A multi-agent based multi-modal coupling co-evolution simulation method

Abstract

The application discloses a kind of based on multi-agent's multimodal coupling collaborative evolution simulation method, comprising the following steps: S1, obtains multimodal record and carries out pretreatment;S2, process fragment and candidate model are matched, and model capability matrix is constructed;S3, with the results of trial calculation as the adaptation value, adjacent connection relationship is combined, model access path is determined, and is instantiated as model agent;S4, data transmission relationship and feedback back relationship are combined screening, and data transmission edge and feedback back edge are established, and agent interaction atlas is constructed;S5, model agent is scheduled in round, and the evolution state of model agent is updated, and the scheduling state is revised, and the evolution state of this round is obtained;S6, connection residual calculation is carried out, and collaborative evolution simulation result is iteratively scheduled and output.The application can realize multi-model collaborative coupling and path adaptive arrangement, improve simulation continuity and system expansibility.

Description

Technical Field

[0001] This invention relates to the field of environmental simulation technology, and in particular to a multi-agent-based multimodal coupled cooperative evolution simulation method. Background Technology

[0002] With the increasing demand for environmental simulation and complex system extrapolation, many mature models and systems have emerged in existing technologies. For example, the Earth System Model CAS-ESM can be used for multi-sphere Earth system simulation, CLM is mainly used for land surface process simulation, WRF is mostly used for meteorological field simulation, EFDC is mostly used for surface hydrodynamics and water quality simulation, MODFLOW is mostly used for groundwater flow and pollution migration simulation, CMAQ is mostly used for air quality and pollutant diffusion simulation, and TIMES is mostly used for energy system and carbon emission simulation. These models have good applicability in their respective fields and can complete simulation analysis under single media, single process, or specific scenarios.

[0003] However, when dealing with comprehensive simulation tasks involving multiple processes and responses, existing models typically participate in the computation either independently or through manual concatenation. Differences exist in input / output formats, variable semantics, execution steps, and invocation methods between different models. The output of one model often requires manual processing before it can be used as input for the next. The lack of automatic matching and trial-and-error verification mechanisms between models leads to model selection relying primarily on empirical judgment, making it difficult to automatically determine the appropriate model for a specific simulation task.

[0004] Existing technologies also include some couplers or integration platforms, such as UEC couplers designed for specific application scenarios, which can achieve data exchange and joint simulation between models within a certain range. However, this type of coupling method usually has strong scenario limitations, and the coupling path and model interaction relationship are mostly pre-set. When the simulation task changes, the number of models increases, or the feedback relationship becomes more complex, the existing coupling method is difficult to reorganize the model access path, and it is also difficult to uniformly handle the forward data transmission relationship and the feedback feedback relationship.

[0005] Furthermore, existing multi-model collaborative simulation methods typically focus more on the assembly or presentation of model results, lacking dynamic scheduling and iterative correction mechanisms oriented towards the operational process. When discrepancies arise between the simulation results of a certain round and subsequent processes, existing methods struggle to promptly correct model access paths and model interaction relationships. This causes the multi-model coupling process to easily remain at the linear call level, failing to form a stable collaborative evolution simulation process. Existing technologies suffer from at least the following problems: insufficient adaptability between candidate models and task processes; reliance on manual configuration of model connections; difficulty in unified modeling of data transmission and feedback; and a lack of iterative correction based on operational results during the simulation process. These issues affect the continuity, scalability, and collaborative evolution effects of complex simulation tasks.

[0006] Therefore, how to provide a multi-agent-based multimodal coupled cooperative evolution simulation method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a multi-agent-based multimodal coupled cooperative evolution simulation method. This invention fully utilizes dual-tower semantic encoding, Viterbi path search, K-optimal path search, gated state recursion, and Bayesian confidence update techniques. It details the cooperative evolution simulation process of arranging multimodal records into task evolution chains, constructing a model capability matrix based on trial calculations, instantiating candidate models into model agents, and performing round-robin scheduling and feedback correction based on interaction graphs. It has the advantages of accurate model adaptation, clear coupling paths, continuous scheduling process, and strong feedback correction capability.

[0008] A multi-agent-based multimodal coupled cooperative evolution simulation method according to an embodiment of the present invention includes the following steps: S1. Obtain the multimodal records of the task to be simulated and preprocess them to form a task evolution chain; S2. A dual-tower semantic encoder is used to perform connection matching between process segments and candidate models in the task evolution chain, and connection trial calculation is performed on the candidate models that pass the matching to construct the model capability matrix. S3. Using the results of the acceptance trial calculation in the model capability matrix as the adaptation value, and combining the adjacent acceptance relationship of the task evolution chain, the Viterbi path search is used to determine the model access path, and the candidate models in the path are instantiated as model agents. S4. Map the model access path to the preset interaction graph, use K-optimal path search to combine and filter the data transmission relationship and feedback return relationship between model agents, and establish data transmission edge and feedback return edge according to the filtering results to construct the agent interaction graph. S5. The model agents are scheduled in rounds according to the agent interaction graph. The gated state recursion method is used to update the evolution state of the model agents along the data transmission edge. The scheduling state of the associated model agents is corrected according to the feedback transmission edge to obtain the evolution state of this round. S6. Calculate the residuals for the current evolutionary state, and use the Bayesian confidence update method to correct the model access path and agent interaction graph, and iteratively schedule and output the collaborative evolution simulation results.

[0009] Optionally, the preprocessing specifically includes: segmenting the multimodal records of the task to be simulated into blocks and extracting fields, and aligning the time indices of different record blocks using a dynamic time warping method, wherein: The multimodal record is: heterogeneous data record called, transmitted and received by different component models in the simulation task, including original input record, model interface record, model running record and intermediate response record; Based on the field call relationship, the aligned record blocks are labeled with the succession mark, and the labeled record blocks are arranged in a chain according to the sequential call relationship to form a task evolution chain.

[0010] Optionally, S2 specifically includes: S21. Sequentially extract adjacent process segments from the task evolution chain, combine the output of the previous process segment and the input of the next process segment into a process transition text, and combine the input and output of the candidate model into a model transition text, wherein: The candidate model is a pre-accessed component model. After receiving the input content, the candidate model performs local simulation calculations and generates output content for the subsequent process segment to continue. S22. A dual-tower semantic encoder is used to encode the process acceptance text and the model acceptance text respectively to obtain the process acceptance vector and the model acceptance vector, and the similarity between the process acceptance vector and the model acceptance vector is calculated. S23. Connect candidate models whose similarity reaches the preset condition to the corresponding adjacent process segments, and use the output content of the previous process segment to drive the candidate models to perform succession calculations, thereby obtaining the calculation output content of the candidate models, wherein: The trial calculation involves transcribing the output of the previous process segment into input that the candidate model can accept, calling the candidate model to perform a trial run according to a preset running step size, and extracting the output of the candidate model at the end of the trial run as the trial calculation output. S24. Compare the trial output with the input of the next process segment. Based on the similarity and comparison results, determine the candidate model's ability to connect with adjacent process segments and write the ability value into the model capability matrix.

[0011] Optionally, S22 specifically includes: S221. Perform text segmentation on the process text and the model text respectively, map the segmentation results into word vector sequences, and superimpose positional encoding on the word vector sequences to obtain the process text embedding sequence and the model text embedding sequence. S222. The process text is embedded into the sequence input of the process encoding tower of the dual-tower semantic encoder, and the model text is embedded into the sequence input of the model encoding tower of the dual-tower semantic encoder, wherein the process encoding tower and the model encoding tower adopt the Transformer encoding structure with independent parameters and consistent output dimensions. S223. Perform self-attention encoding on the process text embedding sequence through the process encoding tower to obtain process context features, and perform self-attention encoding on the model text embedding sequence through the model encoding tower to obtain model context features; S224. Perform pooling projection on the process context features and model context features respectively to obtain the process connection vector and the model connection vector, and use cosine similarity to calculate the connection similarity between the process connection vector and the model connection vector.

[0012] Optionally, S3 specifically includes: S31. Read each process segment in the order of the task evolution chain, retrieve the acceptance trial calculation results of the corresponding process segment of the candidate model in the model capability matrix, and determine the adaptation value of the candidate model in the corresponding process segment based on the acceptance trial calculation results. S32. Using the order of process segments in the task evolution chain as the recursive direction, the adaptation value of the candidate model in the current process segment is recursively combined with the cumulative adaptation value retained in the previous process segment, and candidate model combinations that do not meet the continuous succession are pruned according to the succession relationship of adjacent process segments. S33. In each process segment, retain the model connection relationship with the highest cumulative adaptation value and record the corresponding source of the previous process segment, until the end process segment of the task evolution chain is reached. S34. Starting from the end process segment of the task evolution chain, backtrack along the recorded source to determine the model access path covering the task evolution chain, and instantiate the candidate model into the model agent of the corresponding process segment according to the model access path, wherein: The model agent is a schedulable execution object of the candidate model on the corresponding process segment. It receives the preceding content of the corresponding process segment, performs local calculations, sends the calculation results back to the subsequent process segments of the task evolution chain, and records its own recursive state for use in the agent interaction graph.

[0013] Optionally, S32 specifically includes: S321. Construct a transition layer between the previous process segment and the current process segment, pair the candidate model recursive state of the previous process segment with the candidate model adaptation value of the current process segment, and generate a transition gate value based on the transition relationship between adjacent process segments. S322. For pairs that pass the acceptance gate value, perform Viterbi recursion. Based on the recursion status of the previous process segment, the fitting value of the current process segment, and the acceptance gate value, calculate the current recursion score. Stop recursion for pairs that fail the acceptance gate value. S323. For multiple current recursive scores corresponding to the same current candidate model, the bundle search pruning method is used to screen the recursive source and establish a source index between the current candidate model and the candidate model of the previous process segment. S324. In the current process segment, organize the current recursive scores and source indices of each candidate model into the current recursive layer, and pass the current recursive layer to the next process segment to continue the Viterbi recursion.

[0014] Optionally, S4 specifically includes: S41. According to the process segments corresponding to each model agent in the model access path, locate the corresponding graph nodes in the preset interaction graph and map the model agents to the corresponding graph nodes to obtain candidate interaction subgraphs. S42. In the candidate interaction subgraph, the graph edges extending in the positive direction along the task evolution chain are set as candidate data transmission edges, the graph edges pointing back to the mapped graph nodes are set as candidate feedback return edges, and the cost of the candidate edges is calculated based on the trial calculation results and the process crossing relationship. S43. Using model access path coverage, continuity of connection and feedback reachability as search constraints, the K-optimal path search method is used to screen candidate path combinations in the candidate interaction subgraph and eliminate candidate path combinations that do not satisfy the connection relationship. S44. Confirm the candidate data transmission edges in the retained candidate path combinations as data transmission edges, confirm the candidate feedback transmission edges in the retained candidate path combinations as feedback transmission edges, and construct the agent interaction graph based on the data transmission edges and feedback transmission edges.

[0015] Optionally, S43 specifically includes: S431. Using the model agent mapping node in the model access path as the constraint node, expand the candidate interaction subgraph into a hierarchical search graph according to the acceptance order of the task evolution chain, and map the candidate data transmission edge and the candidate feedback feedback edge as inter-layer acceptance edge and feedback feedback edge, respectively. S432. In the hierarchical search graph, the K-optimal path search method is used to generate the baseline path, and the candidate path is expanded with the constraint node as the deviation node. Candidate paths that bypass the constraint node, repeatedly pass through the same model agent, or disrupt the succession order are pruned. S433. Perform multi-objective ranking on the pruned candidate paths. The ranking criteria are determined by candidate edge cost, continuity of connection and feedback reachability. In the ranking results, retain the combination of non-dominated paths that meets the preset number. S434. In the combination of non-dominated paths, the relationships of inter-layer receiving edges and feedback edges are merged. Repeated receiving and feedback relationships between agents of the same model are merged to obtain the data transmission relationship and feedback return relationship used to construct the agent interaction graph.

[0016] Optionally, S5 specifically includes: S51. Perform topological layering based on the pointing relationship of data transmission edges in the agent interaction graph, classify the model agents whose previous data transmission edges have completed recursion into the current scheduling round, and determine the scheduling order of the current scheduling round according to the hierarchical order of the edges. S52. For the model agent in the current scheduling round, map the evolution state of its predecessor model agent to the recursive input state along the data transmission edge, and combine the recursive input state with the previous evolution state of the model agent to form a gated recursive state. S53. Using the gated state recursion method, the update gate and retention gate are calculated for the gated recursion state. The recursion input state is introduced according to the update gate, and the previous round of evolution state is retained according to the retention gate. The current model agent's evolution state to be executed is obtained, and the current model agent is driven to complete the local evolution operation in this round. S54. Calculate the feedback gating value based on the feedback back-propagation edge for the local evolution operation results of this round, correct the scheduling state of the associated model agents according to the feedback gating value, and merge the evolution states of each model agent in rounds to obtain the evolution state of this round.

[0017] Optionally, S6 specifically includes: S61. According to the adjacent process segments in the task evolution chain, in the current evolution state, the output content of each model agent and the received content of the successor model agent are compared and the acceptance residual is calculated in combination with the scheduling state of the data transmission edge and the feedback transmission edge. S62. Match the received residual with the preset residual interval to determine the confidence level of this round of scheduling, and use the confidence level as the observation result to enter the Bayesian confidence update process to perform posterior correction on the adaptation value in the model access path and the edge weight in the agent interaction graph. S63. Adjust the retention state of the corresponding model agent in the model access path according to the corrected adaptation value, and adjust the data transmission edge and feedback transmission edge in the agent interaction graph according to the corrected edge weight to obtain the corrected model access path and agent interaction graph. S64. Based on the corrected model access path and agent interaction graph, continue to execute round-by-round scheduling. When the residuals of consecutive rounds meet the convergence condition, output the evolution state of the last round as the result of the co-evolution simulation.

[0018] Compared with the prior art, the beneficial effects of the present invention are: First, existing multi-model simulation methods typically suffer from problems such as inconsistent interfaces between heterogeneous models, difficulty in quantifying model capabilities, and reliance on human experience for model selection. This invention addresses these issues by orchestrating the task evolution chain of multimodal records and employing a dual-tower semantic encoder and a process acceptance calculation to construct a model capability matrix. This allows the process acceptance capabilities of candidate models to be calculated and verified, avoiding reliance on experience alone and improving the accuracy of model selection and process acceptance.

[0019] Secondly, existing model coupling methods mostly employ fixed serial connections or static configurations, which are difficult to adapt to the data transfer and feedback requirements between multiple models in complex simulation tasks. This invention determines model access paths through Viterbi path search, instantiates candidate models as model agents, and then constructs an agent interaction graph by combining a pre-defined interaction graph and K-optimal path search. This transforms the coupling relationship between models from a static configuration to a dynamic organization based on inheritance relationships and path combinations, improving the flexibility and scalability of model collaborative invocation.

[0020] Finally, existing simulation systems typically lack a continuous feedback correction mechanism during operation, making it difficult to adjust the model links in a timely manner once they are determined. This invention uses gated state recursion for round-robin scheduling and updates and corrects the model access path and agent interaction graph based on the inherited residuals and Bayesian confidence, enabling the multimodal coupled simulation process to have continuous iteration and adaptive correction capabilities, thereby improving the continuity, stability, and reliability of the results of the co-evolution simulation. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a multi-agent-based multimodal coupling cooperative evolution simulation method proposed in this invention; Figure 2 This invention proposes a dual-tower semantic encoding for matching and model capability matrix construction based on a multi-agent multimodal coupling cooperative evolution simulation method. Figure 3 This invention provides the output and iterative correction of the cooperative evolution simulation results of a multi-agent-based multimodal coupling cooperative evolution simulation method. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3A multi-agent-based multimodal coupled cooperative evolution simulation method includes the following steps: S1. Obtain the multimodal records of the task to be simulated and preprocess them to form a task evolution chain; S2. A dual-tower semantic encoder is used to perform connection matching between process segments and candidate models in the task evolution chain, and connection trial calculation is performed on the candidate models that pass the matching to construct the model capability matrix. S3. Using the results of the acceptance trial calculation in the model capability matrix as the adaptation value, and combining the adjacent acceptance relationship of the task evolution chain, the Viterbi path search is used to determine the model access path, and the candidate models in the path are instantiated as model agents. S4. Map the model access path to the preset interaction graph, use K-optimal path search to combine and filter the data transmission relationship and feedback return relationship between model agents, and establish data transmission edge and feedback return edge according to the filtering results to construct the agent interaction graph. S5. The model agents are scheduled in rounds according to the agent interaction graph. The gated state recursion method is used to update the evolution state of the model agents along the data transmission edge. The scheduling state of the associated model agents is corrected according to the feedback transmission edge to obtain the evolution state of this round. S6. Calculate the residuals for the current evolutionary state, and use the Bayesian confidence update method to correct the model access path and agent interaction graph, and iteratively schedule and output the collaborative evolution simulation results.

[0024] In this embodiment, preprocessing specifically includes: segmenting the multimodal records of the task to be simulated into blocks and extracting fields; and aligning the time indices of different record blocks using a dynamic time warping method, wherein: Multimodal records are heterogeneous data records called, transmitted, and received by different component models in the task to be simulated, including original input records, model interface records, model running records, and intermediate response records; Based on the field calling relationship, the aligned record blocks are labeled with the succession mark, and the labeled record blocks are arranged in a chain according to the sequential calling relationship to form a task evolution chain; In this embodiment, the multimodal records of the task to be simulated are initially sorted according to the call timestamp and the source of the record, and are divided into blocks with 8 to 24 scheduling time steps as a record window, with 20% to 40% overlap between adjacent record windows. After extracting fields for each record block, the field name, field value, field source, and field destination are encoded into a 256-dimensional field vector, and abnormal record blocks with a missing field ratio greater than 30% are deleted. Subsequently, a dynamic time warping method was used to align the time indexes of different record blocks. The search bandwidth of dynamic time warping was set to 10% to 25% of the record window length, and the alignment cost was composed of field vector similarity and time interval difference. After completing the time index alignment, the record blocks are marked according to the calling relationship between the field source and the field destination, and the record blocks with no less than 2 consecutive calls are retained as valid process segments; The process segments are arranged in a chain according to their order of invocation, forming a task evolution chain containing 4 to 15 process segments.

[0025] In this embodiment, S2 specifically includes: S21. Sequentially extract adjacent process segments from the task evolution chain, combine the output of the previous process segment and the input of the next process segment into a process transition text, and combine the input and output of the candidate model into a model transition text, wherein: The candidate model is a pre-connected component model. After receiving the input content, the candidate model performs local simulation calculations and generates output content for the subsequent process segment to continue. S22. A dual-tower semantic encoder is used to encode the process acceptance text and the model acceptance text respectively to obtain the process acceptance vector and the model acceptance vector, and the similarity between the process acceptance vector and the model acceptance vector is calculated. S23. Connect candidate models whose similarity reaches the preset condition to the corresponding adjacent process segments, and use the output content of the previous process segment to drive the candidate models to perform succession calculations, thereby obtaining the calculation output content of the candidate models, wherein: The trial calculation is performed by transcribing the output of the previous process segment into input that the candidate model can accept, calling the candidate model to perform the trial run according to the preset running step size, and extracting the output of the candidate model at the end of the trial run as the trial calculation output. S24. Compare the trial output with the input of the next process segment. Based on the similarity and comparison results, determine the candidate model's ability to connect with adjacent process segments and write the ability value into the model capability matrix.

[0026] In this embodiment, S22 specifically includes: S221. Perform text segmentation on the process text and the model text respectively, map the segmentation results into word vector sequences, and superimpose positional encoding on the word vector sequences to obtain the process text embedding sequence and the model text embedding sequence. S222. The process text is embedded into the sequence input of the process encoding tower of the dual-tower semantic encoder, and the model text is embedded into the sequence input of the model encoding tower of the dual-tower semantic encoder, wherein the process encoding tower and the model encoding tower adopt the Transformer encoding structure with independent parameters and consistent output dimensions. S223. Perform self-attention encoding on the process text embedding sequence through the process encoding tower to obtain process context features, and perform self-attention encoding on the model text embedding sequence through the model encoding tower to obtain model context features; S224. Perform pooling projection on the process context features and model context features respectively to obtain the process connection vector and the model connection vector, and use cosine similarity to calculate the connection similarity between the process connection vector and the model connection vector. In this embodiment, before the process receiving text and the model receiving text enter the dual-tower semantic encoder, the text is first segmented according to the field separator, the call identifier and the process boundary identifier. The segmented text fragments are mapped to word vector sequences. The word vector dimension is set to 256 to 768 dimensions, and position encoding is superimposed on the word vector sequence so that the field order and process order in the same receiving text are preserved. Both the process coding tower and the model coding tower use 3 to 6 Transformer coding layers, with 4 to 8 attention heads set in each Transformer coding layer. The network parameters of the two coding towers are independent of each other, and the output dimensions are consistent. During the encoding process, the process encoding tower performs self-attention encoding on the content related to the preceding and following processes in the process text embedding sequence, and the model encoding tower performs self-attention encoding on the content connecting the input and output in the model text embedding sequence, thereby obtaining process context features and model context features respectively. Average pooling and linear projection are performed on the process context features and model context features to obtain process and model inheritance vectors with dimensions ranging from 256 to 1024. The system uses cosine similarity to calculate the similarity between the two models. When the similarity reaches the preset range of 0.65 to 0.85, the corresponding candidate model is marked as a matching model and enters the subsequent connection trial calculation.

[0027] In this embodiment, S3 specifically includes: S31. Read each process segment in the order of the task evolution chain, retrieve the acceptance trial calculation results of the corresponding process segment of the candidate model in the model capability matrix, and determine the adaptation value of the candidate model in the corresponding process segment based on the acceptance trial calculation results. S32. Using the order of process segments in the task evolution chain as the recursive direction, the adaptation value of the candidate model in the current process segment is recursively combined with the cumulative adaptation value retained in the previous process segment, and candidate model combinations that do not meet the continuous succession are pruned according to the succession relationship of adjacent process segments. S33. In each process segment, retain the model connection relationship with the highest cumulative adaptation value and record the corresponding source of the previous process segment, until the end process segment of the task evolution chain is reached. S34. Starting from the end process segment of the task evolution chain, backtrack along the recorded source to determine the model access path covering the task evolution chain, and instantiate the candidate model into the model agent of the corresponding process segment according to the model access path, wherein: The model agent is a schedulable execution object of the candidate model on the corresponding process segment. It receives the preceding content of the corresponding process segment, performs local calculations, returns the calculation results to the subsequent process segments of the task evolution chain, and records its own recursive state for use in the agent interaction graph.

[0028] In this embodiment, S32 specifically includes: S321. Construct a transition layer between the previous process segment and the current process segment, pair the candidate model recursive state of the previous process segment with the candidate model adaptation value of the current process segment, and generate a transition gate value based on the transition relationship between adjacent process segments. S322. For pairs that pass the acceptance gate value, perform Viterbi recursion. Based on the recursion status of the previous process segment, the fitting value of the current process segment, and the acceptance gate value, calculate the current recursion score. Stop recursion for pairs that fail the acceptance gate value. S323. For multiple current recursive scores corresponding to the same current candidate model, the bundle search pruning method is used to screen the recursive source and establish a source index between the current candidate model and the candidate model of the previous process segment. S324. In the current process segment, organize the current recursive scores and source indices of each candidate model into the current recursive layer, and pass the current recursive layer to the next process segment to continue the Viterbi recursion.

[0029] In this embodiment, S4 specifically includes: S41. According to the process segments corresponding to each model agent in the model access path, locate the corresponding graph nodes in the preset interaction graph and map the model agents to the corresponding graph nodes to obtain candidate interaction subgraphs. S42. In the candidate interaction subgraph, graph edges extending in the positive direction along the task evolution chain are designated as candidate data transfer edges, and graph edges pointing back to mapped graph nodes are designated as candidate feedback return edges. The cost of the candidate edges is calculated based on the trial calculation results and the process crossing relationship, specifically including: The model access path is expanded into a hierarchical sequence according to the process of the task evolution chain. The graph nodes in the candidate interaction subgraph are mapped to the corresponding levels, and the connection direction and cross-level span of the graph edge are determined according to the hierarchical changes of the nodes at both ends of the graph edge. The graph edges whose receiving direction increases along the hierarchical sequence are marked as candidate data transmission edges, the graph edges whose receiving direction points back to the mapped level are marked as candidate feedback transmission edges, and graph edges that do not fall within the receiving range of the model access path are removed. Based on the acceptance trial calculation results of the model agents at both ends of the candidate edge, the semantic consistency verification and difference measurement are performed on the trial calculation output of the starting model agent and the acceptance input of the ending model agent to obtain the acceptance credibility of the candidate edge. Candidate edge costs are generated based on acceptance credibility, acceptance direction, and cross-layer span, and corresponding cost labels are assigned to candidate data transmission edges and candidate feedback return edges respectively. S43. Using model access path coverage, continuity of connection and feedback reachability as search constraints, the K-optimal path search method is used to screen candidate path combinations in the candidate interaction subgraph and eliminate candidate path combinations that do not satisfy the connection relationship. S44. Confirm the candidate data transmission edges in the retained candidate path combinations as data transmission edges, confirm the candidate feedback transmission edges in the retained candidate path combinations as feedback transmission edges, and construct the agent interaction graph based on the data transmission edges and feedback transmission edges.

[0030] In this embodiment, S43 specifically includes: S431. Using the model agent mapping node in the model access path as the constraint node, expand the candidate interaction subgraph into a hierarchical search graph according to the acceptance order of the task evolution chain, and map the candidate data transmission edge and the candidate feedback feedback edge as inter-layer acceptance edge and feedback feedback edge, respectively. S432. In the hierarchical search graph, the K-optimal path search method is used to generate the baseline path, and the candidate path is expanded with the constraint node as the deviation node. Candidate paths that bypass the constraint node, repeatedly pass through the same model agent, or disrupt the succession order are pruned. S433. Perform multi-objective ranking on the pruned candidate paths. The ranking criteria are determined by candidate edge cost, continuity of connection and feedback reachability. In the ranking results, retain the combination of non-dominated paths that meets the preset number. S434. In the combination of non-dominated paths, the relationships of inter-layer receiving edges and feedback edges are merged. Repeated receiving and feedback relationships between agents of the same model are merged to obtain the data transmission relationship and feedback return relationship used to construct the agent interaction graph.

[0031] In this embodiment, S5 specifically includes: S51. Perform topological layering based on the pointing relationship of data transmission edges in the agent interaction graph, classify the model agents whose previous data transmission edges have completed recursion into the current scheduling round, and determine the scheduling order of the current scheduling round according to the hierarchical order of the edges. S52. For the model agent in the current scheduling round, map the evolution state of its predecessor model agent to the recursive input state along the data transmission edge, and combine the recursive input state with the previous evolution state of the model agent to form a gated recursive state. S53. Using the gated state recursion method, the update gate and retention gate are calculated for the gated recursion state. The recursion input state is introduced according to the update gate, and the previous round of evolution state is retained according to the retention gate. The current model agent's evolution state to be executed is obtained, and the current model agent is driven to complete the local evolution operation in this round. S54. Calculate the feedback gating value based on the feedback back-propagation edge for the local evolution operation results of this round, correct the scheduling state of the associated model agents according to the feedback gating value, and merge the evolution states of each model agent in rounds to obtain the evolution state of this round.

[0032] In this embodiment, S6 specifically includes: S61. According to the adjacent process segments in the task evolution chain, in the current evolution state, the output content of each model agent and the received content of the successor model agent are compared and the acceptance residual is calculated in combination with the scheduling state of the data transmission edge and the feedback transmission edge. S62. Match the received residual with the preset residual interval to determine the confidence level of this round of scheduling, and use the confidence level as the observation result to enter the Bayesian confidence update process to perform posterior correction on the adaptation value in the model access path and the edge weight in the agent interaction graph. S63. Adjust the retention state of the corresponding model agent in the model access path according to the corrected adaptation value, and adjust the data transmission edge and feedback transmission edge in the agent interaction graph according to the corrected edge weight to obtain the corrected model access path and agent interaction graph. S64. Based on the corrected model access path and agent interaction graph, continue to execute round-by-round scheduling. When the residuals of consecutive rounds meet the convergence condition, output the evolution state of the last round as the result of the co-evolution simulation.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-process collaborative simulation scenario in a watershed. This scenario involves multiple continuous processes, including meteorological driving forces, hydrodynamic and water quality processes, subsurface migration, soil transformation, and ecological responses. Traditionally, technicians determine the order of component model calls based on experience and then manually transfer the outputs of different models. Because the inputs, outputs, and step sizes of different models differ, the output of one model often cannot be directly adopted by the next. Simultaneously, intermediate results from ecological responses, soil conditions, and water quality changes can negatively impact the scheduling status of related models. Traditional linear concatenation methods struggle to reflect these feedback relationships in a timely manner, easily leading to process breakpoints, repeated model calls, and fluctuations in simulation results across multiple rounds.

[0034] In this embodiment, the system received a total of 4620 multimodal records, of which original input records accounted for 43.5%, model interface records for 11.6%, model running records for 27.8%, and intermediate response records for 17.1%. The system first preprocessed these records, dividing consecutive records into 15 record blocks, each covering 12 scheduling time steps, with adjacent record blocks retaining an overlap of 4 scheduling time steps. After dynamic time warping and alignment, the system identified 10 valid process segments and formed a task evolution chain according to field call relationships. When manually sorting the records, there were 7 breakpoints in the same batch of records; after processing using this invention, only 1 weakly connected segment was retained, and this weakly connected segment was subsequently taken over by a substitute model in the model access path search.

[0035] Subsequently, the system performed connection matching on 22 candidate component models. Both the process connection vector and model connection vector output by the dual-tower semantic encoder were represented using 512 dimensions, with a connection similarity threshold set to 0.74. After semantic matching, 15 candidate models met the threshold condition. The system further performed connection trial calculations on the matched candidate models. For each process segment, three sets of probe inputs were selected for short-range execution, and the trial calculation output was compared with the input content of the next process segment. After the trial calculations, 41 effective connection relationships were formed in the model capability matrix, including 20 high-connection-capability relationships, 14 medium-connection-capability relationships, and 7 low-connection-capability relationships. Relationships below the connection threshold were not included in the subsequent model access path search.

[0036] During the model integration phase, the system uses the adaptation values ​​from the model capability matrix and the adjacent adaptation relationships in the task evolution chain to determine the model integration path using Viterbi path search. After recursion and backtracking, the system forms a model integration path covering 10 process segments, in which 8 candidate models are instantiated into 10 model agents. Compared with the manual configuration method, the manual method has 3 instances of duplicate model calls, 2 instances of model adaptation gaps, and 4 instances of manual transcription nodes in the 10 process segments; the model integration path obtained by this invention does not have any adaptation gaps, and the average path adaptation value is improved from 0.66 to 0.84.

[0037] During the interaction graph construction phase, the system maps model access paths to a preset interaction graph and uses K-optimal path search to combine and filter data transmission and feedback relationships between model agents, where K is set to 5. After filtering, 17 data transmission edges and 6 feedback edges are formed. Traditional methods typically only retain positive sequential relationships and fail to identify feedback relationships; this implementation uses feedback edges to transmit ecological response results, soil state results, and water quality response results back to the associated model agents, enabling subsequent scheduling to be corrected based on intermediate response states. Statistics show that 4 out of the 6 feedback edges triggered scheduling state corrections during round-robin scheduling.

[0038] During the round-robin scheduling phase, the system performs 13 rounds of collaborative evolution simulation according to the agent interaction graph, and updates the evolutionary state of the model agents using a gated state recursion method. Rounds 1 to 4 mainly complete forward acceptance, rounds 5 to 9 introduce feedback correction, and the acceptance residual tends to stabilize after round 10. The average acceptance residual in the initial rounds is 0.192, which decreases to 0.108 in round 6 and to 0.064 in round 10, and finally remains below 0.060 for three consecutive rounds. In contrast, the average acceptance residual of the traditional manual configuration method is 0.223, and there is still fluctuation above 0.15 after round 10. In terms of the complete simulation time, the traditional method takes about 168 minutes, while this invention takes about 104 minutes; the number of model acceptance failures decreases from 18 to 5; the scheduling waiting rate decreases from 33.2% to 19.4%; and the continuity score of the simulation results improves from 76.8 to 92.7.

[0039] To verify the applicability of this invention to the collaborative simulation of natural processes and social response processes, the same method was applied to a natural-social collaborative simulation scenario in a certain region. In this scenario, the energy consumption, emissions, and response data generated by the social activity component model need to be continuously integrated with the meteorological, pollution diffusion, and ecological response models. The system received 3860 multimodal records, forming 8 process segments, and configured 16 candidate models. After dual-tower integration matching and integration trial calculations, 29 sets of effective integration relationships were formed; after Viterbi path search, 6 types of candidate models were instantiated into 8 model agents; after K-optimal path search, 12 data transmission edges and 4 feedback return edges were formed. After 10 rounds of phased scheduling, the average integration residual decreased from the initial 0.174 to 0.057, the number of model path adjustments decreased from 9 under manual configuration to 3, and the complete simulation time decreased from 126 minutes to 81 minutes.

[0040] To verify the scalability of this invention, a new candidate model for water ecological response was added to the aforementioned multi-process collaborative simulation scenario in the watershed. Traditional methods require reconfiguring input / output paths and manually adjusting existing call relationships, resulting in 7 connection failures during testing and a total connection verification time of approximately 18 hours. Using this invention, the system incorporates the new candidate model into the connection matching process, achieving a connection similarity of 0.79 through dual-tower semantic encoding. After 3 sets of probe input connection trials, 4 valid connection relationships were added to the model capability matrix. Subsequently, the system instantiated the new candidate model as a model agent and integrated it into the agent interaction graph through Viterbi path search and K-optimal path search. After the new model was integrated, the number of data transmission edges increased from 17 to 19, and the number of feedback edges increased from 6 to 7. There were 2 connection failures, and the total connection verification time was approximately 6.5 hours.

[0041] The above results show that the present invention can reduce the connection breaks caused by manual configuration and improve the continuity of the model access path in simulation scenarios with complex multimodal recording, a large number of candidate models, and coexistence of forward data transmission and feedback transmission relationships. Furthermore, it enhances the stability of multi-model collaborative evolution simulation through agent interaction graphs and round-by-round state recursion.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-agent-based multimodal coupled cooperative evolution simulation method, characterized in that, Includes the following steps: S1. Obtain the multimodal records of the task to be simulated and preprocess them to form a task evolution chain; S2. A dual-tower semantic encoder is used to perform connection matching between process segments and candidate models in the task evolution chain, and connection trial calculation is performed on the candidate models that pass the matching to construct the model capability matrix. S3. Using the results of the acceptance trial calculation in the model capability matrix as the adaptation value, and combining the adjacent acceptance relationship of the task evolution chain, the Viterbi path search is used to determine the model access path, and the candidate models in the path are instantiated as model agents. S4. Map the model access path to the preset interaction graph, use K-optimal path search to combine and filter the data transmission relationship and feedback return relationship between model agents, and establish data transmission edge and feedback return edge according to the filtering results to construct the agent interaction graph. S5. The model agents are scheduled in rounds according to the agent interaction graph. The gated state recursion method is used to update the evolution state of the model agents along the data transmission edge. The scheduling state of the associated model agents is corrected according to the feedback transmission edge to obtain the evolution state of this round. S6. Calculate the residuals for the current evolutionary state, and use the Bayesian confidence update method to correct the model access path and agent interaction graph, and iteratively schedule and output the collaborative evolution simulation results.

2. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 1, characterized in that, The preprocessing specifically includes: segmenting the multimodal records of the task to be simulated into blocks and extracting fields; aligning the time indices of different record blocks using a dynamic time warping method, wherein: The multimodal record refers to the heterogeneous data record called, transmitted, and received by different component models in the simulation task, including the original input record, model interface record, model running record, and intermediate response record; Based on the field call relationship, the aligned record blocks are labeled with the succession mark, and the labeled record blocks are arranged in a chain according to the sequential call relationship to form a task evolution chain.

3. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 1, characterized in that, S2 specifically includes: S21. Sequentially extract adjacent process segments from the task evolution chain, combine the output of the previous process segment and the input of the next process segment into a process transition text, and combine the input and output of the candidate model into a model transition text, wherein: The candidate model is a pre-accessed component model. After receiving the input content, the candidate model performs local simulation calculations and generates output content for the subsequent process segment to continue. S22. A dual-tower semantic encoder is used to encode the process acceptance text and the model acceptance text respectively to obtain the process acceptance vector and the model acceptance vector, and the similarity between the process acceptance vector and the model acceptance vector is calculated. S23. Connect candidate models whose similarity reaches the preset condition to the corresponding adjacent process segments, and use the output content of the previous process segment to drive the candidate models to perform succession calculations, thereby obtaining the calculation output content of the candidate models, wherein: The trial calculation involves transcribing the output of the previous process segment into input that the candidate model can accept, calling the candidate model to perform a trial run according to a preset running step size, and extracting the output of the candidate model at the end of the trial run as the trial calculation output. S24. Compare the trial output with the input of the next process segment. Based on the similarity and comparison results, determine the candidate model's ability to connect with adjacent process segments and write the ability value into the model capability matrix.

4. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 3, characterized in that, S22 specifically includes: S221. Perform text segmentation on the process text and the model text respectively, map the segmentation results into word vector sequences, and superimpose positional encoding on the word vector sequences to obtain the process text embedding sequence and the model text embedding sequence. S222. The process text is embedded into the sequence input of the process encoding tower of the dual-tower semantic encoder, and the model text is embedded into the sequence input of the model encoding tower of the dual-tower semantic encoder, wherein the process encoding tower and the model encoding tower adopt the Transformer encoding structure with independent parameters and consistent output dimensions. S223. Perform self-attention encoding on the process text embedding sequence through the process encoding tower to obtain process context features, and perform self-attention encoding on the model text embedding sequence through the model encoding tower to obtain model context features; S224. Perform pooling projection on the process context features and model context features respectively to obtain the process connection vector and the model connection vector, and use cosine similarity to calculate the connection similarity between the process connection vector and the model connection vector.

5. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 1, characterized in that, S3 specifically includes: S31. Read each process segment in the order of the task evolution chain, retrieve the acceptance trial calculation results of the corresponding process segment of the candidate model in the model capability matrix, and determine the adaptation value of the candidate model in the corresponding process segment based on the acceptance trial calculation results. S32. Using the order of process segments in the task evolution chain as the recursive direction, the adaptation value of the candidate model in the current process segment is recursively combined with the cumulative adaptation value retained in the previous process segment, and candidate model combinations that do not meet the continuous succession are pruned according to the succession relationship of adjacent process segments. S33. In each process segment, retain the model connection relationship with the highest cumulative adaptation value and record the corresponding source of the previous process segment, until the end process segment of the task evolution chain is reached. S34. Starting from the end process segment of the task evolution chain, backtrack along the recorded source to determine the model access path covering the task evolution chain, and instantiate the candidate model into the model agent of the corresponding process segment according to the model access path, wherein: The model agent is a schedulable execution object of the candidate model on the corresponding process segment. It receives the preceding content of the corresponding process segment, performs local calculations, sends the calculation results back to the subsequent process segments of the task evolution chain, and records its own recursive state for use in the agent interaction graph.

6. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 5, characterized in that, Specifically, S32 includes: S321. Construct a transition layer between the previous process segment and the current process segment, pair the candidate model recursive state of the previous process segment with the candidate model adaptation value of the current process segment, and generate a transition gate value based on the transition relationship between adjacent process segments. S322. For pairs that pass the acceptance gate value, perform Viterbi recursion. Based on the recursion status of the previous process segment, the fitting value of the current process segment, and the acceptance gate value, calculate the current recursion score. Stop recursion for pairs that fail the acceptance gate value. S323. For multiple current recursive scores corresponding to the same current candidate model, the bundle search pruning method is used to screen the recursive source and establish a source index between the current candidate model and the candidate model of the previous process segment. S324. In the current process segment, organize the current recursive scores and source indices of each candidate model into the current recursive layer, and pass the current recursive layer to the next process segment to continue the Viterbi recursion.

7. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 1, characterized in that, S4 specifically includes: S41. According to the process segments corresponding to each model agent in the model access path, locate the corresponding graph nodes in the preset interaction graph and map the model agents to the corresponding graph nodes to obtain candidate interaction subgraphs. S42. In the candidate interaction subgraph, the graph edges extending in the positive direction along the task evolution chain are set as candidate data transmission edges, the graph edges pointing back to the mapped graph nodes are set as candidate feedback return edges, and the cost of the candidate edges is calculated based on the trial calculation results and the process crossing relationship. S43. Using model access path coverage, continuity of connection and feedback reachability as search constraints, the K-optimal path search method is used to screen candidate path combinations in the candidate interaction subgraph and eliminate candidate path combinations that do not satisfy the connection relationship. S44. Confirm the candidate data transmission edges in the retained candidate path combinations as data transmission edges, confirm the candidate feedback transmission edges in the retained candidate path combinations as feedback transmission edges, and construct the agent interaction graph based on the data transmission edges and feedback transmission edges.

8. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 7, characterized in that, Specifically, S43 includes: S431. Using the model agent mapping node in the model access path as the constraint node, expand the candidate interaction subgraph into a hierarchical search graph according to the acceptance order of the task evolution chain, and map the candidate data transmission edge and the candidate feedback feedback edge as inter-layer acceptance edge and feedback feedback edge, respectively. S432. In the hierarchical search graph, the K-optimal path search method is used to generate the baseline path, and the candidate path is expanded with the constraint node as the deviation node. Candidate paths that bypass the constraint node, repeatedly pass through the same model agent, or disrupt the succession order are pruned. S433. Perform multi-objective ranking on the pruned candidate paths. The ranking criteria are determined by candidate edge cost, continuity of connection and feedback reachability. In the ranking results, retain the combination of non-dominated paths that meets the preset number. S434. In the combination of non-dominated paths, the relationships of inter-layer receiving edges and feedback edges are merged. Repeated receiving and feedback relationships between agents of the same model are merged to obtain the data transmission relationship and feedback return relationship used to construct the agent interaction graph.

9. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 1, characterized in that, S5 specifically includes: S51. Perform topological layering based on the pointing relationship of data transmission edges in the agent interaction graph, classify the model agents whose previous data transmission edges have completed recursion into the current scheduling round, and determine the scheduling order of the current scheduling round according to the hierarchical order of the edges. S52. For the model agent in the current scheduling round, map the evolution state of its predecessor model agent to the recursive input state along the data transmission edge, and combine the recursive input state with the previous evolution state of the model agent to form a gated recursive state. S53. Using the gated state recursion method, the update gate and retention gate are calculated for the gated recursion state. The recursion input state is introduced according to the update gate, and the previous round of evolution state is retained according to the retention gate. The current model agent's evolution state to be executed is obtained, and the current model agent is driven to complete the local evolution operation in this round. S54. Calculate the feedback gating value based on the feedback back-propagation edge for the local evolution operation results of this round, correct the scheduling state of the associated model agents according to the feedback gating value, and merge the evolution states of each model agent in rounds to obtain the evolution state of this round.

10. The multi-agent-based multimodal coupled cooperative evolution simulation method according to claim 1, characterized in that, S6 specifically includes: S61. According to the adjacent process segments in the task evolution chain, in the current evolution state, the output content of each model agent and the received content of the successor model agent are compared and the acceptance residual is calculated in combination with the scheduling state of the data transmission edge and the feedback transmission edge. S62. Match the received residual with the preset residual interval to determine the confidence level of this round of scheduling, and use the confidence level as the observation result to enter the Bayesian confidence update process to perform posterior correction on the adaptation value in the model access path and the edge weight in the agent interaction graph. S63. Adjust the retention state of the corresponding model agent in the model access path according to the corrected adaptation value, and adjust the data transmission edge and feedback transmission edge in the agent interaction graph according to the corrected edge weight to obtain the corrected model access path and agent interaction graph. S64. Based on the corrected model access path and agent interaction graph, continue to execute round-by-round scheduling. When the residuals of consecutive rounds meet the convergence condition, output the evolution state of the last round as the result of the co-evolution simulation.

Images