A multi-agent collaborative reasoning oriented feasibility closure identification method and system

Abstract

This invention provides a method and system for identifying feasible closures in multi-agent collaborative reasoning. The method includes the following steps: locating the scene region where the target task is located, obtaining the reasoning tasks of all agents in the scene region, and constructing a first set of reasoning tasks; based on the dependencies between the reasoning tasks in the first set of reasoning tasks, constructing a first relation graph with each reasoning task as a node, and locating the target task in the first relation graph; determining the reasoning feasibility closure of the target task based on the first relation graph, and identifying the closure in the first relation graph; the step of determining the reasoning feasibility closure of the target task includes performing a proof-of-contrast necessity determination based on the position of the target task in the first relation graph, and determining the closure in the first relation graph through the proof-of-contrast necessity determination. This solution determines closures based on logical / semantic level relationships, enabling timely identification and processing.

Description

Technical Field

[0001] This invention relates to the field of multi-agent collaborative reasoning technology, and in particular to a feasibility closure identification method and system for multi-agent collaborative reasoning. Background Technology

[0002] In recent years, multi-agent systems have been widely applied in scenarios such as vehicle-road cooperation, multi-unmanned swarms, and edge collaborative reasoning. These scenarios typically involve multiple agents undertaking sub-tasks such as perception, modeling, reasoning, and control, forming a collaborative reasoning chain through message interaction and state sharing. In engineering implementation, this collaborative process is often abstracted as a structured workflow of "task set + dependencies": each reasoning sub-task acts as a node, and the data flow / sequence constraints between tasks act as directed dependency edges, thus forming a task dependency graph oriented towards the reasoning goal. The system needs to continuously complete collaborative reasoning and output executable control or decision results under dynamic environments and resource-constrained conditions.

[0003] For the organization and execution of the aforementioned collaborative reasoning tasks, existing technologies typically offer solutions from the perspective of "resource scheduling / task orchestration." Firstly, collaborative services deployed for edge computing often employ centralized or hierarchical orchestration control strategies: the edge-side control platform or orchestrator receives state / observation data from multiple agents, completes task splitting, resource allocation, and task distribution, and performs state synchronization or load migration between multiple edge nodes when necessary. This type of solution can support cross-node operation of multi-agent tasks, but its core is still driven by performance indicators such as resource utilization, throughput, or latency, focusing on "how to allocate resources / how to schedule tasks," while lacking an independent mechanism to determine the feasibility of the reasoning structure itself. Secondly, in terms of task structure modeling, existing technologies typically use directed graphs / directed acyclic structures (DAGs) to describe task dependencies and conduct critical path analysis, priority ranking, or partitioned execution accordingly: for example, determining scheduling priorities based on critical path length or task weights, or dividing tasks into parallelizable subgraphs based on dependencies to improve overall efficiency. These methods can optimize execution order at the performance level, but their judgment target is mainly performance optimization (such as minimizing completion time or meeting latency constraints), and does not use "whether the inference can still be established" as a criterion. Therefore, when a structural gap occurs in the collaborative inference link (such as missing key dependencies, unavailability of necessary tasks, or changes in dependency topology), the system may still be able to execute some sub-tasks and the performance indicators may not be significantly degraded, but the inference goal can no longer be established at the logical / semantic level. Thirdly, in dynamic environments, existing systems usually reschedule or reconfigure in a way that is event-driven or triggered by performance thresholds. For example, when queue backlog, excessive latency, excessive node load, or degraded link quality are observed, migration, scaling up or down, or priority adjustment are triggered. This type of triggering mechanism is essentially still based on performance signals and is suitable for "performance degradation" scenarios, but it is not sensitive to "structural failure" scenarios: even if the performance is still acceptable, if a key link in the inference link is missing, the inference goal may still be unattainable, and this type of failure is not necessarily accompanied by obvious performance anomalies, making it difficult for the system to take timely structural reconstruction measures.

[0004] However, existing methods primarily rely on performance metrics such as latency, throughput, and load to drive task splitting, scheduling, and reconfiguration, lacking a structural judgment mechanism based on whether the inference objective is still valid at the logical / semantic level. When critical dependencies are missing, necessary tasks are unavailable, or dependencies change, the system may still maintain some performance metrics, but the inference objective may become structurally unattainable, making it difficult to identify and address structural failures in a timely manner. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a feasibility closure identification method for multi-agent collaborative reasoning, in order to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides a feasibility closure identification method for multi-agent cooperative reasoning, the method comprising the following steps: Locate the scene area where the target task is located, obtain the reasoning tasks of all agents in the scene area, and construct a first set of reasoning tasks; Based on the dependencies between the reasoning tasks in the first set of reasoning tasks, each reasoning task is used as a node to construct a first relational graph, and the target task is located in the first relational graph. Based on the first relationship graph, the feasibility closure of the target task is determined, and the closure in the first relationship graph is determined. The step of determining the closure of the reasoning feasibility of the target task includes performing a proof-by-contrast necessity determination based on the position of the target task in the first relation graph, and determining the closure in the first relation graph through the proof-by-contrast necessity determination.

[0007] Using the above scheme, in the step of determining the closure of reasoning feasibility, the first set of reasoning tasks and the set of reasoning dependencies between tasks are used as input to construct a first relational graph to describe the logical order and semantic constraints of the reasoning tasks. The first relational graph is used to explicitly depict the necessary preconditions and state transit relationships between reasoning tasks. Then, by using the position of the target task in the first relational graph, reverse reasoning is performed, and the closure is determined based on the logical / semantic relationship of the reasoning target, which can be identified and handled in a timely manner.

[0008] In some embodiments of the present invention, the method further includes encapsulating and uploading a set of closures constructed from the closures in the first relational graph, so that the agent can complete the target task based on the closures in the set of closures.

[0009] In some embodiments of the present invention, the step of determining the inference feasibility closure of the target task includes a reverse extension determination. In the reverse extension determination step, the node where the target task is located is taken as the root node for reverse extension. Whenever the extension reaches a node, the inference task corresponding to that node is added to the result set, and the first relationship graph is updated based on the result set.

[0010] In some embodiments of the present invention, in the step of extending backward from the node where the target task is located as the root node, and adding the inference task corresponding to the node to the result set whenever the extension reaches a node, the node where the target task is located is used as the root node for the backward extension. Whenever the extension reaches a node, the inference task corresponding to the node is obtained, and the inference task is retrieved in the result set. If the inference task does not exist in the result set, the inference task is added to the result set; if the inference task already exists in the result set, the inference task is added to the result set.

[0011] In some embodiments of the present invention, in the step of performing a proof-of-contrast necessity determination based on the position of the target task in the first relation graph, the starting node and the dependent nodes adjacent to the node corresponding to the target task in the first relation graph are located. For any node in the first relation graph that is not the starting node and is not a dependent node, a second relation graph is constructed. Based on the second relation graph, it is determined whether the node currently being tried to be removed is a necessary node, and an attempt is made to remove each node that is not the starting node and is not a dependent node.

[0012] In some embodiments of the present invention, in the steps of removing any non-starting node and non-dependent node from the first relational graph, constructing a second relational graph, determining whether the node currently being attempted to be removed is a necessary node based on the second relational graph, and attempting to remove each non-starting node and non-dependent node, if any dependent node in the second relational graph does not have a complete path from the starting node to the dependent node, then the node currently being attempted to be removed is a necessary node, and another non-starting node and non-dependent node is removed from the first relational graph, the second relational graph is reconstructed, and it is determined again whether the node is a necessary node; if any dependent node in the second relational graph has a complete path from the starting node to the dependent node, then the node currently being attempted to be removed is a non-necessary node, and a further attempt is made to remove another non-starting node and non-dependent node from the current second relational graph, constructing a new second relational graph, until all non-starting nodes and non-dependent nodes have been traversed; the first or second relational graph after traversing all non-starting nodes and non-dependent nodes is the third relational graph.

[0013] In some embodiments of the present invention, the step of determining the closure of the reasoning feasibility of the target task based on the first relation graph further includes minimum completeness convergence. In the step of minimum completeness convergence, the reasoning task corresponding to the node in the third relation graph obtained from the first relation graph is taken as the closure, and a closure set is constructed.

[0014] A second aspect of the present invention also provides a feasibility closure identification system for multi-agent collaborative reasoning. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0015] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned feasibility closure identification method for multi-agent cooperative reasoning.

[0016] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0017] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0018] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0019] Figure 1 This is a schematic diagram of one implementation of the feasibility closure identification method for multi-agent cooperative reasoning in this scheme; Figure 2 This is a schematic diagram illustrating another implementation of the feasibility closure identification method for multi-agent collaborative reasoning in this scheme. Figure 3 This is a schematic diagram of the overall architecture of this solution; Figure 4 This is a schematic diagram of the closure identification process for the feasibility of this solution. Figure 5 A schematic diagram illustrating the workflow for determining the feasibility closure of this solution; Figure 6 This is a schematic diagram of the process for maintaining the validity of inference closures in this solution; Figure 7 This diagram illustrates the joint deployment of multiple MECs in a multi-vehicle collaborative merging scenario. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0021] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0022] This application addresses the common "task decomposability assumption" design approach in existing multi-agent collaborative reasoning systems. It finds that while existing technologies typically divide or schedule tasks based on execution cost, critical path length, resource utilization, or performance metrics, they lack a structural determination mechanism to determine whether the reasoning remains logically valid. When certain tasks are removed due to environmental changes, dependency breaks, or execution failures, the system may still maintain a runnable state in terms of performance metrics, but it may be unable to achieve the original reasoning goal at the logical or semantic level, resulting in structural failure. This type of structural failure does not necessarily manifest as increased latency or decreased throughput, making it difficult to explicitly characterize using traditional performance evaluation models.

[0023] To address the aforementioned issues, this application proposes a method and system for identifying reasoning feasibility closures in multi-agent collaborative reasoning scenarios. This method does not rely on performance metrics but uses "whether the reasoning objective remains valid at the structural level" as the sole criterion. It identifies the smallest indivisible set of tasks under the current dependency structure and maintains the structural validity of this set under dynamic environmental changes.

[0024] like Figure 1 As shown, this invention proposes a feasibility closure identification method for multi-agent cooperative reasoning, the steps of which include: Step S100: Locate the scene area where the target task is located, obtain the reasoning tasks of all agents in the scene area, and construct a first set of reasoning tasks; Step S200: Based on the dependencies between the reasoning tasks in the first set of reasoning tasks, construct a first relation graph with each reasoning task as a node, and locate the target task in the first relation graph; In its implementation, this scheme proposes a reasoning structure modeling and closure determination scheme for multi-agent collaborative reasoning scenarios. Its core lies in constructing a reasoning feasibility determination model with "whether the reasoning is still valid" as the sole criterion, used to identify a set of reasoning tasks that need to be considered holistically at the structural level. In this scheme, the reasoning structure modeling module takes the collaborative reasoning task set T = {t1,t2,…,tn} and the set of reasoning dependencies between tasks as input to construct a reasoning structure model describing the logical order and semantic constraints of the reasoning tasks. This reasoning structure model is used to explicitly characterize the necessary preconditions and state transit relationships between reasoning tasks. The reasoning feasibility determination module receives any subset of reasoning tasks S⊆T and the collaborative reasoning objective G, and calls the reasoning feasibility determination function F(S,G) to determine whether, under the condition of retaining only the task set S, a consistent reasoning result satisfying the reasoning objective G can still be generated. When F(S,G) = true, the current task set S is determined to have structural reasoning feasibility and is marked as a candidate set of reasoning feasibility closures; when F(S,G) = false, the task set is determined to not have the structural conditions for collaborative reasoning and further processing of the set is terminated.

[0025] The above-described determination process does not depend on the execution order of the inference tasks, the allocation of computing power, or the communication conditions, but only on the consistency constraints between the inference dependencies and the inference goal. By performing inference feasibility determination on the complete task set and its subsets, this scheme can output at least one candidate set of closures related to the inference goal, providing structured input for subsequent closure boundary convergence and rule enhancement processing.

[0026] Step S300: Based on the first relationship graph, determine the reasoning feasibility closure of the target task and determine the closure in the first relationship graph; Step S400, the step of determining the reasoning feasibility closure of the target task includes performing a proof-of-contrast necessity determination based on the position of the target task in the first relation graph, and determining the closure in the first relation graph through the proof-of-contrast necessity determination.

[0027] Using the above scheme, in the step of determining the closure of reasoning feasibility, the first set of reasoning tasks and the set of reasoning dependencies between tasks are used as input to construct a first relational graph to describe the logical order and semantic constraints of the reasoning tasks. The first relational graph is used to explicitly depict the necessary preconditions and state transit relationships between reasoning tasks. Then, by using the position of the target task in the first relational graph, reverse reasoning is performed, and the closure is determined based on the logical / semantic relationship of the reasoning target, which can be identified and handled in a timely manner.

[0028] like Figure 2As shown, in some embodiments of the present invention, the method further includes step S500, which involves encapsulating and uploading the closure set constructed from the closures in the first relational graph, so that the agent can complete the target task based on the closures in the closure set.

[0029] In some embodiments of the present invention, the step of determining the inference feasibility closure of the target task includes a reverse extension determination. In the reverse extension determination step, the node where the target task is located is taken as the root node for reverse extension. Whenever the extension reaches a node, the inference task corresponding to that node is added to the result set, and the first relationship graph is updated based on the result set.

[0030] In some embodiments of the present invention, in the step of extending backward from the node where the target task is located as the root node, and adding the inference task corresponding to the node to the result set whenever the extension reaches a node, the node where the target task is located is used as the root node for the backward extension. Whenever the extension reaches a node, the inference task corresponding to the node is obtained, and the inference task is retrieved in the result set. If the inference task does not exist in the result set, the inference task is added to the result set; if the inference task already exists in the result set, the inference task is added to the result set.

[0031] In some embodiments of the present invention, in the step of performing a proof-of-contrast necessity determination based on the position of the target task in the first relation graph, the starting node and the dependent nodes adjacent to the node corresponding to the target task in the first relation graph are located. For any node in the first relation graph that is not the starting node and is not a dependent node, a second relation graph is constructed. Based on the second relation graph, it is determined whether the node currently being tried to be removed is a necessary node, and an attempt is made to remove each node that is not the starting node and is not a dependent node.

[0032] In some embodiments of the present invention, in the steps of removing any non-starting node and non-dependent node from the first relational graph, constructing a second relational graph, determining whether the node currently being attempted to be removed is a necessary node based on the second relational graph, and attempting to remove each non-starting node and non-dependent node, if any dependent node in the second relational graph does not have a complete path from the starting node to the dependent node, then the node currently being attempted to be removed is a necessary node, and another non-starting node and non-dependent node is removed from the first relational graph, the second relational graph is reconstructed, and it is determined again whether the node is a necessary node; if any dependent node in the second relational graph has a complete path from the starting node to the dependent node, then the node currently being attempted to be removed is a non-necessary node, and a further attempt is made to remove another non-starting node and non-dependent node from the current second relational graph, constructing a new second relational graph, until all non-starting nodes and non-dependent nodes have been traversed; the first or second relational graph after traversing all non-starting nodes and non-dependent nodes is the third relational graph.

[0033] In some embodiments of the present invention, the step of determining the closure of the reasoning feasibility of the target task based on the first relation graph further includes minimum completeness convergence. In the step of minimum completeness convergence, the reasoning task corresponding to the node in the third relation graph obtained from the first relation graph is taken as the closure, and a closure set is constructed.

[0034] In some embodiments of the present invention, the method further includes inference closure validity maintenance. In the inference closure validity maintenance step, the changes in the scene region are collected at preset time intervals. If there is an increase or decrease in inference tasks in the scene region, the closure set is adjusted and maintained.

[0035] In some embodiments of the present invention, the step of maintaining the validity of the reasoning closure further includes adding or removing nodes in the third relation graph based on the added or removed reasoning tasks, maintaining the third relation graph using the reverse extension determination and the proof by contradiction necessity determination, and obtaining the maintained closure set.

[0036] In the specific implementation process, a closure boundary convergence mechanism based on minimum completeness constraints is introduced for the candidate set of reasoning feasibility closures to eliminate redundant reasoning tasks and determine the unique reasoning feasibility closure.

[0037] Necessity test module based on proof by contradiction As input, for each of these reasoning tasks Perform the necessity checks in sequence.

[0038] Specifically, for any reasoning task Construct candidate subsets And call the reasoning feasibility determination function. Determine whether collaborative reasoning still holds true after the reasoning task is removed. At that time, determine the reasoning task The necessary tasks for the establishment of reasoning, and mark them as those that must be retained; when At that time, determine the reasoning task Tasks deemed unnecessary are marked as removable. Based on these markings, the minimum completeness convergence module performs a boundary update operation on the closure candidate set, retaining only all inference tasks marked as necessary, thus forming a task set. .when satisfy If there is no reasoning task that can be further removed without affecting the validity of the reasoning, then... It is determined to be the final inference feasibility closure.

[0039] Through the above-described proof-by-contradiction necessity test and boundary convergence process, this scheme can determine the unique and minimal closure that satisfies the reasoning feasibility constraint based on a unified decision rule when there are multiple candidate sets or ambiguous boundaries.

[0040] like Figure 4 As shown, before performing closure recognition for reasoning feasibility, this application first formally models the task relationships in a multi-agent collaborative reasoning scenario, constructing a computable structural abstract model. This part constitutes the input foundation of the closure recognition algorithm, and its core objective is to transform the original semantic description of the collaborative reasoning relationship into a directed structural graph expression, making the subsequent logical judgment process verifiable and computable.

[0041] In the specific implementation process, the system first obtains the set of tasks T involved in the current collaborative reasoning session. This set represents all subtasks participating in the same reasoning goal, and is defined as: in, Indicates the first i Each reasoning subtask. It has at least the following structural attributes: task identifier Input set Output set Implementing entity and dependency sets . Indicates task Required input data set Indicates task The set of output results. Dependency set. This represents the set of precursor tasks that must be completed first at the logical level. The above attributes do not involve execution costs or resource consumption; they are only used to characterize logical dependencies.

[0042] The system simultaneously constructs a set D of inter-task dependencies. Dependencies are defined as directed pairs: like( , If )∈D, then it represents the task. The execution is logically dependent on the task. The output result. This dependency reflects the logical propagation direction of the inference chain, rather than the physical execution order or time constraints.

[0043] Based on the task set T and dependency relationship D, construct a collaborative reasoning dependency graph: in The graph is a directed graph structure, where nodes represent tasks and edges represent dependencies. This graph is used to depict the structural propagation paths between inference tasks, providing a basis for subsequent feasibility determination functions.

[0044] To characterize the structural relationship between the task and the final inference goal, the system introduces an inference goal node G. The inference goal G can represent the final decision output, state determination result, or response generation result. In the graph structure, the goal node is typically a terminal node.

[0045] To determine whether a task participates in the formation of the reasoning goal's structure, this application defines a reachability function: Among them, Reach ( G)=1 indicates that there exists a dependency graph G_task from the task. A directed path to target G; Reach( `G)=0` indicates that the path does not exist. This function can be implemented using graph traversal algorithms, such as depth-first search or breadth-first search. This section does not improve the traversal algorithm itself, but only uses it as a tool for determining the structure.

[0046] Based on the reachability function, generate a candidate set of closures: This candidate set represents all tasks that have a structural relationship with the target G. This step is used to reduce the search space for subsequent closure identification, prevent irrelevant tasks from entering the proof-by-contrast stage, and thus improve the stability and efficiency of the algorithm.

[0047] After structural modeling is completed, the system forms the following data flow structure: (T,D,G)→Construct →Calculate Reach( ,G)→Generate .

[0048] This structural modeling mechanism has the following technical characteristics: First, it achieves a formal transformation from semantic reasoning description to graph structure model, making the reasoning structure computable and verifiable; Second, tasks irrelevant to the goal are filtered out through reachability functions to avoid performance indicators from participating in the screening process and to ensure structural purity. Third, the collaborative reasoning scenario is abstracted into a directed structural model, providing a rigorous input for subsequent proof-of-contradiction necessity tests and minimum completeness convergence; Fourth, maintain decoupling from execution control and resource scheduling logic, so that the modeling mechanism can be deployed independently in different levels of systems.

[0049] Through the aforementioned collaborative reasoning structure modeling mechanism, this application realizes the structural abstraction of multi-agent collaborative reasoning relationships, providing a strict, unified, and computable structural input foundation for the subsequent closure recognition stage.

[0050] like Figure 5 As shown, after completing the collaborative reasoning structure modeling and generating the closure candidate set... Next, this application enters the core stage of closure determination—the reasoning feasibility closure determination mechanism. The goal of this stage is to identify the smallest set of tasks that are logically and semantically indivisible from the candidate set. This mechanism satisfies both the "structural feasibility" and "minimum" constraints. It does not rely on task weight ranking, critical path length, or performance cost assessment, but uses "whether the reasoning objective still holds" as the sole criterion, constructing the necessity determination logic through a proof-of-contrast deletion test.

[0051] To characterize the structural impact of task subsets on the reasoning objective, this application defines a reasoning feasibility determination function: in, Let S represent a subset of tasks and G represent the inference goal. The physical meaning of F(S, G) is to determine whether the inference goal G is still reachable from the dependency graph and maintains dependency integrity, given the structure of only retaining the subset of tasks S.

[0052] Specifically, F(S, G) = true if and only if the following structural condition is satisfied: (1) In the subgraph There exists a directed path from at least one starting task to the target G; (2) There is no dependency break on the path, that is, for any side ,like ∈S, then ∈S; (3) The target node G still has complete preceding dependencies in the subgraph.

[0053] Otherwise, F(S, G) = false.

[0054] The aforementioned decision function constitutes the first constraint condition for closure recognition, which can be called the "structural feasibility constraint".

[0055] To identify necessary tasks in the task set, this application introduces the deletion operator Remove(C, ti), defined as: Its physical meaning is: to delete a single task from the task set C. And retain the remaining tasks to form a new subset. Based on this operator, the candidate set... Each task in Construct a subset of proofs by contradiction: Then calculate .like This indicates that during deletion The post-inference objective no longer holds, therefore This is a necessary task for the reasoning to hold true; if This means This is a non-essential task.

[0056] Based on this, a necessity marker function is defined: Where N(ti) = 1 represents the task This is a necessary task.

[0057] By iterating through all tasks in the candidate set, we obtain the necessary task set: The above steps constitute the second constraint system for closure recognition, which can be called "necessity constraint".

[0058] To ensure that the set of closures satisfies both completeness and minimality, this application introduces a minimum completeness constraint: Completeness constraints: Right now It can guarantee that the reasoning objective is valid.

[0059] Minimality constraint In other words, removing any task will cause the inference to fail.

[0060] like If both of the above conditions are met, then it is determined as the final set of closures: If the conditions are not met, update the candidate set: The necessity check process is repeated until the set converges.

[0061] To characterize the convergence process, this application defines a finite-step convergence rule: In each iteration, if any unnecessary tasks are removed, the size of the set is strictly reduced; since the task set is finite, the algorithm will terminate within a finite number of steps.

[0062] This convergence rule constitutes the third constraint for closure recognition, which can be called the "minimal completeness constraint".

[0063] At the algorithm implementation level, the closure determination process can be represented as the following execution flow: First, input the candidate set. Next, each task is deleted and its feasibility is determined; then, a new set is constructed based on the necessity flag; finally, a convergence check is performed and the result is output. This process constitutes... Figure 3 The internal workflow is shown.

[0064] Compared with existing methods based on critical path analysis or task prioritization, this mechanism has the following technical features: First, the sole criterion is whether the reasoning is still valid, thus avoiding interference from performance metrics. Second, the necessary structural tasks are identified through a proof-of-contrast deletion mechanism, rather than through weighted sorting. Third, the minimum completeness constraint ensures that the closure set cannot be further divided; Fourth, a clear constraint system and convergence logic are established to make the identification process verifiable and deterministic.

[0065] Through the above mechanism, this application achieves accurate identification of the minimum indivisible task set in a multi-agent collaborative reasoning structure, providing a structural basis for subsequent scheduling, control, or resource allocation.

[0066] In some embodiments of the present invention, the method further includes inference closure validity maintenance. In the inference closure validity maintenance step, the changes in the scene region are collected at preset time intervals. If there is an increase or decrease in inference tasks in the scene region, the closure set is adjusted and maintained.

[0067] In some embodiments of the present invention, the step of maintaining the validity of the reasoning closure further includes adding or removing nodes in the third relation graph based on the added or removed reasoning tasks, maintaining the third relation graph using the reverse extension determination and the proof by contradiction necessity determination, and obtaining the maintained closure set.

[0068] In the specific implementation process, this solution introduces a closure state awareness and failure handling mechanism to manage the validity of the current reasoning feasibility closure when the collaborative reasoning environment changes.

[0069] In this scheme, the environment state perception module periodically acquires the collaborative reasoning environment state set E, which includes agent node availability, reasoning task execution status, and information on changes in reasoning dependencies. Let the currently used reasoning feasibility closure be... .

[0070] The closure failure determination module is based on the current closure. Given the set of environment states E, call the closure validity determination function. This is used to determine whether the closure still satisfies the reasoning feasibility constraint under the current environmental conditions.

[0071] when When the system maintains the current closure, it continues to execute the collaborative reasoning process based on that closure; when When the current closure is determined to be structurally invalid, the collaborative reasoning process based on the closure is immediately terminated to avoid continuing the reasoning task when the reasoning structure is no longer valid.

[0072] After a closure fails, the closure re-identification trigger module re-acquires the currently available set of inference tasks and their dependencies, and sequentially calls the inference feasibility closure determination process described in Scheme 1 and the closure boundary convergence process described in Scheme 2 to generate a new inference feasibility closure for use in the subsequent collaborative inference process.

[0073] like Figure 6As shown, after obtaining the set of closures with reasoning feasibility, this application further provides a dynamic validity maintenance mechanism for closures. This mechanism continuously determines whether the closure structure is still valid during multi-agent collaborative reasoning and triggers a closure re-identification process when a closure fails, thereby ensuring that the closure result remains consistent with the dynamic environment state. This maintenance mechanism belongs to the dynamic maintenance layer in the "structural modeling layer—logic determination layer—dynamic maintenance layer" architecture of this application. It is decoupled from the closure determination module and does not depend on specific scheduling strategies or resource configuration methods.

[0074] In its implementation, the system defines a discrete time step k to characterize the state changes during collaborative reasoning. At time k, the system collects an environmental state set E(k), which includes at least the online state of the task / agent, the validity of dependencies, and the update identifier of the reasoning target. The environmental state can be obtained through agent reporting, monitoring module collection, or push from the upper-level control unit. This application does not limit the method of environmental state collection; the focus is on using E(k) to determine the validity of the closure structure.

[0075] To map the environmental state to the structural model, the system defines a task state function: in Indicates task At any moment Can be used in reasoning; Indicates task At any moment Unable to participate in reasoning (e.g., agent offline, task execution failure, or resource unreachable). Based on this, construct the current set of valid tasks: And based on this, update the current set of valid dependencies: This leads to the construction of a dynamic dependency graph: The technical significance of the above process lies in: explicitly reflecting dynamic environmental changes in the reasoning structure diagram, so that subsequent validity judgments are based on the current real structure rather than the initial static structure.

[0076] Based on this, the system defines a closure validity determination function: in Represents the current set of closures. This represents the current environment state. The physical meaning of this function is to determine whether the current closure still satisfies the reasoning feasibility constraint in a dynamic environment. To ensure the feasibility and verifiability of the validity determination, this application will... The criteria for determining validity are broken down into two categories of necessary conditions: closure member availability and closure structure feasibility. Specifically, a closure is considered valid if and only if the following conditions are met: (1) Closure member availability condition: ; (2) Feasibility conditions for closure structure: .

[0077] Where F(·) is the reasoning feasibility determination function defined in the aforementioned closure determination mechanism, and its determination is based on the current valid structure graph. Corresponding sub Figure 1 Consistency and goal attainability. If either of the above conditions is not met, a decision is made. This indicates that the closure has failed.

[0078] When the system determines that the closure is valid, that is... If the current set of closures remains unchanged, the monitoring cycle continues; if the system determines that a closure is invalid, then... If this occurs, the closure re-identification process is triggered. After triggering, the system uses the currently valid task set... Current valid dependency set And taking the inference target G as input, the closure determination algorithm Closure_Refutation_Model(·) is called to generate a new set of closures: And The closure state continues into the validity maintenance process at the next moment. To uniformly characterize this dynamic maintenance process, this application provides closure state transition rules: when hour, ; when hour, .

[0079] The above rules form a decision function The event-triggered closed-loop control logic enables the closure structure to be updated adaptively as the environment changes, while avoiding repeated closure recognition calculations when the closure is still valid.

[0080] From the perspective of information flow, the data flow of this maintenance mechanism is: E(k) → State(ti, k) → → → The control flow is as follows: when V=false, closure re-identification and update are triggered. When V=true, the closure is maintained and monitoring continues. By separating the data flow from the control flow, this application decouples the dynamic maintenance module from the closure determination module, enabling the system to have both continuous operation capability and maintain the independence and reusability of the closure recognition logic.

[0081] Through the aforementioned dynamic validity maintenance mechanism for closures, this application achieves continuous consistency assurance of closure structure in a dynamic environment of multi-agent collaborative reasoning: closures remain unchanged when the environment is stable to reduce unnecessary reconstruction; when key members fail or dependent structures break, causing reasoning to be invalid, the system can promptly trigger re-identification and output new feasible closures, thereby ensuring that the closure result always matches the current reasoning structure.

[0082] like Figure 3 As shown, the overall process of this solution includes: Graph Model Construction: Construct a collaborative reasoning dependency graph model, abstracting multi-agent collaborative tasks into a directed structural graph, and formally expressing the task set and dependencies; Closure candidate generation: Generate a closure candidate set based on the reachability relations of the graph structure, filter out tasks that have no structural connection with the reasoning target, and reduce the recognition search space; Necessity check: Introduce a necessity check mechanism based on proof by contradiction, and determine the structural necessity of the task by performing a deletion test on a subset; Minimal complete convergence: The closure boundary is determined according to the minimum completeness convergence rule, and the final closure set that satisfies the structural feasibility and minimumity constraints is generated; Dynamic validity maintenance: Perform closure validity determination when the environment changes, and trigger the re-identification process when the closure fails.

[0083] The core of this method lies in establishing a structural determination function to determine whether the reasoning is still valid, and identifying the smallest indivisible structural set through a proof-by-contrast deletion mechanism, thereby avoiding indirect inference based on performance metrics.

[0084] In its implementation, this invention provides a reasoning feasibility closure recognition system for multi-agent collaborative reasoning scenarios. For ease of implementation and description, the system's functional structure is divided into a collaborative reasoning structure perception module, a reasoning feasibility closure determination module, and a reasoning closure validity maintenance module. The system forms a closed-loop workflow through these three modules, enabling the generation of closure candidates for collaborative reasoning tasks, the regularization of closure boundaries, and the management of closure validity in dynamic scenarios. The reasoning feasibility closure determination module is the core module, the collaborative reasoning structure perception module is the supporting module, and the reasoning closure validity maintenance module is an optional extension module.

[0085] Compared with existing methods based on performance optimization or critical path analysis, the technical solution of this application has the following technical features: First, it establishes a logical feasibility model based on whether the reasoning is still valid, avoiding indirect inference based on performance indicators; Second, it identifies the structural necessity of tasks through a proof-of-contradiction deletion mechanism, rather than screening through weight ranking or priority scoring; Third, it introduces a minimum completeness convergence rule to ensure that the closure set simultaneously satisfies structural feasibility and minimumity constraints; Fourth, it constructs a dynamic state transition model to achieve continuous validity maintenance of closures under environmental changes.

[0086] Through the above technical solution, this application realizes a complete closed-loop mechanism from "task set structure input" to "closure structure recognition output" and then to "dynamic maintenance and update", providing a structured recognition method for multi-agent collaborative reasoning systems that is independent of specific scheduling strategies and resource allocation methods.

[0087] like Figure 7 As shown, this embodiment constructs a multi-vehicle collaborative inference scenario to illustrate the deployment location and usage of the inference feasibility closure recognition system of this application in a real multi-agent collaborative inference system. This embodiment does not involve performance comparison, but is only used to clarify the working position, input and output objects, and closed-loop operation process of the technical solution of this application in the system.

[0088] In this embodiment, the multi-agent layer includes ramp vehicle agent A, main road preceding vehicle agent B, main road following vehicle agent C, roadside unit (RSU) agent D, and map / event agent E. Each agent periodically reports status / observation information to the edge computing node. The vehicle agent reports status / observation information such as position, speed, acceleration, lane position, and relative distance; the RSU agent reports roadside observation information; and the map / event agent reports event and map information. This information serves as input for collaborative reasoning, supporting the generation of reasoning objectives and structural modeling for the current decision-making cycle.

[0089] This embodiment employs a multi-MEC joint deployment approach, including at least two edge computing nodes: "ramp-side MEC - ramp" and "main road-side MEC - main road." Both nodes deploy and run the inference feasibility closure identification system described in this application, and synchronize information through a collaborative / synchronous link to ensure consistency in cross-regional collaborative inference. Ramp vehicle agent A prioritizes uploading state / observation information to the ramp MEC; main road vehicle agents B and C, as well as RSU agent D and map / event agent E, prioritize uploading state / observation information to the main road MEC. Through this proximity-based access method, a structured inference input is formed, primarily based on the local MEC.

[0090] In each decision cycle, the MEC-ramp and MEC-main road construct the inference task structure for the current cycle based on the received state / observation information, forming a task set, dependencies, and structured input corresponding to the inference target. They then execute the closure confirmation process of this application at their respective nodes, outputting the closure result. The closure result characterizes the minimum necessary task structure set required to ensure the inference target is valid at the structural level under the current inference target constraint. The coordination / synchronization link between the two MECs is used to synchronize the necessary structural information related to this closure identification, enabling the closure result to adapt to cross-MEC collaborative inference scenarios.

[0091] In this embodiment, the closure result is sent as a structured output to the upper-level control / scheduling module. The upper-level control / scheduling module is an existing module used to generate control commands based on the structural constraints of the received closure result and the states / observations of multiple vehicles, and then outputs the control commands accordingly. , , The corresponding vehicle actuators A, B, and C execute the commands. After completing the execution of the control commands, each vehicle actuator generates an execution receipt and environmental status information, forming an environmental state E(k). E(k) is fed back to the MEC side to trigger closure state updates and closed-loop operation.

[0092] During operation, when changes in vehicle status, roadside observations, or event / map information cause the current closure state to no longer meet the conditions for establishing the inference objective, the system of this application triggers a closure re-identification and update process based on the environmental state E(k). Structural consistency is maintained between the MEC-ramp and MEC-main road via a cooperative / synchronous link, thereby updating the new closure result and sending it again to the upper-level control / scheduling module. This closed-loop mechanism ensures that the closure result remains valid despite environmental changes, and the closure identification system always operates within the MEC edge node. The upper-level control / scheduling module only acts as a consumer of the closure result; this application does not limit its specific implementation method.

[0093] As can be seen from the above implementation methods, the working position of the technical solution of this application in the multi-vehicle collaborative merging scenario is as follows: it is deployed inside the edge collaborative reasoning platform (MEC), performs closure confirmation on the structural input formed by the state / observation reported by multiple agents and outputs the closure result; the closure result is provided as a structured constraint to the upper-layer control / scheduling module for control command generation; the vehicle execution receipt forms the environmental state E(k) and feeds it back to the MEC to trigger the closure update, thereby constituting a closed-loop collaborative reasoning process of multi-agent—multi-MEC—control execution.

[0094] Existing technologies also suffer from the problems of indivisible minimum structure being unidentifiable and unmaintainable: In multi-agent collaborative reasoning, there is an objective set of indivisible tasks. Removing any task from this set may cause the reasoning objective to fail. However, existing critical path analysis, task partitioning, or priority ranking methods are difficult to provide strict boundaries and computable identification rules for this set. At the same time, under dynamic environments and multi-MEC joint deployment conditions, there is a lack of closed-loop triggering and consistency maintenance mechanisms for structural changes, which can easily lead to inconsistencies between local pruning and global reasoning objectives. This solution can achieve minimum completeness convergence and maintain the validity of the reasoning closure.

[0095] This invention also provides a feasibility closure identification system for multi-agent collaborative reasoning. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0096] The system adopts a layered architecture, including a collaborative reasoning structure perception module, a reasoning feasibility closure determination module, and a reasoning closure validity maintenance module. These three modules form a progressive architecture of "structure modeling layer - logic judgment layer - dynamic maintenance layer," which is used to complete the entire processing flow from reasoning structure abstraction and closure identification to dynamic maintenance.

[0097] The collaborative reasoning structure awareness module, belonging to the structure modeling layer, receives task set information, inter-task dependencies, and inference target description information from the multi-agent collaborative reasoning scenario, and formalizes them into a computable graph structure model. The module's inputs include a task set T, a dependency set D, and an inference target G. Task set T represents the set of all subtasks participating in the current inference target, dependency set D represents the logical dependencies between tasks, and inference target G represents the final inference output node. The module outputs a collaborative reasoning dependency graph. In this graph, nodes represent tasks, and edges represent directed dependencies. This graph structure is used to characterize the logical propagation direction of the inference chain and provides a structural foundation for subsequent closure identification. This module only performs structural abstraction and reachability marking, and does not involve task execution costs, resource consumption, or scheduling priority information.

[0098] The reasoning feasibility closure determination module belongs to the logic decision layer and is the core improvement of this application. This module receives the collaborative reasoning dependency graph. And using the inference goal G as input, a closure candidate set is generated through reachability filtering. Furthermore, it performs a necessity test by contradiction and a minimum completeness convergence rule based on the candidate set. The output of this module is a set of closures that satisfy both structural feasibility and minimumity constraints. Specifically, this module constructs a reasoning feasibility judgment function F(S,G) to determine whether the reasoning objective G still holds true given only a subset of tasks S. It identifies the structural necessity of each task in the candidate set by performing a deletion test. Finally, it verifies the closure set through a minimum constraint to ensure it cannot be further divided. This module does not rely on task priority ranking or critical path length analysis, but completes closure identification entirely based on logical feasibility judgment.

[0099] The inference closure validity maintenance module belongs to the dynamic maintenance layer. It continuously monitors changes in the environment state during system operation and determines the validity of the current closure structure. The set of environment states is denoted as E(k), where k represents the discrete time step. Environment states include information such as changes in the task's online state, updates to dependencies, and changes in the inference target. This module defines the closure validity determination function. This is used to determine whether the current closure still satisfies the structural feasibility constraints under the current environmental conditions. When the determination result is valid, the system maintains the current closure structure; when the determination result is invalid, the closure re-identification process is triggered, that is, the inference feasibility closure determination module is called again to generate a new set of closures. Through the aforementioned state transition rules, the system forms a dynamic closed-loop mechanism, enabling the expansion from static closure identification to continuous structure maintenance.

[0100] At the system information flow level, the data flow path is: task set T and dependency relationship D are input into the structure awareness module to construct... ; The module determines the closure by inputting the inference target G and outputting the closure set. ; The structural constraints for system operation are output to the upper-level scheduling or control module. The control flow path is: environment state E(k) input validity maintenance module; if the closure is determined to be invalid, the closure re-identification process is triggered; if the closure is valid, the current state is maintained. The data flow and control flow are logically independent but interconnected, forming a dual-path structure.

[0101] The architecture has the following structural features: First, the module boundaries are clear, and the responsibilities of structural modeling, logical judgment and dynamic maintenance are separated; Second, the closure identification logic is decoupled from resource scheduling or execution control, and can be deployed independently on edge nodes or central control units; Third, an explicit state transition mechanism is introduced to enable the system to run continuously; Fourth, the closure output results are unique and verifiable, satisfying the structural feasibility and minimum constraints.

[0102] Through the above-mentioned overall system architecture design, this application constructs a structure recognition system with logical feasibility determination as its core, realizing a complete processing system from collaborative reasoning structure abstraction to closure recognition and then to dynamic maintenance and updating, providing a structured input basis for subsequent scheduling strategies, resource orchestration or control decisions.

[0103] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned feasibility closure identification method for multi-agent cooperative reasoning. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0104] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0105] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0106] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feasibility closure identification method for multi-agent cooperative reasoning, characterized in that, The steps of the method include: Locate the scene area where the target task is located, obtain the reasoning tasks of all agents in the scene area, and construct a first set of reasoning tasks; Based on the dependencies between the reasoning tasks in the first set of reasoning tasks, each reasoning task is used as a node to construct a first relational graph, and the target task is located in the first relational graph. Based on the first relationship graph, the feasibility closure of the target task is determined, and the closure in the first relationship graph is determined. The step of determining the closure of the reasoning feasibility of the target task includes performing a proof-by-contrast necessity determination based on the position of the target task in the first relation graph, and determining the closure in the first relation graph through the proof-by-contrast necessity determination.

2. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 1, characterized in that, The method further includes encapsulating and uploading the closure set constructed from the closures in the first relationship graph, so that the agent can complete the target task based on the closures in the closure set.

3. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 1 or 2, characterized in that, The step of determining the inference feasibility closure of the target task includes a reverse extension determination. In the reverse extension determination step, the node where the target task is located is taken as the root node for reverse extension. Whenever the extension reaches a node, the inference task corresponding to that node is added to the result set, and the first relationship graph is updated based on the result set.

4. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 3, characterized in that, In the step of extending backward from the node where the target task is located as the root node, and adding the inference task corresponding to the node to the result set whenever the extension reaches a node, the node where the target task is located is used as the root node for the backward extension. Whenever the extension reaches a node, the inference task corresponding to the node is obtained, and the inference task is retrieved in the result set. If the inference task does not exist in the result set, the inference task is added to the result set. If the reasoning task already exists in the result set, then add the reasoning task to the result set.

5. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 3, characterized in that, In the step of performing a proof-of-contrast necessity determination based on the position of the target task in the first relation graph, the starting node and the dependent nodes adjacent to the node corresponding to the target task in the first relation graph are located. For any node in the first relation graph that is not a starting node and is not a dependent node, a second relation graph is constructed. Based on the second relation graph, it is determined whether the node currently being tried to be removed is a necessary node, and attempts are made to remove each node that is not a starting node and is not a dependent node.

6. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 5, characterized in that, In the steps of removing any non-starting node and non-dependent node from the first relational graph, constructing a second relational graph, determining whether the node to be removed is a necessary node based on the second relational graph, and attempting to remove each non-starting node and non-dependent node, if there is no complete path from the starting node to the dependent node in the second relational graph, then the node to be removed is a necessary node. For removing another non-starting node and non-dependent node from the first relational graph, the second relational graph is reconstructed, and it is determined again whether the node is a necessary node. If every dependent node in the second relational graph has a complete path from the starting node to the dependent node, then the node currently being tried to remove is a non-essential node. In the current second relational graph, we further try to remove a node that is neither a starting node nor a dependent node, and construct a new second relational graph, until all nodes that are neither starting nodes nor dependent nodes have been traversed. The first or second relational graph after traversing all nodes that are neither starting nodes nor dependent nodes is the third relational graph.

7. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 6, characterized in that, Based on the first relation graph, the feasibility closure of the reasoning for the target task is determined. The step of determining the closure in the first relation graph further includes minimum completeness convergence. In the step of minimum completeness convergence, the reasoning task corresponding to the node in the third relation graph obtained from the first relation graph is taken as the closure, and a closure set is constructed.

8. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 6, characterized in that, The method further includes maintaining the validity of inference closures. In the step of maintaining the validity of inference closures, changes in the scene region are collected at preset intervals. If there is an increase or decrease in inference tasks in the scene region, the closure set is adjusted and maintained.

9. The feasibility closure identification method for multi-agent cooperative reasoning according to claim 8, characterized in that, The steps for maintaining the validity of the reasoning closure also include adding or removing nodes in the third relation graph based on the added or removed reasoning tasks, maintaining the third relation graph using the reverse extension judgment and the proof by contradiction necessity judgment, and obtaining the maintained closure set.

10. A feasibility closure recognition system for multi-agent cooperative reasoning, characterized in that, The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method as described in any one of claims 1 to 9.

Images