Methods, devices, electronic equipment and storage media for complex event evolution simulation

Abstract

This invention provides a method, apparatus, electronic device, and storage medium for complex event evolution deduction. It generates a macro-planning tree based on scene description information, containing temporal stages and macro-objectives. Under the constraint of the macro-objective at the current time step, it drives multiple agents to generate micro-level actions, thereby determining candidate states for the next moment. It extracts micro-evolutionary trends and calculates a consistency metric between these trends and the macro-objective. When the consistency metric is low, it triggers dynamic adjustments to actions or plans to obtain the next state of the target. Finally, it updates the deduction results of the complex event, constructing a complete closed-loop deduction framework that unifies macro-planning, micro-behavior generation, state transition synthesis, and value backtracking arbitration. This significantly reduces deviations caused by the emergence of multi-agent behaviors in long-chain deductions and improves the credibility, stability, and interpretability of deduction results in open environments.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for deducing the evolution of complex events. Background Technology

[0002] With the increasing capabilities of Large Language Models (LLMs) in natural language reasoning, planning, and generation, their application in complex system event extrapolation, strategy simulation, and risk evolution analysis is becoming increasingly widespread. These applications, such as public safety situation assessment, social system evolution, and supply chain impact simulation, typically require models to simulate and predict complex events with multi-agent, nonlinear, and long-chain characteristics in open and dynamic environments, using multiple steps and paths. Currently, the main technical approach in related fields is to directly utilize LLMs or combine them with traditional planning algorithms for event extrapolation. LLM-based extrapolation methods, by inputting a scenario description, allow the model to autonomously generate subsequent event development sequences.

[0003] However, in the process of extrapolation, when the micro-behavioral trends deviate from the macro-planning, the extrapolation may proceed along an invalid or low-quality path, consuming computational resources without obtaining feasible extrapolation results. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for complex event evolution deduction, which addresses the shortcomings of existing technologies where, when micro-behavioral trends deviate from macro-planning, the deduction may proceed along an invalid or low-quality path, failing to yield feasible deduction results.

[0005] This invention provides a method for deducing the evolution of complex events, comprising: Based on the scenario description information, a macro-planning tree is generated that includes time-series stages and macro-objectives; Under the constraints of the macro-objective at the current time step, drive multiple agents to generate micro-behavioral actions, and determine the candidate state for the next time step based on the micro-behavioral actions and the current state; Based on the candidate state extraction of the next time step, the micro-evolution trend is obtained, and the consistency metric between the micro-evolution trend and the macro-objective of the current time step is calculated. If the consistency metric is less than the preset consistency threshold, the micro-behavioral action or the macro-planning tree is adjusted, and the candidate state for the next time step is re-acquired to calculate the consistency metric until the consistency metric is not less than the preset consistency threshold, and the target state for the next time step is obtained. Based on the next state of the target, update the deduction results of complex events.

[0006] According to a method for complex event evolution deduction provided by the present invention, the step of triggering adjustments to the micro-level behavioral actions or the macro-level planning tree when the consistency metric value is less than a preset consistency threshold includes: If the consistency metric is less than the preset consistency threshold, the arbitration level is determined based on the conflict intensity and micro-confidence of the micro-evolutionary trend. According to the arbitration level, resampling signals are sent to the multi-agent to adjust the micro-behavioral actions or reconstruct the macro-planning tree.

[0007] According to a method for predicting the evolution of complex events provided by the present invention, determining the candidate state for the next moment based on the microscopic behavioral actions and the current state includes: The micro-behavioral actions and the current state are input into the large language model to obtain multiple preliminary candidate states output by the large language model; Extract the physical properties and numerical boundaries of each of the preliminary candidate states, and verify whether each of the preliminary candidate states satisfies the preset rules; the preset rules include conservation rules and boundary condition rules. The preliminary candidate states that satisfy the preset rules are determined as the candidate states for the next time step.

[0008] According to a method for extrapolating the evolution of complex events provided by the present invention, updating the extrapolation results of the complex event based on the next state of the target includes: Starting from the next state of the target, the evolution trajectory of multiple future time steps is predicted; Extract the multidimensional value of the evolutionary trajectory during the pre-simulation process; the multidimensional value includes at least two of the following: goal achievement degree, system risk entropy, and state stability parameter value; The value score of the evolutionary trajectory is obtained by weighting the multidimensional values. The tree node score is updated based on the value score. The tree node score is compared with a preset score threshold. If the tree node score is lower than the preset score threshold, the branch corresponding to the evolution trajectory is pruned. The state nodes corresponding to the branches that were not pruned are added as valid nodes to the inference results to complete the update.

[0009] According to a method for complex event evolution deduction provided by the present invention, the step of generating a macro-planning tree containing temporal stages and macro-objectives based on scenario description information includes: Analyze the scene description information and extract the set of key variables and global boundary conditions; At each time step, multiple candidate planning branches are generated, and for each candidate planning branch, the parameter values ​​of time series coverage, variable closure degree, and constraint satisfiability are extracted. The consistency score of each candidate planning branch is obtained by weighted summing of the parameter values ​​of the temporal coverage, variable closure, and constraint satisfiability of each candidate planning branch. A preset number of candidate planning branches with the highest consistency scores are retained to construct the macro-planning tree.

[0010] According to a method for complex event evolution deduction provided by the present invention, the step of driving multiple agents to generate microscopic behavioral actions under the constraint of the macroscopic objective at the current time step includes: Under the constraint of the macroscopic objective at the current time step, the same agent is sampled multiple times in parallel to obtain an initial action set containing multiple initial actions; Calculate the semantic similarity between each initial action, and cluster the initial actions in the initial action set based on the semantic similarity to obtain multiple action clusters; The initial action corresponding to the cluster center of the action cluster with the highest frequency of occurrence is selected as the candidate action, and if no logical mutual exclusion conflict is detected in the candidate action, the candidate action is used as the micro behavior action.

[0011] The present invention also provides a complex event evolution deduction device, comprising: The macro-planning unit generates a macro-planning tree that includes time-series stages and macro-objectives based on scenario description information; The micro-generation unit, under the constraint of the macro-objective at the current time step, drives multiple agents to generate micro-behavioral actions, and determines the candidate state for the next time step based on the micro-behavioral actions and the current state. The consistency measurement unit extracts the micro-evolution trend based on the candidate state at the next time step, and calculates the consistency measurement value between the micro-evolution trend and the macro-objective at the current time step. The adjustment unit, when the consistency metric value is less than the preset consistency threshold, triggers the adjustment of the micro-behavioral action or the macro-planning tree, and re-acquires the candidate state of the next moment to calculate the consistency metric value until the consistency metric value is not less than the preset consistency threshold, and obtains the target state of the next moment. The deduction unit updates the deduction results of complex events based on the next moment's state of the target.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the complex event evolution deduction method described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the complex event evolution deduction method as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the complex event evolution deduction method as described above.

[0015] The complex event evolution deduction method, apparatus, electronic device, and storage medium provided by this invention generate a macro-planning tree containing temporal stages and macro-objectives based on scene description information. Under the constraint of the macro-objective at the current time step, it drives multiple agents to generate micro-behavioral actions, thereby determining the candidate state for the next moment, extracting the micro-evolution trend and calculating its consistency metric with the macro-objective. When the consistency metric is low, it triggers dynamic adjustments to the actions or plans to obtain the state of the target at the next moment, and finally updates the deduction results of complex events. This effectively overcomes the problem of the lack of explicit environment models in the prior art, which makes direct deduction impossible. It constructs a complete closed-loop deduction framework that unifies macro-planning, micro-behavioral generation, state transition synthesis, and value backtracking arbitration. This greatly reduces the deviation caused by the emergence of multi-agent behaviors in long-chain deductions and improves the credibility, stability, and interpretability of the deduction results in an open environment. Attached Figure Description

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

[0017] Figure 1 This is a flowchart illustrating the complex event evolution deduction method provided by the present invention; Figure 2 This is a flowchart illustrating the conflict arbitration method provided by the present invention; Figure 3 This is a flowchart illustrating the state transition method provided by the present invention; Figure 4 This is a flowchart illustrating the complex event deduction method based on generative hierarchical Monte Carlo programming provided by the present invention. Figure 5This is a schematic diagram of the complex event deduction system based on generative hierarchical Monte Carlo programming provided by the present invention; Figure 6 This is a schematic diagram of the complex event evolution deduction device provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

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

[0019] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0020] To address the aforementioned problems, this invention provides a method for extrapolating the evolution of complex events, enabling interpretable, multi-path, and cross-scale consistent long-chain extrapolation in an open environment. Figure 1 This is a flowchart illustrating the complex event evolution deduction method provided by the present invention, as shown below. Figure 1 As shown, the method includes: Step 110: Based on the scenario description information, generate a macro-planning tree that includes time-series stages and macro-objectives.

[0021] Here, scenario description information refers to user-defined text data used to describe the initial state of a complex event, such as the event text, initial state, simulation period, and global boundary conditions such as the total resource limit. Temporal stages refer to the multiple evolutionary phases of a complex event divided chronologically during its evolution. Macro-objectives refer to the global expected results or target directions that the system needs to achieve within each temporal stage. Therefore, the macro-planning tree here refers to a structured tree model with temporal constraints and goal orientation, used to provide global skeletal constraints for subsequent simulations.

[0022] Specifically, first, the input scenario description information is obtained. Then, the scenario description information is input into a large language model, which is used for parsing to extract the set of key variables and the approximate stage divisions. A cluster search strategy is used to generate multiple candidate planning branches, and the optimal branch is retained based on the consistency score. This generates a macro-planning tree containing an ordered set of time-series stages and corresponding macro-objectives, providing a global skeleton constraint for the deduction process and preventing logical divergence over long time spans.

[0023] Step 120: Under the constraint of the macroscopic objective at the current time step, drive the multi-agent to generate microscopic behaviors and actions, and determine the candidate state for the next time step based on the microscopic behaviors and actions and the current state.

[0024] Here, the current time step refers to the specific time node currently in the simulation process. Here, multi-agent refers to the set of simulation agents representing different organizational or individual nodes, including macro-agents and micro-agents. Here, micro-behavioral actions refer to the specific operational instructions or behavioral actions generated by the multi-agents based on local perception and role settings. Additionally, the current state here refers to the state snapshot information containing all entity attribute values ​​at the current time step. Here, the candidate states for the next time step refer to the set of states that may transition to future time steps after executing micro-behavioral actions.

[0025] Specifically, at the current time step, based on the macro-objective corresponding to the macro-programming tree as a constraint, and combined with the current state and the roles of the multiple agents, prompts are constructed and injected into the context to drive the multiple agents to sample and generate micro-level actions in parallel. Here, the format of the prompts can be: [World State] + [Agent Role] + [Macro-Constraint] -> Generate Action.

[0026] Next, the micro-level actions and the current state are input into the large language model to synthesize preliminary candidate states for the next time step. It should be noted that in an environment without an explicit physics engine, the large language model is used as a probability generator to calculate state transitions, and abnormal states that violate physical or logical constraints are filtered out through hard constraint rules to finally determine the legal candidate states for the next time step.

[0027] It should be noted that by combining macro-planning with micro-behavior and achieving state transitions without an explicit environment model, the dependence of traditional deduction on explicit environment models is effectively overcome, ensuring the physical and logical accuracy of the generated candidate states.

[0028] Step 130: Based on the candidate state extraction of the next time step, the micro-evolution trend is obtained, and the consistency metric between the micro-evolution trend and the macro-objective of the current time step is calculated.

[0029] Here, the micro-evolutionary trend refers to the state change vector calculated based on the candidate state at the next time step relative to the current state, reflecting the direction of group evolution resulting from micro-behavior. The consistency metric is a quantitative indicator that measures the degree of alignment between the actual evolutionary direction caused by micro-behavior and the target direction expected by macro-planning; it can be calculated using methods such as cosine similarity.

[0030] Specifically, based on the state difference between the candidate state at the next moment and the current state, the state change vector is calculated as the micro-evolution trend. Then, the target direction expected by the current macro-planning can be extracted as the macro-target vector, and the cosine similarity between the micro-evolution trend and the macro-target vector can be calculated. This cosine similarity is used as a consistency metric to monitor the consistency between micro-emergence and the macro-preset trajectory in real time.

[0031] It should be noted that by introducing a consistency measurement mechanism, the deviation between the emerging results of the underlying micro-behavioral groups and the top-level macro-planning can be quantitatively evaluated in real time, providing an accurate basis for judgment in dynamic closed-loop control.

[0032] Step 140: If the consistency metric is less than the preset consistency threshold, the micro-behavioral action or the macro-planning tree is adjusted, and the candidate state for the next time step is re-acquired to calculate the consistency metric until the consistency metric is not less than the preset consistency threshold, and the target state for the next time step is obtained.

[0033] Here, the target next moment state refers to the legitimate next moment state that, after conflict arbitration and consistency verification, fully conforms to the system evolution logic and remains consistent with the macro-objective.

[0034] Specifically, first, the calculated consistency metric is compared with a preset consistency threshold. If the consistency metric is less than the preset consistency threshold, a conflict is determined to have occurred. At this point, a hierarchical arbitration mechanism is triggered based on the conflict intensity and micro-confidence of the micro-evolutionary trend. For example, when the conflict intensity is low, the macro-planning tree remains unchanged, and a rejection signal is issued to force the multi-agents to resample and generate corrected micro-behavioral actions; when the micro-evolutionary trend has strong inertia and the conflict intensity is high, the subsequent macro-planning tree is dynamically reconstructed.

[0035] Then, the candidate state for the next time step is resynthesized after adjustment and the consistency metric is calculated iteratively until the inconsistency is eliminated, so that the consistency metric reaches or exceeds the preset consistency threshold, and the state at this time is output as the target state for the next time step.

[0036] It should be noted that by dynamically coordinating macro-planning and micro-behavioral trends through the dynamic conflict arbitration mechanism, a dynamic balance between top-down control and bottom-up emergence is achieved, which significantly reduces the deviation and error accumulation caused by the emergence of multiple agent behaviors in long-chain simulations and ensures the stability of the simulations.

[0037] Step 150: Update the deduction results of complex events based on the next moment state of the target.

[0038] Here, complex events refer to dynamic evolutionary processes characterized by openness, nonlinearity, and feedback coupling, such as public safety incidents and the evolution of financial risks. Here, the deduction results refer to cognitive output data that includes multiple possible evolutionary paths, key turning points, and multi-dimensional value scores.

[0039] Specifically, the verified next-time state of the target is added as a valid node to the current evolutionary path, and rapid pre-simulation and multi-dimensional value assessment of future time steps are continued based on this next-time state. For example, improved upper confidence bound algorithms can be used to update path scores, prune low-value or high-risk branches, and continuously supplement the inference results of complex events with high-quality evolutionary paths and state nodes retained from the exploration, thereby continuously expanding and updating the inference results of complex events. For example, the inference results can be output in the form of an inference tree.

[0040] It should be noted that by constructing a closed-loop extrapolation framework that combines macro-planning guidance with micro-behavioral emergence, a macro-micro consistency measurement and a dynamic conflict arbitration mechanism are introduced into the state evolution process without an explicit environmental model. Based on this, the extrapolation results are continuously updated, effectively breaking through the dependence of traditional extrapolation methods on explicit environmental models and achieving high-fidelity long-chain extrapolation in an open system. Furthermore, the extrapolation results supported by rigorous quantitative evaluation and arbitration mechanisms provide a clear causal link and structured evolutionary path for the evolution of complex events, significantly enhancing the interpretability, verifiability, and decision-making ability of the final extrapolation results.

[0041] The method provided in this invention generates a macro-planning tree containing temporal stages and macro-objectives based on scene description information. Under the constraint of the macro-objective at the current time step, it drives multiple agents to generate micro-behavioral actions, thereby determining the candidate state for the next moment. It extracts the micro-evolutionary trend and calculates the consistency metric between it and the macro-objective. When the consistency metric is low, it triggers dynamic adjustments to the actions or plans to obtain the state of the target at the next moment. Finally, it updates the deduction results of complex events. This method effectively overcomes the problem of the lack of explicit environment models in the prior art, which makes direct deduction impossible. It constructs a complete closed-loop deduction framework that unifies macro-planning, micro-behavioral generation, state transition synthesis, and value backtracking arbitration. This greatly reduces the deviation caused by the emergence of multi-agent behaviors in long-chain deductions and improves the credibility, stability, and interpretability of deduction results in open environments.

[0042] Based on any of the above embodiments, in step 140, if the consistency metric value is less than a preset consistency threshold, triggering an adjustment to the micro-behavioral action or the macro-planning tree includes: If the consistency metric is less than the preset consistency threshold, the arbitration level is determined based on the conflict intensity and micro-confidence of the micro-evolutionary trend. According to the arbitration level, resampling signals are sent to the multi-agent to adjust the micro-behavioral actions or reconstruct the macro-planning tree.

[0043] Specifically, in step one, if the consistency metric is less than a preset consistency threshold, the arbitration level is determined based on the conflict intensity and micro-confidence level of the micro-evolutionary trend. Conflict intensity refers to the severity of the deviation of micro-evolution from the expected macro-evolutionary direction. Micro-confidence level refers to the systematic inertia or deterministic probability exhibited by the micro-group behavior evolution. Arbitration level refers to the different levels of response strategies set to resolve the inconsistency between macro and micro levels.

[0044] In detail, firstly, a consistency metric is calculated, such as cosine similarity (Sim). The formula is the dot product of the micro-evolutionary trend vector and the macro-target vector, divided by the product of their magnitudes. For example, it can be calculated using the following formula: ; In the formula, The cosine similarity between the micro-evolutionary trend and the macro-objective at the current time step is a consistency metric. Represents a vector indicating the micro-evolutionary trend; This represents the macroscopic target vector.

[0045] If the calculated consistency metric is less than a preset consistency threshold, a conflict is determined between the macro and micro levels. Subsequently, the conflict intensity and the micro-level confidence level of the micro-evolutionary trend are comprehensively evaluated to determine the arbitration level. For example, in a complex information dissemination and intervention scenario in the public safety field, at a certain time step, the evolution trend of the micro-agent shows an exponential increase in network heat index, while the current stage of the macro-programming tree requires a stable period. Calculations show that the consistency score between the two is only 0.4, less than the preset arbitration trigger threshold of 0.75. At this point, a conflict is successfully detected, and the conflict intensity and micro-level confidence level are further evaluated based on the momentum of the increasing heat, thereby accurately determining the current arbitration level.

[0046] It should be noted that by quantifying consistency and introducing a comprehensive assessment of conflict intensity and micro-confidence when conflict occurs, the underlying causes of inconsistencies can be accurately analyzed, providing a basis for taking appropriate intervention measures in the future.

[0047] Step two involves sending resampling signals to the multi-agent group according to the arbitration level to adjust micro-level behaviors or reconstruct the macro-level planning tree. Here, the resampling signal refers to the control signal sent to the lower-level multi-agent group, forcing them to discard the currently generated set of behaviors and regenerate new behavioral instructions. In other words, this reconstruction refers to the process of regenerating subsequent stage divisions and objective constraints based on the latest current state when the original top-level planning is deemed invalid.

[0048] In detail, the hierarchical arbitration logic is executed according to the determined arbitration level. If it is determined to be at the first level of arbitration, i.e., micro-correction, it indicates that the conflict intensity is low and the micro-confidence is low. It is regarded as a random perturbation. At this time, the macro-program remains unchanged, and a resampling signal, such as a Reject signal, is sent to the multi-agent to force the micro-agent to regenerate the action under the addition of punitive constraints.

[0049] If the arbitration level is determined to be Level 2, which is a macro-adjustment, it indicates that the conflict intensity is high and the micro-trend has strong inertia. Therefore, the original macro-plan is deemed to have failed, and the interface of the planning module is called, with the current state passed in as a new starting point for reconstruction.

[0050] In one embodiment, taking the public safety information dissemination scenario as an example, if the arbitrator determines that the micro-trend of the surge in popularity has strong inertia and reaches the level of secondary arbitration, then a macro-adjustment strategy is executed to postpone the stable decline period in the macro-planning to a later time step and dynamically insert a new macro-objective, such as forced noise reduction intervention. Subsequently, the micro-behavioral actions of the control node are regenerated based on the reconstructed macro-planning.

[0051] For example, in a distributed resource transmission network scenario, macro-planning restricts the load rate of a single link from being too high, but the micro-optimal solution requires a backup link to operate at full capacity. Assessment reveals that rigidly adhering to the macro-constraint will lead to network-wide paralysis. Therefore, the arbitrator chooses to reconstruct the macro-planning tree, temporarily relaxing the constraint but adding time-series restrictions, thereby reconstructing a feasible planning path under sudden disturbances.

[0052] It should be noted that different approaches are taken for different degrees of yaw, which can not only correct local random noise, but also adapt to the strong inertia of the group emergence trend, thus avoiding the problem of system collapse caused by rigid rule constraints in traditional simulations.

[0053] The method provided in this invention dynamically determines the arbitration level based on the conflict intensity and micro-confidence of the micro-evolutionary trend when the consistency metric value is low, and flexibly chooses to issue a resampling signal to correct micro-actions or directly reconstruct the macro-planning tree accordingly. This constructs a dynamic conflict arbitration mechanism, which effectively resolves the logical conflict between macro-planning and micro-behavioral emergence, significantly reduces the accumulation of chain errors in the deduction of long-chain complex systems, and enables the deduction process to achieve a robust dynamic balance between top-down control and bottom-up emergence.

[0054] In one embodiment, Figure 2 This is a flowchart illustrating the conflict arbitration method provided by the present invention, as shown below. Figure 2 As shown, the method includes: Specifically, the conflict arbitration method takes into account the state change vector calculated based on the candidate states at the next time step, i.e., the micro-evolutionary trend, and the target direction expected by the current macro-planning, i.e., the macro-target vector. First, it calculates the consistency metric between the micro-evolutionary trend and the macro-target, i.e., it calculates the cosine similarity. The system determines whether the calculated consistency metric, Sim, is less than a preset consistency threshold. If the consistency metric is not less than the threshold, the system determines that the micro-behavior follows the macro-planning, and the two are consistent, thus maintaining the original path.

[0055] If the consistency metric is less than the threshold, a conflict determination is triggered, further examining the micro-confidence of the micro-evolutionary trend and the system inertia (i.e., conflict intensity). The hierarchical arbitration logic is determined based on the conflict intensity and micro-confidence: when the determination is Level 1 (micro-correction), it indicates low conflict intensity and low micro-confidence, and is judged as a random disturbance. The action executed at this time is to maintain the macro-plan unchanged, sending a Reject signal downstream to force micro-agent resampling.

[0056] When it is determined to be Level 2 (macro adjustment), it indicates that the conflict intensity is high and the micro trend has strong inertia (system inertia). At this time, the macro planning is determined to be invalid. The executed Action is to call the interface of the initial planning module (S1), pass in the current state as the new starting point, and reconstruct the subsequent macro planning tree, such as modifying the target threshold and extending the stage time. Here, S1 refers to step 110, generating a macro planning tree containing time sequence stages and macro goals based on the scenario description information.

[0057] When the condition is determined to be Level 3 (path bifurcation), it indicates an irreconcilable conflict that cannot be easily corrected. The action executed in this case is to generate a bifurcation node on the deduction tree, preserving two parallel paths: one of forced control and the other of adapting to the loss of control, thus creating a parallel world branch. The final output is an arbitration command. It includes commands such as Continue, Resample_Micro, Replan_Macro, and Fork.

[0058] Based on any of the above embodiments, in step 120, determining the candidate state for the next moment based on the microscopic behavior and the current state includes: The micro-behavioral actions and the current state are input into the large language model to obtain multiple preliminary candidate states output by the large language model; Extract the physical properties and numerical boundaries of each of the preliminary candidate states, and verify whether each of the preliminary candidate states satisfies the preset rules; the preset rules include conservation rules and boundary condition rules. The preliminary candidate states that satisfy the preset rules are determined as the candidate states for the next time step.

[0059] Specifically, in step one, the micro-level actions and the current state are input into the large language model to obtain multiple preliminary candidate states output by the large language model. Here, preliminary candidate states refer to the set of possible states for a future moment that the large language model initially infers based on probability distributions without an explicit physics engine.

[0060] In practice, the current state is obtained, such as a snapshot of the state containing all entity attribute values. Additionally, a set of candidate micro-level actions generated by the multi-agent system is acquired. Then, the micro-level actions and the current state are input into a large language model, which is used as the state transition function for generative transfer computation. Specifically, the state transition function can be expressed as: ; In the formula, This represents any initial candidate state output by the large language model; Representing a large language model; This represents the current state, that is, the state at time t. This represents microscopic actions or behaviors. It's understandable that by inputting the microscopic actions and the current state into this function, N possible next-time states are generated. These are used as multiple initial candidate states, which can be denoted as... .

[0061] It should be noted that by using a generative large language model, a broad and reliable state evolution space can be constructed in an environment without a physics engine. This eliminates the need for traditional numerical differential equation solvers and breaks through the strong dependence of traditional deduction methods on explicit environment models.

[0062] Step two involves extracting the physical attributes and numerical boundaries of each preliminary candidate state, and verifying whether each preliminary candidate state satisfies the preset rules. Here, the preset rules include conservation rules and boundary condition rules. These preset rules refer to inviolable objective laws or hard constraints in the real world or specific business scenarios. Physical attributes typically involve indicators such as entity inflow, outflow, storage capacity, or network topology connections; the numerical boundaries involve thresholds such as upper and lower limits of entity-related indicators.

[0063] In practical implementation, each generated preliminary candidate state is traversed, and hard constraint filters are established for verification against inviolable rules such as resource conservation, topological connectivity, and numerical boundaries. For example, algebraic verification is performed to validate conservation rules, and the calculation formula can be: ; That is, total input Subtract total output Equal to the change in storage To verify conservation.

[0064] Simultaneously, boundary checks are performed to verify boundary condition rules. For example, verifying whether key variables meet the requirements. .

[0065] In one embodiment, considering the topology evolution and robustness optimization scenario of a distributed resource transmission network, when a micro-proxy request from a node increases the transmission volume to another core transmission node by 50%, the storage capacity of that core transmission node and its preset capacity limit are extracted from the preliminary candidate states. Hard constraint checks are performed using boundary condition rules. If it is found that the request would cause the storage capacity of the core transmission node to overflow, i.e., exceed the capacity limit, violating the physical constraint that the capacity limit cannot be exceeded, then the preliminary candidate state is determined to not meet the preset rules.

[0066] It should be noted that the hard constraint priority strategy is enforced during the state generation process, which strictly controls the physical logic and numerical boundaries of the large language model, effectively avoiding the problem that generative models are prone to producing illusions or generating data that violates physical common sense.

[0067] Step 3: The preliminary candidate states that meet the preset rules are determined as the candidate states for the next time step.

[0068] In practice, after each preliminary candidate state is verified by the hard constraint filter, if it successfully passes the algebraic and boundary checks of the conservation rules and boundary condition rules, it is marked as a valid state and directly determined as a candidate state for the next time step, and aggregated into the set of valid next states. For preliminary candidate states that fail the checks, the evolutionary path is directly blocked. It should be noted that if all preliminary candidate states fail the checks, a backtracking and resampling mechanism is triggered, regenerating the microscopic behaviors and actions and re-synthesizing them with added punitive cues. This ensures that all candidate states entering the subsequent value backtracking and deduction process possess absolute physical and logical correctness and rationality, greatly improving the purity of the underlying evolutionary data.

[0069] The method provided in this invention generates multiple preliminary candidate states by calling a large language model as a state transition function, and introduces a hard constraint checker containing conservation rules and boundary condition rules to strictly verify the physical properties and numerical boundaries of the preliminary candidate states, thereby selecting legal and reliable candidate states for the next time step. This constructs a generative state transition mechanism that prioritizes hard constraints. Without the need to pre-build complex explicit mathematical or physical simulation models, it cleverly integrates the generalization generation capability of the large language model with the rigid constraints of hard rules, realizing high-fidelity and high-reliability state transitions and closed-loop deduction in complex open systems.

[0070] In one embodiment, Figure 3 This is a flowchart illustrating the state transition method provided by the present invention, as shown below. Figure 3 As shown, the method includes: When performing state transition synthesis, the current state is first obtained. The micro-level actions (A) generated by the micro-agent are used as input. Next, the current state and actions are input into the large language model, where probabilistic generation is performed using LLM. Specifically, the large language model is used as the state transition function F(S,A)->Candidates for generative transition, thereby generating a candidate state list containing multiple possible next-time states. .

[0071] Subsequently, each candidate state in the candidate state list is traversed and input into a hard constraint filter for rigorous validation. During validation, the hard constraint filter invokes a pre-established rule base, which contains inviolable hard constraints such as physical conservation (e.g., input-output conservation in algebraic validation), topological connectivity, and numerical boundaries. If a candidate state successfully passes the hard constraint filter's validation, it is marked as a valid state and added to the valid state set for subsequent value backtracking and deduction.

[0072] If a candidate state violates the aforementioned hard constraint rules, it will be directly filtered out and blocked. Further, it is determined whether the generation of the current batch has all failed, i.e., all candidate states have failed the validation. If all fail, an error callback mechanism is triggered, which then performs punitive resampling, requiring the agent to regenerate the micro-behavioral actions with added punitive prompts. These actions are then input back into the large language model for a new round of probability generation until a valid set of states, i.e., the candidate states for the next time step, is synthesized.

[0073] Understandably, this process effectively achieves generative state transitions without an explicit physics engine, and strictly guarantees the correctness of physics and logic through a hard constraint priority strategy.

[0074] Based on any of the above embodiments, step 150 includes: Starting from the next state of the target, the evolution trajectory of multiple future time steps is predicted; Extract the multidimensional value of the evolutionary trajectory during the pre-simulation process; the multidimensional value includes at least two of the following: goal achievement degree, system risk entropy, and state stability parameter value; The value score of the evolutionary trajectory is obtained by weighting the multidimensional values. The tree node score is updated based on the value score. The tree node score is compared with a preset score threshold. If the tree node score is lower than the preset score threshold, the branch corresponding to the evolution trajectory is pruned. The state nodes corresponding to the branches that were not pruned are added as valid nodes to the inference results to complete the update.

[0075] Specifically, in step one, starting from the target's next state, the evolution trajectory is rehearsed for multiple future time steps. In this step, the evolution trajectory refers to a summary record of a series of continuous changes in the system state over a consecutive period of time, starting from the currently determined legal state.

[0076] In practical implementation, the next-time state of the target, which has passed the prior conflict arbitration and hard constraint verification, is taken as the starting point, and a lightweight large language model is used for rapid pre-simulation. Specifically, a generative expansion strategy is adopted to quickly deduce the summary trajectory of the next-time state of each target for the next n time steps, thereby obtaining the evolution trajectory of multiple future time steps. For example, n=3~10.

[0077] Step two involves extracting the multidimensional value of the evolutionary trajectory during the pre-evolution process. This multidimensional value includes at least two of the following: goal achievement degree, system risk entropy, and state stability parameter values. This embodiment of the invention does not specifically limit the value of these. In this step, multidimensional value refers to a set of quantitative indicators used to comprehensively evaluate the merits of a particular evolutionary trajectory. Goal achievement degree refers to the normalized value of the evolutionary trajectory ultimately approaching or achieving the macro-planning goal. System risk entropy is an indicator that measures the uncertainty or potential collapse risk of the system during its evolution. State stability parameter values ​​are typically represented by the inverse of the variance of the state vector, reflecting the stability of the system's evolution process.

[0078] In practical implementation, for each evolutionary trajectory obtained from the pre-simulation, the value indicators parameters of the aforementioned multiple dimensions during the pre-simulation process are extracted. For example, the degree of achievement of macro goals, system risk entropy, and state stability parameter values ​​are extracted. Combining the complex information dissemination and intervention simulation scenarios in the field of public safety, for the pre-simulated public opinion evolution trajectory, the achievement of the goal of information coverage, the system risk entropy of uncontrolled heat, and the state stability parameter of network node emotions are extracted and used as multi-dimensional value input data for evaluating the pre-simulated trajectory.

[0079] Step three involves weighted calculation of the multidimensional values ​​to obtain the value score of the evolutionary trajectory. In this step, the value score refers to the final score used to characterize the confidence and preference of the evolutionary branch after integrating the multidimensional value indicators.

[0080] In the specific calculation, a multidimensional value function is used to comprehensively evaluate the evolutionary trajectory. The calculation formula is as follows: ; In the formula, This represents a multidimensional value function used to calculate the value score. Indicates the degree of goal achievement; The reciprocal of the variance of the vector representing the next state of the target; Represents the system risk entropy; This represents the value of the state stability parameter; These represent the weighting coefficients of the corresponding indicators, which can be adjusted based on the actual situation.

[0081] In one embodiment, considering the complex information dissemination and intervention scenarios in the public safety field, value scores are calculated for the evolutionary trajectories under different intervention strategies. For example, the value score of the evolutionary trajectory is calculated using the following formula, as shown below: ; In the formula, The index represents the level of public opinion heat; the lower the level of public opinion heat, the less impact this index has on the overall value. The greater the contribution; The reciprocal of the variance of the vector representing the next state of the target; It indicates the level of conflict or inconsistency within the system, such as a conflict between macro-level planning and micro-level behavioral trends.

[0082] In another embodiment, taking the topology evolution and robustness optimization scenario of a distributed resource transmission network as an example, the pre-simulation of the evolution trajectory of maintaining the existing topology results in cascading congestion after three steps, leading to extremely low target achievement and stability and extremely high risk. At this time, the overall value score obtained by weighted calculation is extremely low. On the other hand, the pre-simulation of the trajectory of activating the backup high-cost link, although the cost increases in the short term, the network throughput remains stable in the long term, and the value score obtained by weighted calculation is relatively high.

[0083] Step 4: Update the tree node score based on the value score, compare the tree node score with a preset score threshold, and prune the branch corresponding to the evolutionary trajectory if the tree node score is lower than the preset score threshold. Here, the tree node score refers to the node confidence score used to guide the direction of evolutionary exploration and utilization in the inference tree search framework. Here, the preset score threshold refers to the dynamic critical pruning line used to eliminate low-value or high-risk evolutionary paths.

[0084] In practice, the calculated value score is backpropagated, and the tree node score of the corresponding state node on the evolution path is updated based on the improved Upper Confidence Bound for Trees (UCT) algorithm. The update formula can be set as the average value score of the node plus the exploration constant multiplied by the square root of the total number of visits to the parent node (taking the natural logarithm) divided by the number of visits to the current node. For example, the specific formula for updating the tree node score can be expressed as: ; In the formula, Represents the score of a tree node; This represents the cumulative value score; N represents the number of visits; C represents the exploration constant; and T represents the total number of visits to the parent node.

[0085] After the update is complete, compare the tree node's score with a preset score threshold. If the tree node's score is lower than the preset score threshold, directly prune the branch corresponding to that evolutionary trajectory. For example, in information propagation simulation, decisively prune low-score, dangerous branches, such as those where the popularity rises infinitely and causes system collapse, to avoid invalid simulations.

[0086] It should be noted that by using the tree node scoring update and dynamic pruning mechanism, low-value and high-risk evolutionary paths are effectively eliminated, ensuring that limited computing resources are concentrated on the in-depth exploration of the optimal evolutionary path.

[0087] Step 5: Add the state nodes corresponding to the branches that were not pruned as valid nodes to the inference results to complete the update. Specifically, for high-value or high-confidence evolution trajectories retained after pre-simulation, value calculation, and pruning, extract the state nodes that were not pruned from their corresponding state evolution paths. These high-quality nodes, selected through multiple layers of screening, are then formally stored and added to the multi-path inference tree or structured inference result model for the current complex event. With continuous iterations and value backtracking, valid nodes are continuously added to expand the evolution paths, thereby continuously updating the inference results.

[0088] Understandably, gradually building and improving high-quality multi-path structured inference results lays the data foundation for the final output of an interpretable evolution report containing probabilities and key turning points.

[0089] The method provided in this invention performs rapid multi-timestep pre-simulation starting from the next state of the target, extracts multi-dimensional values ​​such as target achievement degree, system risk entropy, and state stability, and performs weighted aggregation to obtain a value score. Then, based on the score, the confidence of tree nodes is updated and dynamic branch pruning is performed. Finally, high-value nodes that have not been pruned are added to the inference results. The generative state transition process is cleverly embedded in the Monte Carlo search framework. Through multi-dimensional value backtracking and upper confidence bound algorithm to guide the exploration direction, the long chain inference process has a robust and verifiable convergence control capability, effectively avoids the path exploration from getting stuck in local optima, and greatly improves the reliability and decision-making of multi-path inference results.

[0090] Based on any of the above embodiments, step 110 includes: Analyze the scene description information and extract the set of key variables and global boundary conditions; At each time step, multiple candidate planning branches are generated, and for each candidate planning branch, the parameter values ​​of time series coverage, variable closure degree, and constraint satisfiability are extracted. The consistency score of each candidate planning branch is obtained by weighted summing of the parameter values ​​of the temporal coverage, variable closure, and constraint satisfiability of each candidate planning branch. A preset number of candidate planning branches with the highest consistency scores are retained to construct the macro-planning tree.

[0091] Specifically, the first step is to parse the scenario description information and extract the set of key variables and global boundary conditions. In this step, the set of key variables refers to the underlying core parameters extracted from the unstructured initial text to construct the system's state space. Here, global boundary conditions refer to the hard constraints that must not be violated throughout the entire complex event deduction cycle, such as global thresholds like time limits and total resource limits.

[0092] In specific implementation, the system receives unstructured scene description text as input and performs structured mapping and parsing using a large language model. Key variables related to entity states are extracted, and global boundary conditions for the simulation cycle are identified. In one embodiment, considering the evolution simulation scenario of a sudden flood disaster, when the input scene description is "a sudden flood disaster in a city, requiring rescue within 72 hours," parsing can accurately extract a set of key variables related to the number of trapped people, water levels, etc., as well as a time limit of 72 hours. The global boundary conditions of ).

[0093] Next, step two is executed, generating multiple candidate planning branches at each time step. For each candidate planning branch, the parameters of temporal coverage, variable closure, and constraint satisfiability are extracted. In this step, candidate planning branches refer to the macro-evolutionary paths with different stage divisions and goal settings that the large language model attempts to generate under the search strategy. Here, temporal coverage is used to characterize whether there are uncovered gaps or logical overlaps in the generated plan on the time axis. Here, variable closure is used to measure whether the macro-goals set within a stage incorrectly reference undefined variables outside the set of key variables. Here, constraint satisfiability is used to determine whether the current plan exceeds the limits of the global boundary conditions.

[0094] In practical implementation, a bundle search strategy and a large language model are combined. A specific bundle width parameter is set at each generation step, such as bundle width K=5, to generate multiple candidate planning branches in parallel. Subsequently, the logical structure and timeline of each candidate planning branch are traversed, extracting parameters for detecting time windows or overlaps, variable closure for detecting undefined variable references, and constraint satisfiability for detecting violations of global boundaries. For example, when traversing the timeline of the planning tree, if overlapping timestamps of adjacent stages are detected (…),… ) or there is a fault gap ( This is reflected in the parameter value of time series coverage.

[0095] It should be noted that by using bundle search to generate multiple candidate plans in parallel and extracting multi-dimensional logical constraint evaluation parameters, it is possible to comprehensively and precisely measure the performance of each generated plan in terms of logical rigor and feasibility.

[0096] Further, in step three, the time series coverage, variable closure, and constraint satisfiability parameter values ​​of each candidate planning branch are weighted and summed to obtain the consistency score of each candidate planning branch.

[0097] In practical implementation, corresponding weight coefficients are assigned to the extracted multidimensional parameters. A weighted summation algorithm is used to score the time series coverage, variable closure, and constraint satisfiability parameters of each candidate planning branch, thus calculating the consistency score for each candidate branch. The consistency score of any candidate branch can be calculated using the following formula, as shown below: ; In the formula, This represents the consistency score of any candidate branch; The parameter value representing the timing coverage; The parameter value representing the degree of closure of a variable; The parameter value representing the satisfiability of the constraint; These represent the weights of each dimension.

[0098] It should be noted that by using a multi-dimensional parameter weighted summation method to scientifically quantify the consistency of planning branches, a strict planning quality evaluation standard was constructed, effectively eliminating the illusionary planning that may be generated by large language models, which may have logical contradictions or be unrealistic.

[0099] Finally, step four is executed, retaining a preset number of candidate planning branches with the highest consistency scores to construct the macro-planning tree.

[0100] In practice, all generated candidate planning branches are sorted in descending order based on the calculated consistency score. A predetermined number of candidate planning branches with the highest scores (i.e., Top-K) are selected and retained, while inferior branches with lower scores are removed. The retained optimal candidate planning branches are used as the final top-level framework to construct a macro-planning tree containing an ordered set of stages, a global objective, and resource constraints.

[0101] In one embodiment, combining the complex information dissemination and intervention scenarios in the field of public safety, the final constructed optimal macro-programming tree includes two ordered temporal stages and the macro-objective corresponding to each temporal stage. Stage 1 (t∈[0,7]): diffusion rise period, the macro-objective is to maximize information coverage; Stage 2 (t∈[7,14]): steady decline period, the macro-objective is to reduce the system entropy.

[0102] Understandably, by selecting the best of the best and retaining high-quality candidate branches to construct the final macro-planning tree, the foresight and global consistency of the top-level design in the deduction process are ensured, and logical deviations in the long chain deduction are prevented from occurring in the initial stage.

[0103] The method provided in this invention extracts a set of key variables and global boundary conditions by parsing scenario description information. When generating multiple candidate planning branches in parallel, it extracts time series coverage, variable closure degree, and constraint satisfiability parameters respectively, and performs weighted summation to obtain a consistency score. Finally, it retains the candidate branch with the highest score to construct a macro-planning tree. This method efficiently and rigorously transforms unstructured text into a structured planning model with strict time series constraints and goal orientation. It provides a highly logically consistent global skeleton support for the long-chain evolution of complex systems from the source, and greatly improves the accuracy, rationality, and anti-divergence ability of the initial direction of deduction.

[0104] Based on any of the above embodiments, in step 120, under the constraint of the macroscopic objective at the current time step, driving multiple agents to generate microscopic behavioral actions includes: Under the constraint of the macroscopic objective at the current time step, the same agent is sampled multiple times in parallel to obtain an initial action set containing multiple initial actions; Calculate the semantic similarity between each initial action, and cluster the initial actions in the initial action set based on the semantic similarity to obtain multiple action clusters; The initial action corresponding to the cluster center of the action cluster with the highest frequency of occurrence is selected as the candidate action, and if no logical mutual exclusion conflict is detected in the candidate action, the candidate action is used as the micro behavior action.

[0105] Specifically, first, step one is executed: under the constraint of the macroscopic objective at the current time step, multiple parallel samples are performed on the same agent to obtain an initial action set containing multiple initial actions. In this step, an initial action refers to an unverified temporary operation instruction generated and computed in a single instance under specific contextual constraints. Here, the initial action set refers to the set of temporary instructions aggregated by repeated sampling for a single agent.

[0106] In practical implementation, the macro-level stage constraints (i.e., macro-level objectives) of the current time step in the macro-programming tree can be extracted. This is combined with the current state snapshot containing all entity attribute values ​​and a set of agent profiles defining the micro-level individual behavioral logic to construct a context for triggering the generation of prompt words. Under the constraints of this macro-level objective, a large language model is used to perform m parallel samplings on the same agent, such as micro-level agents representing specific nodes. Typically, m is set to a value of 3 to 5 times, thereby generating an initial action set containing m initial actions.

[0107] For example, in complex information dissemination and intervention scenarios in the field of public safety, as a micro-agent, such as an ordinary netizen as a dissemination node, based on the current local public opinion state, it generates different initial actions such as forwarding and commenting, or remaining silent multiple times according to probability, which together constitute the initial action set of the dissemination node.

[0108] Understandably, by obtaining rich instruction samples through multiple parallel sampling mechanisms, we can fully explore the diverse distribution of generative large language models in behavioral inference, laying a data foundation for subsequently selecting the most logical behavioral decisions.

[0109] Next, step two is executed: the semantic similarity between each initial action is calculated, and the initial actions in the initial action set are clustered based on the semantic similarity to obtain multiple action clusters. In this step, an action cluster refers to a logical category formed by grouping initial actions with the same or highly similar execution intentions based on the similarity of their natural language meanings.

[0110] In practice, for the generated initial action set, natural language processing techniques or vector embedding models are used to calculate the semantic similarity between each pair of initial action texts. Then, based on the calculated semantic similarity scores, a clustering algorithm is used to automatically group all initial actions in the initial action set. Actions with similar semantic distances and expressing core meanings are grouped together. Figure 1 The initial actions are grouped into the same category, thereby integrating the discrete initial action set into multiple structured action clusters.

[0111] Understandably, semantic similarity clustering analysis can effectively identify and summarize the core behavioral intent in multi-source sampling instructions, overcoming the randomness and discreteness of the output results of the generative model in natural language expression.

[0112] Finally, in step three, the initial action corresponding to the cluster center of the action cluster with the highest frequency of occurrence is selected as the candidate action. If no logical mutual exclusion conflict is detected among the candidate actions, the candidate action is adopted as the micro-behavioral action. In this step, the candidate action refers to the preliminary decision instruction that represents the highest generation probability and the strongest group consensus after clustering statistics. Here, logical mutual exclusion conflict refers to two or more contradictory instructions generated simultaneously for the same executing entity that are absolutely impossible to simultaneously hold true in terms of physical operation or business logic.

[0113] In practice, the number of initial actions contained in each action cluster obtained from the above clustering is counted, and a clustering voting mechanism is executed to select the action cluster with the most actions and the highest frequency of occurrence. The initial action represented by the cluster center of this high-frequency action cluster is extracted as a candidate action. Next, a self-consistent filtering mechanism is introduced to perform strict logical verification on the candidate action. Taking information propagation network deduction as an example, if it is detected that the candidate action requires the same propagation node to perform both silent and high-frequency forwarding behaviors at the same time step, which have obvious logical mutual exclusion conflicts, it is directly eliminated; conversely, if it is detected that the candidate action is logically consistent and does not have logical mutual exclusion conflicts, its rationality is confirmed, and it is formally established as the final output micro-behavioral action.

[0114] It should be noted that by using clustering voting and strict logical mutual exclusion verification, not only are the optimal generation actions with a high probability retained, but contradictory behaviors that violate common sense are also thoroughly filtered out, ensuring the legality of micro-behavioral instructions at the business and physical levels.

[0115] The method provided in this invention, under the constraint of a macroscopic objective, performs multiple parallel samplings on the same agent to obtain an initial action set, and clusters multiple action clusters based on semantic similarity. Then, the cluster center of the high-frequency action cluster is selected as a candidate action, and after logical mutual exclusion detection to confirm that there is no conflict, it is established as the final microscopic behavior action. The method introduces a self-consistency filtering mechanism and cluster voting logic, which effectively eliminates the logical errors or low-probability anomalies that are very easy to occur when large language models generate agent behavior. This greatly improves the rationality and purity of the emergence of multi-agent microscopic behavior, and provides a solid and reliable microscopic evolutionary driving force for the long chain deduction process.

[0116] Based on any of the above embodiments Figure 4 This is a flowchart illustrating the complex event deduction method based on generative hierarchical Monte Carlo programming provided by the present invention, as shown below. Figure 4 As shown, the method includes: The process begins with inputting a scenario description and boundary conditions. The system receives user-defined unstructured event text, initial states, and global boundary constraints such as the simulation period. Then, in the S1 Macro Planning stage, the system uses a large language model combined with a cluster search strategy to generate a planning tree that includes temporal stage divisions and global macro objectives. This provides a global skeleton constraint to prevent logical divergence in subsequent simulations.

[0117] Next, the S2-level agent behavior generation (HMAS Execution) is executed. Under the guidance of macro instructions and the constraints of the macro goal of the current time step, macro agents and micro agents are instantiated, and the multi-agents are driven to generate specific candidate action sets through a self-consistent filtering mechanism to achieve micro behavior emergence.

[0118] After generating the actions, the process enters the S3 state transition and hard constraint check (TransitionSynthesis) stage. The system synthesizes the aforementioned set of micro-actions and the current state into the generative large language model to synthesize the candidate state for the next time step. During this process, the system strictly implements a hard constraint priority filtering strategy, verifying against inviolable physical rules such as resource conservation and numerical boundaries, blocking illegal paths to generate a legal next state. After obtaining a legal state, the S4 generative MCTS value backtracking is performed. The system performs multi-step rapid pre-playback starting from the current legal state and extracts indicators such as goal achievement, system risk entropy, and state stability for multi-dimensional value evaluation (V). Subsequently, the upper confidence bound algorithm is used to update the path node scores and prune low-value or high-risk branches.

[0119] Following value backtracking, the process enters the core S5 Dynamic Conflict Arbiter stage, where the system monitors the macro-micro consistency between the micro-evolutionary trend and the macro-planning objective in real time. If the calculated consistency score is lower than a preset threshold, the system will perform tiered arbitration based on the conflict intensity and micro-confidence of the micro-trend: when the conflict intensity is low, Level 1 micro-correction is performed, i.e., the macro-plan remains unchanged, a rejection signal is sent to the system, and the process reverts to the S2 stage, requiring the multi-agents to resample and generate actions. When the micro-trend has strong inertia and the conflict is severe enough to cause the original plan to fail, Level 2 macro-reconstruction is performed, and the system reverts to the S1 stage and regenerates the macro-planning tree starting from the current latest state.

[0120] When the macro and micro evolution reach a consensus or are in an acceptable state, the process finally advances to the S6 structured cognitive output stage. The system will extract the key turning points where the value function gradient changes drastically from the selected evolutionary paths and output a structured deduction tree containing multi-dimensional indicators and complete evolutionary paths, thereby realizing interpretable, multi-path long-chain deduction in an open environment.

[0121] In one specific embodiment, this embodiment aims to simulate the propagation and diffusion process of specific information among multiple types of nodes in an open network environment, and to deduce the evolutionary path under different intervention strategies without explicit dynamic equations. The scenario setting includes: System environment: Constructing an information propagation network containing multiple types of agents. Agent types include: information source nodes (e.g., media), propagation nodes (e.g., netizens), and control nodes (e.g., management departments). State space. It is composed of the network topology, the sentiment / activation state vectors of each node, and the global public opinion heat index. Parameter configuration: Monte Carlo search beam width. Number of single-step samplings UCT exploration constant Macroeconomic consistency threshold Arbitration trigger threshold Total simulation budget Second rollout.

[0122] The specific execution method includes: S1: Macro Planning Generation. System Input Description (Emergency event text) and boundaries (Derivation cycle 14 days). Generate a macro-programming tree using a Large Language Model (LLM). It consists of two main phases: Phase P1 ( Phase P2: The diffusion rise phase, with the goal of maximizing information coverage. (This refers to) a stable decline phase, with the objective of reducing system entropy. A global constraint is defined: the global heat index. .

[0123] S2: Hierarchical Proxy Behavior Generation (HMAS). At the current time step... The macro-level agent (control node) generates a strategy vector and issues an authoritative announcement; the micro-level agent (propagation node), based on the current local state, generates actions according to probability and forwards the message while commenting or remaining silent. The system performs self-consistent voting on the generated set of actions, eliminating logically conflicting actions (such as a single node simultaneously executing both silent and high-frequency forwarding).

[0124] S3: Transition Synthesis. Input the action set into the generative model to synthesize a candidate state set for the next time step. Hard constraint check: Checks whether the candidate state violates physical constraints (such as the number of node connections cannot exceed the upper limit). If it violates the constraint, a backtracking resampling is triggered.

[0125] S4: Generative MCTS Value Backtracking (GMCP). For candidate states... conduct The generative Rollout (preview) is performed step by step. A multidimensional value function is calculated to obtain a value score. The value score is updated using the UCT algorithm to obtain a path score, and low-value branches, such as those whose popularity increases indefinitely and leads to system crashes, are pruned. S5: Arbitration mechanism intervenes. Conflict detection: In At that time, the evolutionary trend of micro-agents showed that the heat index increased exponentially, while macro-planning required a period of stability. Calculations revealed a consistency score. This triggers arbitration. Conflict resolution: The arbitrator determines that micro-trends have strong inertia, therefore implementing a macro-adjustment strategy: postponing the smooth decline period in the macro-planning to... The system dynamically inserts a new macro-level objective to force noise reduction intervention. Based on the adjusted macro-level plan, the system regenerates the behavior of the control nodes.

[0126] S6: Structured output. Outputs a multi-path derivation tree, displaying the main path: in The introduction of strong intervention nodes at opportune moments caused the system's heat to... After reaching its peak, the price quickly fell back. Key turning points were marked as intervention points.

[0127] In another embodiment, this embodiment aims to simulate the topology evolution and optimal path reconstruction process of a system in a complex distributed resource network (such as a supply chain network, power dispatch network, or computing power network) when a core transmission node experiences a performance blockage (Shock). The scenario setting includes: Network definition: Graph structure. . For nodes (source node, relay node, and sink node). For transmission links. Input disturbance: Within a specific time window, some edges... The impedance increases instantaneously to infinity, simulating a power outage or link interruption. Macro-level objective: Network overall throughput stability. Total network transmission cost increment (Cost Delta) Hard constraint: any node Storage capacity It cannot be negative and cannot exceed the capacity limit. .

[0128] The method execution flow includes: S1: Macro-level planning tree generation. The system generates a planning tree for network topology reconstruction. The initial plan includes local cache buffering mode and neighborhood load balancing mode.

[0129] S2: Micro-level agent decision-making. Each node agent generates specific resource scheduling requests based on local awareness. For example, a node... The request will be sent to the node Increased transmission volume .

[0130] S3: Hard-constraint-first transition synthesis. The system synthesizes the network state for the next time step. In this step, the hard-constraint checker identifies nodes. The request will cause the node Storage overflow ( Processing mechanism: The system automatically rejects this micro-behavior and forcibly triggers the hard constraint priority rule to limit the transmission rate.

[0131] S4: MCTS Value Assessment and Pruning. The system performs Rollout prediction of future network conditions. Branch A: Maintain the existing topology, only adjust traffic. Prediction result: After 3 steps, cascading congestion occurs in the network, and the value function... Extremely low. Branch B: Activate the backup high-cost link (Hidden Edge). Prediction: Short-term cost increases, but long-term throughput remains stable; value function... High. The system is based on The value is used to prune branch A.

[0132] S5: Tiered arbitration process. Trigger mechanism: Macro-planning limits "single link load rate". However, the micro-optimal solution (branch B) requires a backup link to be running at full capacity. A conflict occurred. Arbitration was implemented: the arbitrator's assessment revealed that strictly adhering to macro-level constraints would lead to a network-wide shutdown (high risk). Therefore, the arbitrator chose to relax the macro-level constraints: temporarily allowing link load rates to reach [a certain level]. However, a timing constraint is added requiring a fallback after 3 time steps. Path bifurcation: The system generates and retains two evolution paths, one a high-throughput path with relaxed constraints and the other a low-throughput path with strict constraints, for decision-makers to refer to.

[0133] S6: Output Results. Output the optimal topology evolution path: in At this critical turning point, a structural change occurs in the network topology (link switching). Quantitative metric: The network throughput stability of the main path remains stable. The transmission cost increment is To meet macro-level objectives.

[0134] Based on any of the above embodiments Figure 5 This is a schematic diagram of the complex event deduction system based on generative hierarchical Monte Carlo programming provided by the present invention, as shown below. Figure 5 As shown, the system constructs an end-to-end closed-loop deduction architecture encompassing macro-level planning guidance, micro-level behavioral emergence, generative state backtracking, and dynamic conflict arbitration. Specifically, the system mainly includes an input layer, the GMCP-H core deduction engine, and an output layer.

[0135] The input layer is primarily responsible for receiving initial data from the user, including scene description text (D), global boundary conditions (B), and the world model Prompt. This input data provides the initial environment settings, global constraints, and domain knowledge background for subsequent inference processes.

[0136] The GMCP-H core inference engine is the system's core processing unit, comprising six key modules. First is the Macro Planning Generation module, which receives data from the input layer and uses a large language model combined with a beam search strategy to generate a macro planning tree containing temporal stages, a global set of objectives, and resource constraints, providing a global skeleton constraint for subsequent inferences. Second is the Hierarchical Agent Behavior Generation module (HMASExecution), which instantiates macro and micro agents under the constraints of the current time step of macro planning. Each agent generates a candidate set of actions based on local perception and role settings, and introduces a self-consistency filtering mechanism to eliminate abnormal behaviors. Next is the Transition Synthesis with HardConstraints module, which inputs the action set into the generative model, synthesizes a candidate state set for the next time step, and establishes a hard constraint filter to block paths that violate physical constraints. Next is the GenerativeMCTS value backtracking module, which embeds the aforementioned state generation process into the Monte Carlo Tree Search (MCTS) framework. It utilizes a large language model for rapid rollout and updates path confidence for pruning based on multi-dimensional value assessment and an improved UCT algorithm. The fifth module is the Dynamic Arbiter, a core control component that monitors the consistency between the evolutionary trends of micro-behaviors and the pre-defined target trajectory of macro-planning in real time. When the inconsistency is below a threshold or the micro-trend has strong inertia, it triggers hierarchical arbitration logic, sending resampling signals to subordinate micro-agents or feeding back to reconstruct the macro-planning tree, and even generating path branches. Finally, the Structured Output module is responsible for integrating and analyzing the results after multiple rounds of iterative deduction.

[0137] The output layer receives the final processing results from the core inference engine and outputs structured cognitive information such as structured inference trees, key turning points, and risk / value reports, providing decision-makers with traceable and interpretable inference results that are easy to integrate into various decision-making systems and simulation platforms.

[0138] Based on any of the above embodiments Figure 6 This is a schematic diagram of the complex event evolution deduction device provided by the present invention, as shown below. Figure 6 As shown, the device includes: Macro planning unit 610 generates a macro planning tree that includes time-series stages and macro objectives based on scenario description information.

[0139] The micro-generation unit 620, under the constraint of the macro-objective at the current time step, drives multiple agents to generate micro-behavioral actions, and determines the candidate state for the next time step based on the micro-behavioral actions and the current state. The consistency measurement unit 630 extracts the micro-evolution trend based on the candidate state at the next time step, and calculates the consistency measurement value between the micro-evolution trend and the macro-objective at the current time step.

[0140] The adjustment unit 640, when the consistency metric value is less than the preset consistency threshold, triggers the adjustment of the micro-behavioral action or the macro-planning tree, and re-acquires the candidate state of the next moment to calculate the consistency metric value until the consistency metric value is not less than the preset consistency threshold, thereby obtaining the target state of the next moment.

[0141] The deduction unit 650 updates the deduction results of complex events based on the next moment's state of the target.

[0142] The apparatus provided in this invention generates a macro-planning tree containing temporal stages and macro-objectives based on scene description information. Under the constraint of the macro-objective at the current time step, it drives multiple agents to generate micro-behavioral actions, thereby determining the candidate state for the next moment, extracting the micro-evolutionary trend and calculating its consistency metric with the macro-objective. When the consistency metric is low, it triggers dynamic adjustments to the actions or plans to obtain the state of the target at the next moment, and finally updates the deduction results of complex events. This effectively overcomes the problem of the lack of explicit environment models in the prior art, which makes direct deduction impossible. It constructs a complete closed-loop deduction framework that unifies macro-planning, micro-behavioral generation, state transition synthesis, and value backtracking arbitration. This greatly reduces the deviation caused by the emergence of multi-agent behaviors in long-chain deductions and improves the credibility, stability, and interpretability of deduction results in open environments.

[0143] Based on any of the above embodiments, the adjustment unit is specifically used for: If the consistency metric is less than the preset consistency threshold, the arbitration level is determined based on the conflict intensity and micro-confidence of the micro-evolutionary trend. According to the arbitration level, resampling signals are sent to the multi-agent to adjust the micro-behavioral actions or reconstruct the macro-planning tree.

[0144] Based on any of the above embodiments, the micro-generation unit is specifically used for: The micro-behavioral actions and the current state are input into the large language model to obtain multiple preliminary candidate states output by the large language model; Extract the physical properties and numerical boundaries of each of the preliminary candidate states, and verify whether each of the preliminary candidate states satisfies the preset rules; the preset rules include conservation rules and boundary condition rules. The preliminary candidate states that satisfy the preset rules are determined as the candidate states for the next time step.

[0145] Based on any of the above embodiments, the deduction unit is specifically used for: Starting from the next state of the target, the evolution trajectory of multiple future time steps is predicted; Extract the multidimensional value of the evolutionary trajectory during the pre-simulation process; the multidimensional value includes at least two of the following: goal achievement degree, system risk entropy, and state stability parameter value; The value score of the evolutionary trajectory is obtained by weighting the multidimensional values. The tree node score is updated based on the value score. The tree node score is compared with a preset score threshold. If the tree node score is lower than the preset score threshold, the branch corresponding to the evolution trajectory is pruned. The state nodes corresponding to the branches that were not pruned are added as valid nodes to the inference results to complete the update.

[0146] Based on any of the above embodiments, the macro-planning unit is specifically used for: Analyze the scene description information and extract the set of key variables and global boundary conditions; At each time step, multiple candidate planning branches are generated, and for each candidate planning branch, the parameter values ​​of time series coverage, variable closure degree, and constraint satisfiability are extracted. The consistency score of each candidate planning branch is obtained by weighted summing of the parameter values ​​of the temporal coverage, variable closure, and constraint satisfiability of each candidate planning branch. A preset number of candidate planning branches with the highest consistency scores are retained to construct the macro-planning tree.

[0147] Based on any of the above embodiments, the micro-generation unit is specifically used for: Under the constraint of the macroscopic objective at the current time step, the same agent is sampled multiple times in parallel to obtain an initial action set containing multiple initial actions; Calculate the semantic similarity between each initial action, and cluster the initial actions in the initial action set based on the semantic similarity to obtain multiple action clusters; The initial action corresponding to the cluster center of the action cluster with the highest frequency of occurrence is selected as the candidate action, and if no logical mutual exclusion conflict is detected in the candidate action, the candidate action is used as the micro behavior action.

[0148] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a complex event evolution deduction method. This method includes: generating a macro-planning tree containing temporal stages and macro-objectives based on scene description information; driving multiple agents to generate micro-level actions under the constraints of the macro-objective at the current time step, and determining candidate states for the next time step based on the micro-level actions and the current state; extracting a micro-evolutionary trend based on the candidate states for the next time step, and calculating a consistency metric between the micro-evolutionary trend and the macro-objective at the current time step; if the consistency metric is less than a preset consistency threshold, triggering adjustments to the micro-level actions or the macro-planning tree, and re-acquiring the candidate states for the next time step to calculate the consistency metric until the consistency metric is not less than the preset consistency threshold, thus obtaining the target state for the next time step; and updating the deduction results of the complex event based on the target state for the next time step.

[0149] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0150] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the complex event evolution deduction method provided by the above methods. The method includes: generating a macro-planning tree containing temporal stages and macro-objectives based on scene description information; driving multiple agents to generate micro-behavioral actions under the constraints of the macro-objectives at the current time step, and determining candidate states for the next time step based on the micro-behavioral actions and the current state; extracting micro-evolutionary trends based on the candidate states for the next time step, and calculating a consistency metric between the micro-evolutionary trends and the macro-objectives at the current time step; if the consistency metric is less than a preset consistency threshold, triggering adjustments to the micro-behavioral actions or the macro-planning tree, and re-acquiring the candidate states for the next time step to calculate the consistency metric until the consistency metric is not less than the preset consistency threshold, thereby obtaining the target state for the next time step; and updating the deduction results of the complex event based on the target state for the next time step.

[0151] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the complex event evolution deduction method provided by the above methods. The method includes: generating a macro-planning tree containing temporal stages and macro-objectives based on scene description information; driving multiple agents to generate micro-behavioral actions under the constraints of the macro-objective at the current time step, and determining candidate states for the next time step based on the micro-behavioral actions and the current state; extracting micro-evolutionary trends based on the candidate states for the next time step, and calculating a consistency metric between the micro-evolutionary trends and the macro-objective at the current time step; if the consistency metric is less than a preset consistency threshold, triggering adjustments to the micro-behavioral actions or the macro-planning tree, and re-acquiring the candidate states for the next time step to calculate the consistency metric until the consistency metric is not less than the preset consistency threshold, thereby obtaining the target state for the next time step; and updating the deduction results of the complex event based on the target state for the next time step.

[0152] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0153] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for deducing the evolution of complex events, characterized in that, include: Based on the scenario description information, a macro-planning tree is generated that includes time-series stages and macro-objectives; Under the constraints of the macro-objective at the current time step, drive multiple agents to generate micro-behavioral actions, and determine the candidate state for the next time step based on the micro-behavioral actions and the current state; Based on the candidate state extraction of the next time step, the micro-evolution trend is obtained, and the consistency metric between the micro-evolution trend and the macro-objective of the current time step is calculated. If the consistency metric is less than the preset consistency threshold, the micro-behavioral action or the macro-planning tree is adjusted, and the candidate state for the next time step is re-acquired to calculate the consistency metric until the consistency metric is not less than the preset consistency threshold, and the target state for the next time step is obtained. Based on the next state of the target, update the deduction results of complex events.

2. The method for deducing the evolution of complex events according to claim 1, characterized in that, When the consistency metric value is less than a preset consistency threshold, the step of triggering adjustments to the micro-level behavioral action or the macro-level planning tree includes: If the consistency metric is less than the preset consistency threshold, the arbitration level is determined based on the conflict intensity and micro-confidence of the micro-evolutionary trend. According to the arbitration level, resampling signals are sent to the multi-agent to adjust the micro-behavioral actions or reconstruct the macro-planning tree.

3. The method for deducing the evolution of complex events according to claim 1, characterized in that, The process of determining the candidate state for the next moment based on the microscopic behavior and the current state includes: The micro-behavioral actions and the current state are input into the large language model to obtain multiple preliminary candidate states output by the large language model; Extract the physical properties and numerical boundaries of each of the preliminary candidate states, and verify whether each of the preliminary candidate states satisfies the preset rules; the preset rules include conservation rules and boundary condition rules. The preliminary candidate states that satisfy the preset rules are determined as the candidate states for the next time step.

4. The method for deducing the evolution of complex events according to any one of claims 1 to 3, characterized in that, The process of updating the deduction results of complex events based on the next state of the target includes: Starting from the next state of the target, the evolution trajectory of multiple future time steps is predicted; Extract the multidimensional value of the evolutionary trajectory during the pre-simulation process; the multidimensional value includes at least two of the following: goal achievement degree, system risk entropy, and state stability parameter value; The value score of the evolutionary trajectory is obtained by weighting the multidimensional values. The tree node score is updated based on the value score. The tree node score is compared with a preset score threshold. If the tree node score is lower than the preset score threshold, the branch corresponding to the evolution trajectory is pruned. The state nodes corresponding to the branches that were not pruned are added as valid nodes to the inference results to complete the update.

5. The method for deducing the evolution of complex events according to any one of claims 1 to 3, characterized in that, The process of generating a macro-planning tree based on scenario description information, which includes temporal stages and macro-level objectives, includes: Analyze the scene description information and extract the set of key variables and global boundary conditions; At each time step, multiple candidate planning branches are generated, and for each candidate planning branch, the parameter values ​​of time series coverage, variable closure degree, and constraint satisfiability are extracted. The consistency score of each candidate planning branch is obtained by weighted summing of the parameter values ​​of the temporal coverage, variable closure, and constraint satisfiability of each candidate planning branch. A preset number of candidate planning branches with the highest consistency scores are retained to construct the macro-planning tree.

6. The method for deducing the evolution of complex events according to any one of claims 1 to 3, characterized in that, Under the constraint of the macroscopic objective at the current time step, driving multiple agents to generate microscopic behavioral actions includes: Under the constraint of the macroscopic objective at the current time step, the same agent is sampled multiple times in parallel to obtain an initial action set containing multiple initial actions; Calculate the semantic similarity between each initial action, and cluster the initial actions in the initial action set based on the semantic similarity to obtain multiple action clusters; The initial action corresponding to the cluster center of the action cluster with the highest frequency of occurrence is selected as the candidate action, and if no logical mutual exclusion conflict is detected in the candidate action, the candidate action is used as the micro behavior action.

7. A device for predicting the evolution of complex events, characterized in that, include: The macro-planning unit generates a macro-planning tree that includes time-series stages and macro-objectives based on scenario description information; The micro-generation unit, under the constraint of the macro-objective at the current time step, drives multiple agents to generate micro-behavioral actions, and determines the candidate state for the next time step based on the micro-behavioral actions and the current state. The consistency measurement unit extracts the micro-evolution trend based on the candidate state at the next time step, and calculates the consistency measurement value between the micro-evolution trend and the macro-objective at the current time step. The adjustment unit, when the consistency metric value is less than the preset consistency threshold, triggers the adjustment of the micro-behavioral action or the macro-planning tree, and re-acquires the candidate state of the next moment to calculate the consistency metric value until the consistency metric value is not less than the preset consistency threshold, and obtains the target state of the next moment. The deduction unit updates the deduction results of complex events based on the next moment's state of the target.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the complex event evolution deduction method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the complex event evolution deduction method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the complex event evolution deduction method as described in any one of claims 1 to 6.

Images