A multi-branch parallel deduction and decision optimization method, device and equipment based on a large language model and a storage medium

Abstract

The application provides a multi-branch parallel deduction and decision optimization method and device based on a large language model, equipment and a storage medium. A dynamic evolving battlefield state space is constructed by receiving multi-source real-time situation data, a battlefield emergency is identified through situation analysis, and a large language model is called to generate multiple differentiated decision branches each containing macroscopic purposes and microscopic strategies as a trigger condition. An independent parallel battlefield deduction instance is constructed for each decision branch and computing and simulation resources are allocated, driving the instances to synchronously and parallelly deduce real-time deduction states. Multi-dimensional dynamic evaluation is performed to obtain a dynamic comprehensive score, based on which dynamic pruning is performed on branches with scores continuously below a preset threshold, and the computing and simulation resources released by the branches are reallocated to surviving branches according to weights to improve the calculation accuracy. Finally, high-quality branches with higher scores are extracted, a large language model is called to perform decision fusion on the key action sequences of the branches, and an optimal combat path and instruction sequence are output.

Description

Technical Field

[0001] This invention relates to the field of combat decision support and wargaming technology, and in particular to a method, apparatus, device and storage medium for multi-branch parallel inference and decision optimization based on a large language model. Background Technology

[0002] Current wargaming, command and control, and intelligent mission planning systems traditionally aim to replicate the physical battlefield with high fidelity. They involve pre-setting a single combat plan by humans and then conducting single-path simulations in a simulation environment. The system mainly plays the role of "executing the simulation," and the generation and selection of the plan are highly dependent on human intervention.

[0003] To find relatively optimal solutions within a finite state space, existing techniques employ expert rule bases, state machines, Monte Carlo tree search, or reinforcement learning policy networks for decision inference. These methods are effective for closed problems where states are enumerable and value functions are relatively stable.

[0004] However, in adjoint decision-making within an open adversarial environment, a long-standing and unresolved technical challenge exists: the operational decision space evolves continuously over time, and the merits of any candidate solution cannot be determined at the outset. Instead, they gradually emerge during high-fidelity simulations, under conditions of simultaneous autonomous confrontation with enemy entities. Consequently, their value is a priori unknown and non-stationary. Simultaneously, the total computational and simulation power available for simulation is finite, and adjoint decision-making imposes a hard real-time constraint on simulation convergence. Under this constraint, if a static, uniform approach to computational power allocation is adopted for each candidate solution, a mutually constraining contradiction arises: on the one hand, distributing limited computational power across all candidate solutions to accommodate broad exploration reduces the simulation fidelity of each solution, leading to unreliable evaluation results and difficulty in early differentiation of solution merits; on the other hand, concentrating computational power on a few solutions to ensure simulation fidelity sacrifices broad exploration, potentially missing the final optimal solution, and continuously consuming significant computational power on solutions destined to deteriorate and be eliminated as the adversarial process evolves.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention discloses a method, apparatus, device, and storage medium for multi-branch parallel inference and decision optimization based on a large language model. It aims to solve the technical problem that under limited computing power and real-time constraints, it is difficult to keep computing power focused on truly valuable solutions, thus making it impossible to balance the breadth of exploration and the fidelity of simulation and quickly converge to the optimal decision.

[0007] The first embodiment of the present invention provides a multi-branch parallel inference and decision optimization method based on a large language model, including: The system receives real-time situational data from multiple sources and constructs a dynamically evolving battlefield state space. It performs situational analysis on the battlefield state space to identify sudden battlefield events and uses these events as triggering conditions to call a large language model to generate multiple differentiated decision branches, each containing macro-level objectives and micro-level strategies. An independent parallel battlefield simulation instance is constructed for each decision branch. Computational and simulation resources are allocated to each parallel battlefield simulation instance, and each parallel battlefield simulation instance is driven to perform synchronous and parallel simulations to obtain the real-time simulation status of each decision branch. Based on the real-time simulation status, a multi-dimensional dynamic evaluation is performed on each decision branch to obtain a dynamic comprehensive score for each decision branch; Dynamic pruning is performed based on the dynamic comprehensive score, terminating decision branches whose scores are continuously below a preset threshold, and redistributing the computing and simulation resources released by the terminated decision branches to the surviving branches according to their weights. Extract the high-scoring high-quality branches from the surviving branches, call the large language model to perform decision fusion on the key action sequences of the high-quality branches, and output the final optimal combat path and command sequence. The second embodiment of the present invention provides a multi-branch parallel inference and decision optimization device based on a large language model, comprising: The situation mapping module is used to receive multi-source real-time situation data and construct a dynamically evolving battlefield state space. It performs situation analysis on the battlefield state space to identify battlefield emergencies and uses the battlefield emergencies as triggering conditions to call a large language model to generate multiple differentiated decision branches, each containing macro-objectives and micro-strategies. The parallel battlefield module is used to construct an independent parallel battlefield simulation instance for each decision branch, allocate computing and simulation resources to each parallel battlefield simulation instance, and drive each parallel battlefield simulation instance to perform synchronous and parallel simulations to obtain the real-time simulation status of each decision branch. The dynamic evaluation module is used to perform multi-dimensional dynamic evaluation on each decision branch based on the real-time simulation status, and obtain a dynamic comprehensive score for each decision branch. The dynamic pruning module is used to perform dynamic pruning based on the dynamic comprehensive score, terminate decision branches whose scores are continuously lower than a preset threshold, and redistribute the computing and simulation resources released by the terminated decision branches to the surviving branches according to their weights. The fusion output module is used to extract the high-scoring high-quality branches among the surviving branches, call the large language model to perform decision fusion on the key action sequences of the high-quality branches, and output the final optimal combat path and instruction sequence. The third embodiment of the present invention provides a multi-branch parallel inference and decision optimization device based on a large language model, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the multi-branch parallel inference and decision optimization method based on a large language model as described in any of the above embodiments.

[0008] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by the processor of the device where the computer-readable storage medium is located, to implement the multi-branch parallel inference and decision optimization method based on a large language model as described in any of the above embodiments.

[0009] Beneficial effects: The large language model automatically generates multiple differentiated decision branches, each containing macro-level objectives and micro-level strategies, based on battlefield emergencies. It also constructs independent parallel battlefield simulation instances for each branch for synchronous and parallel simulation, expanding decision-making from a single path to a parallel exploration of the decision space. This allows multiple candidate solutions to be verified simultaneously under the same situation, avoiding the problem of missing better solutions due to a single decision path.

[0010] Based on the real-time simulation status, multi-dimensional dynamic evaluation is performed on each branch, and dynamic pruning is performed on branches that are continuously below the preset threshold according to the score. The released computing and simulation resources are then redistributed to the surviving branches according to their weights. This allows limited computing power to continuously converge on high-value branches as the value of the branches changes dynamically, thereby balancing the breadth of exploration and the fidelity of simulation under the condition of limited computing power, and significantly improving the simulation efficiency and convergence speed.

[0011] Extracting surviving high-quality branches and using a large language model to perform decision fusion on their key action sequences, outputting the optimal combat path and command sequence that combines the advantages of multiple high-quality solutions, realizing the transformation from "manually setting single path deduction" to "autonomous decision convergence", which can quickly and reliably form the optimal decision solution. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a multi-branch parallel inference and decision optimization method based on a large language model provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-branch parallel inference and decision optimization device based on a large language model provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] This invention discloses a method, apparatus, device, and storage medium for multi-branch parallel inference and decision optimization based on a large language model. It aims to solve the technical problem that under limited computing power and real-time constraints, it is difficult to keep computing power focused on truly valuable solutions, thus making it impossible to balance the breadth of exploration and the fidelity of simulation and quickly converge to the optimal decision.

[0016] Please see Figure 1 The first embodiment of the present invention provides a multi-branch parallel inference and decision optimization method based on a large language model, which can be executed by an optimization device (or system), specifically by one or more processors within the optimization device, to at least implement the following steps: S101, receive multi-source real-time situational data and construct a dynamically evolving battlefield state space, perform situational analysis on the battlefield state space to identify battlefield emergencies, and use the battlefield emergencies as triggering conditions to call a large language model to generate multiple differentiated decision branches, each containing macro-objectives and micro-strategies. The system first receives real-time situational data from multiple data sources via a situational mapping module. These data sources include the position, status, and movement information of friendly and enemy combat units, sensor detection information, target information, and battlefield environment information. The system then performs spatiotemporal alignment and unified representation on the heterogeneous data to construct a continuously updated and dynamically evolving battlefield state space that reflects the current confrontation situation and its evolution in real time. Based on this, the system performs situational analysis and feature extraction on the battlefield state space, extracting key features representing the confrontation situation from the continuously updated situational data, and identifying battlefield emergencies accordingly. The identification of battlefield emergencies is based on preset triggering conditions, including at least one of the following: the appearance of a new enemy entity, friendly casualties exceeding a preset threshold (this preset threshold is a trigger threshold set for friendly casualties), and obstruction of the target area. That is, when the situational analysis result meets any of the above triggering conditions, the corresponding battlefield emergency is determined to have occurred and is used as the trigger signal for subsequent branch generation.

[0017] Upon identifying a battlefield emergency, the system invokes a large language model, using this emergency as a trigger. The large language model automatically generates multiple differentiated decision branches based on the current battlefield state space, rather than relying on a pre-defined single solution. Each decision branch comprises two levels: macro-level objectives and micro-level strategies. At the macro-level objectives, different branches are generated for different response orientations. These macro-level objectives include at least one of the following: prioritizing awareness, responding to threats, seizing opportunities, prioritizing loss mitigation, restoring capabilities, and regaining control. This ensures that the generated decision branches are distinguished by their overall intent and cover different response directions. At the micro-level strategies, the macro-level objectives of each decision branch are mapped downwards to... The simulation engine can execute specific task strategy nodes, which include at least one of the following: control of the on / off states of sensors and countermeasures equipment (e.g., opening and closing of electronic jamming switches and detection switches), control of the maneuver and formation following of combat units, and target-oriented task execution instructions. During the generation process, the large language model outputs the task strategy nodes corresponding to each decision branch point by point in a streaming output manner. Each time a decision point is output, the branch generation logic of the application is called to complete the corresponding force allocation and task creation. This translates the semantic strategy output by the large language model into specific tasks that can be executed in parallel battlefield simulation instances, and provides input for building independent parallel battlefield simulation instances for each decision branch in subsequent steps.

[0018] As a more concrete example, when the situation analysis identifies a newly emerging high-value threat target in the battlefield state space, i.e., when the triggering condition for the emergence of a new enemy entity is met, the system triggers the large language model with this battlefield event and generates several differentiated decision branches for the threat. For example, there is a cognitive priority branch with the macro-purpose of obtaining target parameters, which corresponds to sending out detection units to obtain information at close range; a threat response branch with the macro-purpose of suppressing the threat, which corresponds to activating electronic jamming switches for long-range suppression; and a loss mitigation priority branch with the macro-purpose of risk avoidance, which corresponds to adjusting maneuver routes to detour. The macro-purpose of each of the above decision branches is mapped down to task strategy nodes such as detection switch control, electronic jamming switch control, and unit maneuver control, and is generated point by point by the large language model in a streaming output manner. The application-side branch generation logic completes the force allocation and task creation, thereby obtaining multiple differentiated decision branches with mutually distinguishable intentions that can be put into parallel simulation.

[0019] S102, construct an independent parallel battlefield simulation instance for each decision branch, allocate computing and simulation resources to each parallel battlefield simulation instance, and drive each parallel battlefield simulation instance to perform synchronous and parallel simulations to obtain the real-time simulation status of each decision branch. After generating multiple differentiated decision branches from the large language model, the system constructs an independent parallel battlefield simulation instance for each decision branch via a parallel battlefield cluster. This ensures that each decision branch is mapped to an independently running simulation instance, with each simulation instance isolated from and not interfering with each other. This allows for the simulation and verification of the solutions corresponding to different decision branches under the same initial situation. When constructing each parallel battlefield simulation instance, the system uses the battlefield state space as the unified initial situation for each simulation instance and loads the task strategy nodes corresponding to each decision branch at the micro-strategy level into the corresponding simulation instance. This enables each simulation instance to advance the simulation according to the macro-objective and micro-strategy of the decision branch it carries. After construction, the system allocates computing and simulation resources to each parallel battlefield simulation instance. These resources include CPU, GPU, memory, and simulation engine resources to support the operation of the simulation instances. At the beginning of the simulation, initial resources can be allocated equally to each simulation instance so that each decision branch can conduct simulations under the same conditions. Subsequently, the system drives each parallel battlefield simulation instance to conduct simulations in parallel in a synchronous manner. That is, each simulation instance advances synchronously on a unified simulation timeline according to a preset simulation time step, simulating the situational evolution process of its respective decision branch under the given scheme, and outputting the simulation progress of each decision branch in real time. This results in the real-time simulation status of each decision branch. The real-time simulation status is used to characterize the situation and execution status of the corresponding decision branch at each moment during the simulation process, and serves as the input for multi-dimensional dynamic evaluation of each decision branch in subsequent steps.

[0020] To enhance the realism of the simulation process, during the synchronous parallel simulation of each parallel battlefield instance, the system sets up an adversarial agent in each instance to represent the enemy entity. This transforms the enemy entity from a passive recipient of our actions into an autonomous participant with decision-making and counter-measure capabilities. Specifically, the adversarial agent in each parallel battlefield instance continuously senses our real-time situation within that instance and autonomously judges and adjusts its strategy and implements corresponding counter-measures based on the sensed situation. This creates a continuous interactive and mutually influential adversarial process between us and our side in that instance, ensuring that each decision branch is verified under the condition of active enemy confrontation. Consequently, the resulting real-time simulation state more realistically reflects the feasibility and effectiveness of the corresponding decision scheme in actual confrontation. Since each parallel battlefield simulation instance is independent of the others and each has its own adversarial agents, the adversarial agents in different simulation instances can take targeted countermeasures against the different decision branches adopted by our side in their respective simulation instances. This allows the differences in the advantages and disadvantages of each decision branch when facing the enemy's dynamic confrontation to gradually emerge during the simulation process, providing a basis for subsequent dynamic evaluation and dynamic pruning.

[0021] As a more concrete example, after multiple differentiated decision branches such as cognitive priority, threat response, and loss mitigation priority generated in response to the same battlefield emergency are constructed as independent parallel battlefield simulation instances and initial calculation and simulation resources are evenly distributed, each simulation instance proceeds synchronously and in parallel with a unified simulation time step. Within each simulation instance, the adversarial agent takes different countermeasures based on our actions in that instance: for the cognitive priority branch which employs close-range detection, the adversarial agent can autonomously take interceptive countermeasures against the approaching detection unit; for the threat response branch which employs long-range electronic suppression, the adversarial agent can autonomously adjust its detection methods to counter our suppression; for the loss mitigation priority branch which employs detour and avoidance, the adversarial agent can autonomously relocate its position and implement concealment. This allows each decision branch to evolve into different real-time simulation states under the enemy's active countermeasures, providing a basis for subsequent multi-dimensional dynamic evaluation of each decision branch.

[0022] S103, Based on the real-time simulation status, perform multi-dimensional dynamic evaluation on each decision branch to obtain a dynamic comprehensive score for each decision branch; After driving the synchronous and parallel simulation of various parallel battlefield instances and continuously outputting the real-time simulation status of each decision branch, the system, through a dynamic evaluation and pruning center, conducts multi-dimensional dynamic evaluations of each decision branch according to a preset evaluation cycle during the simulation process. This means that the evaluation is not a one-time event after the simulation ends, but rather a quantitative evaluation based on the current real-time simulation status of each decision branch at the corresponding moment of each evaluation cycle, as the simulation progresses. This allows the strengths and weaknesses of each decision branch to be dynamically and continuously characterized as the confrontation situation evolves. Specifically, the system quantitatively analyzes the real-time simulation status of each decision branch from multiple evaluation dimensions. These dimensions include at least two of the following: mission achievement rate, casualty exchange ratio, resource utilization rate, and time efficiency. For each decision branch, the system calculates its quantitative value on each evaluation dimension in each evaluation cycle, then integrates the quantitative values ​​of each dimension and takes into account operational risks to obtain a dynamic comprehensive score for that decision branch at the corresponding moment.

[0023] When performing multi-dimensional dynamic evaluation on each decision branch and obtaining a dynamic comprehensive score, for each decision branch... The system calculates its value using the following formula. Dynamic comprehensive score at any time :

[0024] in, The number of evaluation dimensions; For decision branches In the The quantitative values ​​on each evaluation dimension are used to characterize the current performance of the decision branch on the corresponding evaluation dimension. For example, the quantitative value of the mission achievement rate dimension characterizes the degree to which the decision branch has advanced the mission completion; the quantitative value of the battle loss exchange ratio dimension characterizes the ratio of the losses to the results of the decision branch; the quantitative value of the resource utilization rate dimension characterizes the efficiency of the consumption and utilization of combat resources by the decision branch in the execution of the plan; and the quantitative value of the time efficiency dimension characterizes the quality of the time taken by the decision branch to achieve the goal. For the first The system's basic weights for each evaluation dimension are used to reflect the default relative importance of each evaluation dimension in the scoring; For the first The preference weight coefficients for each evaluation dimension are used to amplify or reduce the system's basic weights for the corresponding dimensions. For decision branches exist The action risk penalty item at any moment is used to quantify and deduct the action risk currently faced by the decision branch; The risk penalty coefficient is used to adjust the weight of the action risk penalty item in the dynamic comprehensive score. As can be seen from the above formula, the dynamic comprehensive score of each decision branch is obtained by subtracting the action risk penalty item adjusted by the risk penalty coefficient from the sum of its weighted quantitative values ​​in each evaluation dimension. Therefore, it can comprehensively reflect the overall performance of each decision branch in multiple dimensions, as well as the accompanying action risks, and it is dynamically updated as the real-time simulation status of each decision branch changes.

[0025] To enable the dynamic comprehensive score to adapt to different decision-making orientations, the preference weight coefficient... It is not fixed, but rather determined according to the received preference configuration parameters and a preset weight adjustment rule. That is, after receiving the preference configuration parameters representing the current decision orientation, the system maps these parameters to preference weight coefficients for each evaluation dimension according to the pre-set adjustment rules. This dynamically amplifies or reduces the system's basic weights for the corresponding dimensions, causing the dynamic comprehensive score to tilt towards the preferred dimension. Specifically, the weight adjustment rule includes: in response to the survival priority configuration parameter, the system increases the preference weight coefficient for the battle loss exchange ratio dimension and simultaneously increases the risk penalty coefficient. This system adjusts the preference weighting coefficient for the task achievement rate dimension in response to the task achievement priority parameter. This results in higher scores for decision branches with higher task completion rates, thus favoring decision branches that are more likely to achieve their goals. Similarly, it adjusts the preference weighting coefficient for the resource utilization rate dimension in response to the cost priority parameter. This results in higher scores for decision branches with higher resource utilization efficiency, thus favoring decision branches with more economical resource consumption. Because the preference weighting coefficients are determined in this way, the resulting dynamic comprehensive score can automatically adjust the relative weights of each dimension based on the received preference configuration parameters within the same scoring model, yielding an evaluation result that matches the current decision orientation.

[0026] As a more concrete example, when multiple differentiated decision branches generated in response to the same battlefield contingency are simultaneously and concurrently deployed into parallel battlefield simulation instances, the system continuously calculates the dynamic comprehensive score of each decision branch according to the evaluation cycle. When the system receives survival priority configuration parameters, it increases the preference weight coefficient of the battle loss exchange ratio dimension and increases the risk penalty coefficient according to the weight adjustment rules. If a decision branch experiences increased losses and a significant rise in action risk during the simulation, then its action risk penalty item will be applied. As the value increases, the corresponding quantification value decreases, causing the dynamic comprehensive score of the decision branch to decrease significantly at the evaluation point. On the other hand, the decision branch with less loss and more stable progress obtains a higher dynamic comprehensive score because of its higher quantification value in the corresponding dimension. This allows the differences in the merits of each decision branch to be continuously and dynamically characterized as the simulation progresses, and the dynamic comprehensive score of each decision branch is output to subsequent steps as the basis for performing dynamic pruning and calculation and simulation resource reallocation for each decision branch.

[0027] S104, Perform dynamic pruning based on the dynamic comprehensive score, terminate decision branches whose scores are continuously lower than a preset threshold, and redistribute the computing and simulation resources released by the terminated decision branches to the surviving branches according to their weights. After the dynamic evaluation and pruning center continuously calculates and outputs the dynamic comprehensive score of each decision branch according to the evaluation cycle, the system performs dynamic pruning on each decision branch accordingly. The purpose is to prevent limited computing and simulation resources from being consumed on decision branches with persistently low scores that have gradually shown disadvantages as the adversarial evolution progresses. Instead, resources are directed to converge on surviving branches with higher value as the value of each decision branch dynamically changes, thereby improving the overall inference efficiency and convergence speed under limited computing power. To avoid pruning recoverable decision branches due to instantaneous fluctuations in scores, the system does not terminate based on a single score. Instead, it adopts a pruning mechanism based on continuous threshold judgment. That is, the performance of each decision branch's score relative to a preset threshold is cumulatively judged over multiple consecutive cycles before deciding whether to terminate.

[0028] Specifically, during the dynamic pruning process, the system maintains a low-score count for each decision branch: in each evaluation cycle, when the dynamic comprehensive score of a decision branch is lower than the preset threshold, the system increments the low-score count for that decision branch; conversely, when the dynamic comprehensive score of the decision branch is not lower than the preset threshold, the low-score count for that decision branch is reset to zero. This allows the low-score count to represent the number of evaluation cycles in which the decision branch has been continuously in a low-score state. When the low-score count of a decision branch reaches a continuous... Each evaluation cycle, i.e., its dynamic comprehensive score is continuous. When the evaluation period is lower than the preset threshold, or when the decision branch meets the absolute stop-loss condition, the system determines that the decision branch should be terminated and terminates it. The absolute stop-loss condition is defined as a preset positive integer, whereby the decision branch incurs an irreversible loss reaching a preset level during the deduction process. In other words, when a decision branch experiences such a severely adverse situation where further deduction is no longer worthwhile, it is not necessary to wait for its low score count to accumulate to a continuous level. The decision branch can be terminated immediately after one evaluation cycle. Upon termination of a decision branch, the system reclaims the computing and simulation resources used by that branch, including CPU, GPU, memory, and simulation engine resources. The reclaimed resources are then pooled into a resource pool as allocable resources that can be reallocated to surviving branches.

[0029] After recovering resources from the terminated decision branches, the system redistributes the resources in the resource pool to the surviving branches according to their weights. Specifically, it allocates the available resources in the resource pool among the surviving branches based on their respective weights, ensuring that the surviving branches with higher scores and greater value receive more additional resources. For each surviving branch... The system determines the additional resources allocated to the surviving branch using the following formula. :

[0030] in, This refers to the total amount of allocatable resources in the resource pool. For the survival branch The allocation weights are determined based on its dynamic comprehensive score. The sum of the weights assigned to each surviving branch. To allocate to surviving branches The additional resources are calculated as follows: As can be seen from the above formula, the additional resources obtained by each surviving branch are proportional to the proportion of its allocated weight in the sum of the allocated weights of all surviving branches. This ensures that the resources released by the terminated branch are allocated skewed according to the value of the surviving branch. After obtaining additional resources, the corresponding surviving branch improves the solution accuracy of its simulation engine. This improvement includes reducing the time step of the surviving branch's inference, allowing it to advance the inference with finer time granularity and perform more refined solutions to its situational evolution. This improves the fidelity of the high-value branch's inference while accelerating its convergence, enabling the system to converge to the optimal decision scheme faster and more reliably.

[0031] As a more concrete example, in a parallel battlefield simulation scenario where multiple differentiated decision branches generated in response to the same battlefield emergency are simultaneously and concurrently simulated, and the system continuously calculates the dynamic comprehensive score of each decision branch according to the evaluation cycle, a certain decision branch may fall into a disadvantageous position during the confrontation, causing its dynamic comprehensive score to continuously fall below a preset threshold. When the cumulative count of its low scores reaches a certain threshold... The system terminates the process within one evaluation cycle; simultaneously, another decision branch meets the absolute stop-loss condition due to irreversible losses reaching a preset level during the simulation, without waiting for its low score count to accumulate to a continuous level. The system immediately terminates the evaluation cycle; the CPU, GPU, memory, and simulation engine resources occupied by the two terminated decision branches are reclaimed and added to the resource pool. The system then allocates the available resources in the resource pool according to the allocation weight determined by each surviving branch based on its dynamic comprehensive score. Resources are allocated to each surviving branch, with higher-scoring branches receiving more additional resources. This reduces their simulation time step and improves the accuracy of their simulation engine, allowing for higher fidelity and faster simulation. This provides a more reliable basis for subsequent steps to extract high-quality surviving branches and perform decision fusion.

[0032] S105, extract the high-scoring high-quality branches from the surviving branches, call the large language model to perform decision fusion on the key action sequences of the high-quality branches, and output the final optimal combat path and command sequence. After the dynamic evaluation and pruning center continuously performs dynamic pruning and reallocation of computational and simulation resources, causing limited resources to converge on high-value surviving branches, decision branches with consistently low scores are gradually terminated as the simulation progresses. Meanwhile, decision branches with higher scores, which gradually demonstrate advantages as the adversarial evolution unfolds, are retained and continue to be simulated with higher solution accuracy until the simulation ends. After the synchronous parallel simulation concludes, the system does not directly select the single branch with the highest score as the final solution. Instead, it extracts several high-scoring, high-quality branches from the surviving branches through a fusion output module and integrates the advantages exhibited by these high-quality branches to obtain a final solution superior to any single branch.

[0033] Specifically, when extracting high-quality branches from the surviving branches and performing decision fusion, the system sorts the surviving branches according to their dynamic comprehensive scores after the simulation, and ranks the branches with the highest dynamic comprehensive scores. The surviving branches are designated as the high-quality branches. These high-quality branches are decision branches that maintain high scores after multiple cycles of simulation and dynamic evaluation under conditions of active enemy confrontation, each demonstrating corresponding advantages in different aspects. Subsequently, the system processes these high-quality branches through a multi-agent collaborative closed loop. This loop consists of a strategic planning agent, a tactical analysis agent, a resource scheduling agent, a risk assessment agent, and an adjudication agent. The strategic planning agent grasps the overall orientation of each high-quality branch from a global intent perspective. The tactical analysis agent analyzes the key actions and effects of each high-quality branch at the specific execution level. The resource scheduling agent evaluates the resource utilization arrangements of each high-quality branch. The risk assessment agent assesses the associated action risks of each high-quality branch. The adjudication agent coordinates and adjudicates the analysis results of each agent. These agents cooperate to form a collaborative closed loop, jointly extracting the key action sequences of each high-quality branch. The key action sequences are a sequence of key actions that play a decisive role in the simulation of the corresponding high-quality branch, arranged in their execution order.

[0034] After extracting the key action sequences of each high-quality branch, the system submits these sequences to a large language model for fusion reasoning. The large language model analyzes, coordinates, and reorganizes the key action sequences from different high-quality branches, identifying and organically combining the advantages of each branch to generate an optimal operational path and command sequence that incorporates the tactical advantages of multiple high-quality branches. The optimal operational path represents the overall action plan obtained after fusion, and the command sequence is an ordered set of specific commands that can be issued and executed corresponding to the optimal operational path. Compared to simply selecting a single branch, this optimal operational path achieves better overall results because it integrates the advantages of multiple high-quality branches in different aspects. Finally, the system outputs the optimal operational path and command sequence, thus completing a closed loop from receiving multi-source real-time situational data, through multi-branch parallel deduction and dynamic convergence, to outputting the optimal decision plan, realizing the transformation from single-path deduction to autonomous decision convergence.

[0035] As a more concrete example, after multiple differentiated decision branches generated in response to the same battlefield emergency are simultaneously and in parallel through parallel battlefield simulations, and undergo multi-cycle dynamic evaluation and pruning, assuming that two surviving branches with higher scores still emerge victorious at the end of the simulation—one demonstrating an advantage in suppressing the enemy, and the other demonstrating an advantage in carrying out the mission against the target—the system will rank these two surviving branches as the top performers in the dynamic comprehensive score. The optimal branch is formed by a multi-agent collaborative closed loop consisting of a strategic planning agent, a tactical analysis agent, a resource scheduling agent, a risk assessment agent, and an adjudication agent. The key action sequences of the two optimal branches are extracted and then fed into a large language model for fusion reasoning. This generates the optimal combat path that uses the suppression action of the former as cover to guide the mission action of the latter in a coordinated manner. Based on this, the corresponding instruction sequence is output and issued, thus obtaining a final decision scheme that combines the advantages of the two optimal branches.

[0036] Please see Figure 2 The second embodiment of the present invention provides a multi-branch parallel inference and decision optimization device based on a large language model, comprising: The situation mapping module 201 is used to receive multi-source real-time situation data and construct a dynamically evolving battlefield state space, perform situation analysis on the battlefield state space to identify battlefield emergencies, and use the battlefield emergencies as triggering conditions to call a large language model to generate multiple differentiated decision branches, each containing macro-objectives and micro-strategies. Parallel battlefield module 202 is used to construct an independent parallel battlefield simulation instance for each decision branch, allocate computing and simulation resources to each parallel battlefield simulation instance, and drive each parallel battlefield simulation instance to perform synchronous and parallel simulations to obtain the real-time simulation status of each decision branch. The dynamic evaluation module 203 is used to perform multi-dimensional dynamic evaluation on each decision branch based on the real-time simulation status, and obtain a dynamic comprehensive score for each decision branch. The dynamic pruning module 204 is used to perform dynamic pruning based on the dynamic comprehensive score, terminate decision branches whose scores are continuously lower than a preset threshold, and redistribute the computing and simulation resources released by the terminated decision branches according to their weights to the surviving branches. The fusion output module 205 is used to extract the high-scoring high-quality branches among the surviving branches, call the large language model to perform decision fusion on the key action sequences of the high-quality branches, and output the final optimal combat path and instruction sequence. The third embodiment of the present invention provides a multi-branch parallel inference and decision optimization device based on a large language model, including a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the multi-branch parallel inference and decision optimization method based on a large language model as described in any of the above embodiments.

[0037] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, which can be executed by the processor of the device where the computer-readable storage medium is located, to implement the multi-branch parallel inference and decision optimization method based on a large language model as described in any of the above embodiments.

[0038] Exemplary examples show that the computer program described in the third and fourth embodiments of the present invention can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in implementing a multi-branch parallel inference and decision optimization device based on a large language model. For example, the apparatus described in the second embodiment of the present invention.

[0039] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the aforementioned multi-branch parallel inference and decision optimization method based on a large language model, connecting various parts of the method through various interfaces and lines.

[0040] The memory can be used to store the computer program and / or modules. The processor, by running or executing the computer program and / or modules stored in the memory, and by calling the data stored in the memory, implements various functions of a multi-branch parallel inference and decision optimization method based on a large language model. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0041] If the implemented module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0042] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0043] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-branch parallel inference and decision optimization method based on a large language model, characterized in that, include: The system receives real-time situational data from multiple sources and constructs a dynamically evolving battlefield state space. It performs situational analysis on the battlefield state space to identify sudden battlefield events and uses these events as triggering conditions to call a large language model to generate multiple differentiated decision branches, each containing macro-level objectives and micro-level strategies. An independent parallel battlefield simulation instance is constructed for each decision branch. Computational and simulation resources are allocated to each parallel battlefield simulation instance, and each parallel battlefield simulation instance is driven to perform synchronous and parallel simulations to obtain the real-time simulation status of each decision branch. Based on the real-time simulation status, a multi-dimensional dynamic evaluation is performed on each decision branch to obtain a dynamic comprehensive score for each decision branch. Specifically, for each decision branch... Calculate using the following formula Dynamic comprehensive score at any time : in, The number of evaluation dimensions; For decision branches In the Quantitative values ​​on an evaluation dimension, wherein the evaluation dimension includes at least two of the following: mission achievement rate, battle loss exchange ratio, resource utilization rate, and time efficiency; For the first The system's basic weights for each evaluation dimension; For the first The preference weight coefficients for each evaluation dimension are used to amplify or reduce the corresponding basic weights of the system. For decision branches Action risk penalty items; The risk penalty coefficient, the preference weight coefficient In response to the received preference configuration parameters, and according to a preset weight adjustment rule, the following adjustments are made: in response to the survival priority configuration parameters, the preference weight coefficient of the battle loss exchange ratio dimension is increased, and the risk penalty coefficient is increased. In response to the task achievement priority configuration parameter, increase the preference weight coefficient of the task achievement rate dimension; in response to the cost priority configuration parameter, increase the preference weight coefficient of the resource utilization dimension. Dynamic pruning is performed based on the dynamic comprehensive score, terminating decision branches whose scores are continuously below a preset threshold, and redistributing the computational and simulation resources released by the terminated decision branches to the surviving branches according to their weights. Specifically: For each decision branch, when its dynamic comprehensive score is lower than the preset threshold, the low score count for that decision branch is incremented; otherwise, the low score count is reset to zero. When the low score count for a decision branch reaches a certain threshold... Once an evaluation cycle has elapsed, or when the decision branch meets the absolute stop-loss condition, the decision branch is terminated, and its occupied CPU, GPU, memory, and simulation engine resources are reclaimed to the resource pool. The absolute stop-loss condition is defined as an irreversible loss that reaches a preset level when the decision branch incurs a predetermined amount of loss. Based on the allocation weight of each surviving branch, the resources in the resource pool are added in an additional amount. Assigned to each surviving branch, where This refers to the total amount of allocatable resources in the resource pool. For the survival branch The allocation weights are determined based on its dynamic comprehensive score. To allocate to surviving branches Additional resources; The surviving branch that receives additional resources improves the solution accuracy of its simulation engine, and the improvement in solution accuracy includes reducing the time step of the surviving branch inference; Extract the high-scoring high-quality branches from the surviving branches, call the large language model to perform decision fusion on the key action sequences of the high-quality branches, and output the final optimal combat path and command sequence.

2. The multi-branch parallel inference and decision optimization method based on a large language model according to claim 1, characterized in that, The process of using the aforementioned battlefield emergency as a trigger condition to call a large language model to generate multiple differentiated decision branches, each containing macro-level objectives and micro-level strategies, is as follows: The occurrence of a new enemy entity, friendly casualties exceeding a preset threshold, or obstruction of the target area are used as the triggering conditions for the battlefield emergency. The macro-level objectives of the generated decision branches include at least one of the following: prioritizing cognition, responding to threats, seizing opportunities, prioritizing loss mitigation, restoring capabilities, and regaining control. Their micro-level strategies are mapped down to specific task strategy nodes, which include at least one of the following: interference switch control, detection switch control, suicide attack command, and entity follow command. These nodes are generated point by point by the large language model in a streaming output manner.

3. The multi-branch parallel inference and decision optimization method based on a large language model according to claim 1, characterized in that, The process of driving the synchronous and parallel simulation of each parallel battlefield instance is as follows: In each of the parallel battlefield simulation instances, an adversarial agent is set up to represent the enemy entity. The adversarial agent has autonomous decision-making and countermeasure capabilities, and adjusts its own strategy and implements countermeasures based on the real-time situation of our side in the parallel battlefield simulation instance, so that the parallel battlefield simulation instance presents a continuous game between the offensive and defensive sides.

4. The multi-branch parallel inference and decision optimization method based on a large language model according to claim 1, characterized in that, The step of extracting high-scoring high-quality branches from the surviving branches and then using a large language model to perform decision fusion on the key action sequences of these high-quality branches is as follows: After the simulation, the top performers will be ranked according to their dynamic comprehensive scores. The surviving branches are considered the high-quality branches; A multi-agent collaborative closed loop, consisting of a strategic planning agent, a tactical analysis agent, a resource scheduling agent, a risk assessment agent, and an adjudication agent, extracts the key action sequences of each high-quality branch and submits them to the large language model for fusion reasoning to generate the optimal combat path and instruction sequence that combines the tactical advantages of multiple high-quality branches.

5. A multi-branch parallel inference and decision optimization device based on a large language model, characterized in that, include: The situation mapping module is used to receive multi-source real-time situation data and construct a dynamically evolving battlefield state space. It performs situation analysis on the battlefield state space to identify battlefield emergencies and uses the battlefield emergencies as triggering conditions to call a large language model to generate multiple differentiated decision branches, each containing macro-objectives and micro-strategies. The parallel battlefield module is used to construct an independent parallel battlefield simulation instance for each decision branch, allocate computing and simulation resources to each parallel battlefield simulation instance, and drive each parallel battlefield simulation instance to perform synchronous and parallel simulations to obtain the real-time simulation status of each decision branch. The dynamic evaluation module is used to perform multi-dimensional dynamic evaluation of each decision branch based on the real-time simulation status, and obtain a dynamic comprehensive score for each decision branch. Specifically, for each decision branch... Calculate using the following formula Dynamic comprehensive score at any time : in, The number of evaluation dimensions; For decision branches In the Quantitative values ​​on an evaluation dimension, wherein the evaluation dimension includes at least two of the following: mission achievement rate, battle loss exchange ratio, resource utilization rate, and time efficiency; For the first The system's basic weights for each evaluation dimension; For the first The preference weight coefficients for each evaluation dimension are used to amplify or reduce the corresponding basic weights of the system. For decision branches Action risk penalty items; The risk penalty coefficient, the preference weight coefficient In response to the received preference configuration parameters, and according to a preset weight adjustment rule, the following adjustments are made: in response to the survival priority configuration parameters, the preference weight coefficient of the battle loss exchange ratio dimension is increased, and the risk penalty coefficient is increased. In response to the task achievement priority configuration parameter, increase the preference weight coefficient of the task achievement rate dimension; in response to the cost priority configuration parameter, increase the preference weight coefficient of the resource utilization dimension. The dynamic pruning module is used to perform dynamic pruning based on the dynamic comprehensive score, terminating decision branches whose scores are continuously below a preset threshold, and redistributing the computing and simulation resources released by the terminated decision branches to the surviving branches according to their weights. Specifically: For each decision branch, when its dynamic comprehensive score is lower than the preset threshold, the low score count for that decision branch is incremented; otherwise, the low score count is reset to zero. When the low score count for a decision branch reaches a certain threshold... Once an evaluation cycle has elapsed, or when the decision branch meets the absolute stop-loss condition, the decision branch is terminated, and its occupied CPU, GPU, memory, and simulation engine resources are reclaimed to the resource pool. The absolute stop-loss condition is defined as an irreversible loss that reaches a preset level when the decision branch incurs a predetermined amount of loss. Based on the allocation weight of each surviving branch, the resources in the resource pool are added in an additional amount. Assigned to each surviving branch, where This refers to the total amount of allocatable resources in the resource pool. For the survival branch The allocation weights are determined based on its dynamic comprehensive score. To allocate to surviving branches Additional resources; The surviving branch that receives additional resources improves the solution accuracy of its simulation engine, and the improvement in solution accuracy includes reducing the time step of the surviving branch inference; The fusion output module is used to extract the high-scoring high-quality branches among the surviving branches, call the large language model to perform decision fusion on the key action sequences of the high-quality branches, and output the final optimal combat path and instruction sequence.

6. A multi-branch parallel inference and decision optimization device based on a large language model, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be executed by the processor to implement a multi-branch parallel inference and decision optimization method based on a large language model as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The device contains a computer program that can be executed by a processor of the device where the computer-readable storage medium is located, to implement the multi-branch parallel inference and decision optimization method based on a large language model as described in any one of claims 1 to 4.

Images