 0.1 to 0.7, used to reflect the degree of deviation in task negotiation between adjacent peer agents.

[0066] In the figure, the pink solid line represents the average task achievement rate curve, and the brown dashed line represents the average negotiation offset curve. As β gradually increases from 0.05 to approximately 0.3, the average task achievement rate rises significantly, from approximately 0.3 to nearly 1.0, indicating that the fusion of contextual feature injection can effectively enhance the task requirement vector and improve task achievement. In the range of β between approximately 0.3 and 0.4, the average task achievement rate reaches its peak, approaching saturation, indicating optimal system coordination performance. At this point, the average negotiation offset remains at a low level, around 0.3, indicating high task negotiation efficiency and small deviation.

[0067] As β further increases beyond 0.4, the average task achievement rate gradually decreases, dropping below 0.1 when β reaches approximately 0.9. This indicates that excessive injection of fused context features leads to an imbalance in task requirement representation, affecting task achievement performance. Correspondingly, the average negotiation offset curve shows a continuous upward trend, gradually climbing from 0.1 to around 0.7, indicating an increase in negotiation differences between adjacent peer agents and a heavier negotiation burden.

[0068] Example 5:

[0069] In its implementation, the knowledge core layer also includes a semantic consistency arbitration module. This module detects version conflicts in the globally consistent SPU semantic topology graph. When the globally consistent SPU semantic topology graph undergoes adaptive adjustment of structural parameters in the autonomous evolution layer, it may generate multiple candidate version topology graphs due to differences in the timing of local view updates by different peer agents, or inconsistencies in topology modification suggestions proposed in parallel by multiple peer agents in the cooperative network layer. The semantic consistency arbitration module continuously monitors the major version number and all registered derivative version numbers of the globally consistent SPU semantic topology graph stored in the knowledge core layer. Version numbers use a triple format, where the triple includes the major version number, minor version number, and revision number. When the semantic consistency arbitration module detects two or more candidate version topology graphs with the same major and minor version numbers but different revision numbers within the same time window, it determines that a version conflict has occurred.

[0070] When a version conflict is detected, the semantic consistency arbitration module generates an arbitration request. The arbitration request includes an arbitration request identifier, a conflict discovery timestamp, a list of version numbers of all candidate version topologies involved in the conflict, and a list of source peer agent identifiers for each candidate version topology. The semantic consistency arbitration module calls the broadcast interface of the cooperative network layer to broadcast the arbitration request to multiple peer agents within the cooperative network layer. It can be understood that peer agents receiving the arbitration request should be those currently active whose local view replicas contain SPU semantic nodes related to the conflict, with the selection criterion being that the intersection of a peer agent's local view replica and any node in the candidate version topology is not empty. The cooperative network layer dynamically determines the broadcast range based on this condition and pushes the arbitration request to every peer agent within that range.

[0071] Each peer agent receiving an arbitration request executes its local version voting logic to generate a version voting result. This local version voting logic is deployed within the collaborative processing unit of each peer agent. Upon receiving an arbitration request, each peer agent extracts a description summary of all candidate version topologies from the knowledge core layer temporary storage area or the collaborative network layer message body. The description summary of a candidate version topology includes a list of newly added SPU semantic node identifiers, a list of deleted SPU semantic node identifiers, and a list of modified SPU semantic edges and their weights. Each peer agent compares the description summary of each candidate version topology with its own local view replica, calculating a compatibility score between each candidate version topology and the local view replica. The compatibility score is derived by calculating the degree of conflict between the operation sequence in the description summary and the existing graph structure in the local view replica, as well as the expected benefit. The degree of conflict is defined as the number of times a deleted SPU semantic node from a candidate version topology appears on the critical path of the local view replica, and the expected benefit is defined as the number of SPU semantic edges added to a candidate version topology that directly connect previously unconnected pairs of SPU semantic nodes in the local view replica. The compatibility score is calculated by subtracting the conflict severity score from the expected benefit score and multiplying it by a penalty coefficient. The penalty coefficient is set to 2, based on the principle that the damage weight of a deletion operation to the local collaborative task is twice the benefit weight of a addition operation, thus protecting existing usable collaborative paths. Peer agents select the candidate version of the topology graph with the highest compatibility score as their supported object and generate a version vote. The version vote includes an arbitration request identifier, peer agent identifiers, the version number of the supported candidate topology graph, and the compatibility score.

[0072] The semantic consistency arbitration module receives version voting results from multiple peer agents. It sets a voting collection timeout of 30 seconds, based on the window length required to ensure at least 95% of peer agents within the broadcast range return voting results under typical network latency. During the timeout, the module caches all received version voting results in the arbitration result queue. After the timeout expires, the module stops receiving data and begins determining the authoritative version topology based on the version voting results.

[0073] In its implementation, the semantic consistency arbitration module determines the authoritative version topology based on the version voting results, including: counting the number of supporting votes in the version voting results for each candidate version topology. The semantic consistency arbitration module iterates through each version voting result in the arbitration result queue, extracts the version number of the supported candidate version topology, and counts each distinct version number. For a candidate version topology, its number of supporting votes equals the number of version voting results with the same version number as that candidate version topology. It's understood that if a vote is cast for a version number other than a candidate version, that vote is marked as invalid and not counted in the statistics.

[0074] The semantic consistency arbitration module uses the candidate version topology graph with the most supporting votes as the authoritative version topology graph. If multiple candidate version topologies have the same number of supporting votes and are all the highest, the semantic consistency arbitration module uses the candidate version topology graph with the larger minor version number as the authoritative version topology graph; if the minor version numbers are the same, the candidate version topology graph with the larger revision number is selected as the authoritative version topology graph. If the number of votes is the same and the version numbers are exactly the same, it is considered a duplicate proposal, and either one can be selected.

[0075] The semantic consistency arbitration module invokes the authoritative version of the topology graph to overwrite and update the globally consistent SPU semantic topology graph. This overwrite update operation is performed in the knowledge core layer, and the specific steps are as follows: The semantic consistency arbitration module obtains the complete graph data of the authoritative version of the topology graph, which includes the authoritative SPU semantic node set, the authoritative SPU semantic edge set, and the weight matrix of the authoritative SPU semantic edges. The semantic consistency arbitration module completely replaces the graph data of the globally consistent SPU semantic topology graph currently stored in the knowledge core layer with the graph data of the authoritative version of the topology graph, and increments the minor version number in the major version number of the globally consistent SPU semantic topology graph by 1, while resetting the revision number to zero. After completing the overwrite update, the semantic consistency arbitration module sends a topology graph update notification to the collaborative network layer, and each peer agent in the collaborative network layer refreshes its deployed local view copy accordingly.

[0076] It is understandable that the collaborative network layer also includes a collaborative result attribution module. This module is used to backtrack the execution path of the final result of SPU circulation tasks in a decentralized collaborative flow, generating an execution path backtracking chain. When an SPU circulation task is completed in a decentralized collaborative flow and its status changes to completed or terminated, the collaborative network layer pushes the final result information of that SPU circulation task to the collaborative result attribution module. The final result information includes a unique task identifier, task status code, task reward value, and the complete task flow path generated during task execution. In the collaborative network layer, the task flow path is synchronously recorded by the autonomous evolution layer during task execution. At this time, the collaborative result attribution module directly retrieves the task flow path corresponding to the task from the autonomous evolution layer. The task flow path is an ordered sequence, where each element records the identifier of an executing peer agent and a global timestamp. The execution path backtracking operation is as follows: according to the global timestamp from the start to the end of the task, each execution step in the task flow path is extracted one by one, and the specific SPU circulation operation instruction executed in that step is marked. Simultaneously, the collaborative result attribution module determines whether there is a collaborative interaction between the preceding and following peer agents for every two consecutive execution steps. If so, it records the collaborative interaction edge. All execution steps and their connections constitute the execution path backtracking chain, which is stored in the form of a directed acyclic graph, with the direction of the directed edges consistent with the time progression direction.

[0077] Based on the execution path backtracking chain, the collaborative outcome attribution module calculates the contribution weight of each participating peer agent. A participating peer agent is a node that appears at least once in the execution path backtracking chain. The contribution weight is quantified by analyzing the marginal contribution of each participating peer agent to the successful execution of the task. The collaborative outcome attribution module uses the Shapley value attribution method to calculate the contribution weight. The execution path backtracking chain is viewed as a collaborative game, with all participating peer agents forming the participant set, and the task payoff representing the total payoff.

[0078] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An AI-native e-commerce network multi-agent SPU collaborative circulation system, characterized in that: include: The knowledge core layer is used to construct SPU standardized knowledge representation units and perform semantic topology mapping processing on the SPU standardized knowledge representation units to generate a globally consistent SPU semantic topology graph. The collaborative network layer is used to invoke the globally consistent SPU semantic topology graph, orchestrate point-to-point interactive tasks for multiple peer agents, and generate a decentralized collaborative flow for SPU circulation tasks. The self-evolution layer is used to perform task trajectory tracking on the circulating tasks in the decentralized collaborative flow, generate task trajectory tensors, and adaptively adjust the structural parameters of the knowledge core layer and the collaborative network layer according to the task trajectory tensors.

2. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 1, characterized in that, The knowledge core layer performs semantic topology mapping processing on the SPU standardized knowledge representation unit to generate a globally consistent SPU semantic topology graph, including: Extract the attribute dimension and relation dimension from the SPU standardized knowledge representation unit to construct an attribute-relation tuple set; Perform topological embedding processing on the attribute-relation tuple set to generate SPU semantic nodes and SPU semantic edges; The SPU semantic nodes are connected according to the SPU semantic edges to generate the globally consistent SPU semantic topology graph.

3. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 2, characterized in that, The collaborative network layer invokes the globally consistent SPU semantic topology graph to orchestrate point-to-point interaction tasks among multiple peer agents, generating a decentralized collaborative flow for SPU circulation tasks, including: Deploy a local view copy of the globally consistent SPU semantic topology graph in each peer agent; Based on the local view copy, calculate the negotiation offset of the SPU flow task between adjacent peer agents; The interaction sequence of adjacent peer agents is adjusted based on the negotiated offset to generate the decentralized cooperative flow.

4. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 3, characterized in that, The autonomous evolution layer performs task trajectory tracking on the flowing tasks in the decentralized collaborative flow, generating task trajectory tensors, including: Each SPU circulation task is assigned a unique task identifier; In the decentralized collaborative flow, each peer agent through which the SPU flow task passes is timestamped according to the unique task identifier to generate a task flow path; Multiple task flow paths are stacked according to the time dimension to generate the task trajectory tensor.

5. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 4, characterized in that, The self-evolutionary layer adaptively adjusts the structural parameters of the knowledge core layer and the collaborative network layer based on the task trajectory tensor, including: Sparseness analysis is performed on the task trajectory tensor to generate a sparseness distribution map; The sparsity distribution map is matched with the globally consistent SPU semantic topology graph to determine the topology subgraph to be optimized. Based on the subgraph to be optimized, the SPU semantic nodes and SPU semantic edges in the globally consistent SPU semantic topology graph are added, deleted, or adjusted.

6. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 5, characterized in that, The collaborative network layer also includes: The interaction context fusion module is used to fuse the SPU circulation context information exchanged between adjacent peer agents in the decentralized collaborative flow, generate a fused context feature vector, and inject the fused context feature vector into the decentralized collaborative flow.

7. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 6, characterized in that, The interaction context fusion module fuses the SPU circulation context information exchanged between neighboring peer agents to generate a fused context feature vector, including: Extract SPU attribute metadata and transaction parameter metadata exchanged between neighboring peer agents; Spatial topology alignment processing is performed on the SPU attribute metadata to obtain a spatial alignment vector; Perform transaction intent weighting processing on the transaction parameter metadata to obtain a weighted intent vector; The spatial alignment vector and the weighted intent vector are concatenated to generate the fusion context feature vector.

8. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 1, characterized in that, The core knowledge layer also includes: The semantic consistency arbitration module is used to perform version conflict detection on the globally consistent SPU semantic topology graph. When a version conflict is detected, an arbitration request is generated and the arbitration request is broadcast to multiple peer agents in the cooperative network layer. Receive the version voting results returned by the multiple peer agents, and determine the authoritative version topology based on the version voting results.

9. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 8, characterized in that, The semantic consistency arbitration module determines the authoritative version topology graph based on the version voting results, including: Count the number of supporting votes in the version voting results obtained for each candidate version topology graph; The candidate version topology graph with the most support votes will be used as the authoritative version topology graph. The authoritative version of the topology graph is invoked to overwrite and update the globally consistent SPU semantic topology graph.

10. The AI-native e-commerce network multi-agent SPU collaborative circulation system according to claim 1, characterized in that, The collaborative network layer also includes: The collaborative result attribution module is used to backtrack the execution path of the final result of the SPU circulation task in the decentralized collaborative flow and generate an execution path backtracking chain. Based on the execution path backtracking chain, the contribution weight of each participating peer agent is calculated, and the task rewards in the decentralized collaborative flow are allocated according to the contribution weight.

Images