Simulation scene generation method and device based on large language model

Abstract

The invention discloses a simulation scene generation method and device based on a large language model, and the method comprises the steps: S1, constructing a fine tuning model based on a preset first large language model and a second large language model; s2, acquiring a user scene description text; s3, processing the user scene description text by utilizing the fine tuning model to obtain an MSDL simulation text set; and S4, performing simulation evaluation processing on the MSDL simulation text set to obtain an optimal evaluation value and an optimal MSDL text. According to the method, deep reasoning can be carried out on a high-level natural language intention in combination with domain knowledge, and a standardized and executable optimal simulation scene and a corresponding evaluation value are automatically generated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of scenario simulation, and particularly relates to a simulation scenario generation method and device based on a large language model. BACKGROUND

[0002] In the current deep integration of artificial intelligence, natural language processing and computer simulation technology, the construction precision and efficiency of simulation scenarios in the fields of military deduction, unmanned system cluster control and city traffic simulation are required to be higher. At present, the scene construction in these fields highly depends on structured languages such as the Coalition Battle Management Language (C-BML) based on XML and various private JSON and LUA scripts. However, the existing technology has many deep problems in actual application, which has become a key bottleneck restricting the performance improvement of simulation systems.

[0003] Firstly, the cognitive gap problem is prominent. The tactical intent proposed by commanders or field experts is a high-level expression with fuzzy and dynamic characteristics, while the machine language recognizable by the simulation system is low-level, accurate and static code form. In the process of converting "tactical intent" into "machine executable code", there is a lack of effective bridge to bridge the difference between the two, which makes it difficult for the real needs of experts to be accurately transmitted to the simulation system, greatly increasing the communication and conversion cost of scene construction.

[0004] Secondly, the scene arrangement faces the problems of combination explosion and rigidity. Complex simulation scenarios need to integrate a large number of entities, events, rules and time and space constraints. When manually writing structured languages, staff need to handle the combination logic of a large number of elements, which is easy to miss. And once the writing is completed, the scene parameters and logic are fixed and cannot quickly respond to tactical intent adjustment or carry out "what-if analysis", making it difficult to adapt to the dynamic changes of simulation requirements.

[0005] Thirdly, it is difficult to effectively carry the tacit knowledge. Simulation scenario description often relies on the experiential tacit knowledge of operators, such as the default combat mode of specific equipment, the traffic rules in different terrains, etc. The existing structured language cannot standardize the expression and storage of such tacit knowledge, and if it is described redundantly through code, it will result in a bloated scene file, and if it is omitted, it will cause information loss, affecting the accuracy and rationality of the simulation results.

[0006] Finally, the lack of intention reasoning capability. Existing simulation tools can only execute explicit instructions given by the user, and cannot "understand" the deep intention behind the instructions. For example, when the user proposes the requirement of "suppressing the enemy air defense system", the tool cannot automatically infer that the task may involve a series of tactical action combinations such as "electronic jamming", "hard-kill strike" and "evasive maneuver", and still needs the user to disassemble the task one by one and write the corresponding code, which seriously limits the automation level and intelligent degree of simulation scene construction.

[0007] In summary, there is an urgent need for a technical solution that can accurately understand high-level natural language intention, combine domain knowledge for deep reasoning, and finally automatically generate standardized and executable simulation scenes, to break through the technical bottlenecks of existing technologies and promote the development of the simulation field to a higher level of intelligence. SUMMARY

[0008] The technical problem to be solved by the present application is to provide a simulation scene generation method and device based on a large language model, which can accurately understand high-level natural language intention, combine domain knowledge for deep reasoning, and automatically generate standardized and executable optimal simulation scenes and corresponding evaluation values.

[0009] To solve the above technical problems, the present application embodiment discloses a simulation scene generation method based on a large language model, which comprises:

[0010] S1, constructing a fine-tuning model based on a preset first large language model and a second large language model;

[0011] S2, obtaining a user scene description text;

[0012] S3, processing the user scene description text using the fine-tuning model to obtain a MSDL simulation text set; the MSDL simulation text set comprises M MSDL texts; M is an integer greater than 1;

[0013] S4, performing simulation evaluation processing on the MSDL simulation text set to obtain an optimal evaluation value and an optimal MSDL text.

[0014] As an optional implementation, in the first aspect of the present application embodiment, the fine-tuning model is constructed based on the preset first large language model and the second large language model, which comprises:

[0015] S11, presetting an MSDL text training set; the MSDL text training set comprises a plurality of MSDL texts;

[0016] S12, training the MSDL text training set by using the first large language model to obtain a scene description text training set; the scene description text training set comprises a scene description text corresponding to each MSDL text in the MSDL text training set;

[0017] S13, fine-tuning the second large language model based on the MSDL text training set and the scene description text training set to obtain the fine-tuning model.

[0018] As an optional implementation, in the first aspect of the embodiment of the present application, the MSDL text comprises an entity layer, a timing layer, a trigger layer and a target layer;

[0019] The entity layer comprises a plurality of entities; the entity comprises an entity ID field, a main type field, a sub-type field, an attribute sequence and an initial state sequence; the attribute sequence comprises a plurality of attribute fields; and the initial state sequence comprises a plurality of initial state fields.

[0020] The timing layer comprises a plurality of simulation events; the simulation event comprises an event ID field, the entity ID field, a start time field, an entity action and / or a dependent event field; the entity action comprises an action type field and / or a plurality of action attribute fields.

[0021] The trigger layer comprises a plurality of triggers; the trigger comprises a trigger ID field, a trigger condition and an action sequence; the trigger condition comprises a condition type field and a plurality of trigger attribute fields; and the action sequence comprises a plurality of entity actions.

[0022] The target layer comprises a plurality of simulation targets; the simulation target comprises a target ID field, a target type field, a readable description field, a completion condition, a score type and a target score; the completion condition comprises a condition type field, the entity ID field, a detection attribute field, an operator field and a target value field; and the value of the score type is 0 or 1.

[0023] As an optional implementation, in the first aspect of the embodiment of the present application, the processing of the user scene description text by using the fine-tuning model to obtain a MSDL simulation text set comprises:

[0024] S31, clarifying the user scene description text by using the fine-tuning model to obtain a clarified scene description text;

[0025] S32, performing initial conversion processing on the clarified scene description text by using the fine-tuning model to obtain an initial MSDL text;

[0026] S33. Using the fine-tuning model, the initial MSDL text is randomly replaced to obtain the MSDL simulation text set.

[0027] As an optional implementation, in the first aspect of the present invention, the step of using the fine-tuning model to clarify the user scenario description text to obtain clarified scenario description text includes:

[0028] S311. Using the fine-tuning model, perform fuzzy detection on the user scenario description text to obtain a fuzzy text set; the fuzzy text set includes several fuzzy texts.

[0029] S312. Based on the fuzzy text set, query a preset knowledge base to obtain a set of associated text sequences; the set of associated text sequences includes the associated text sequence corresponding to each fuzzy text in the fuzzy text set; the associated text sequence includes N associated texts; N is an integer greater than 1;

[0030] S313. Using the fine-tuning model, perform multiple rounds of confirmation processing on the fuzzy text set and the associated text sequence set to obtain the user confirmation text set.

[0031] S314. Using the fine-tuning model, based on the user confirmation text set, the user scenario description text is corrected to obtain the clarified scenario description text.

[0032] As an optional implementation, in the first aspect of the present invention, the step of using the fine-tuning model to perform multiple rounds of confirmation processing on the fuzzy text set and the associated text sequence set to obtain a user confirmation text set includes:

[0033] S3131. Set the initial confirmation text set to an empty set;

[0034] S3132. Set the current fuzzy text to any of the fuzzy texts in the fuzzy text set; set the current associated text sequence to the associated text sequence corresponding to the current fuzzy text in the associated text sequence set;

[0035] S3133. Using the fine-tuning model, perform user confirmation processing on the current fuzzy text and the current associated text sequence to obtain user confirmation text; add the user confirmation text to the user confirmation text set.

[0036] S3134. Delete the current fuzzy text from the fuzzy text set to obtain the fuzzy text set after deletion;

[0037] S3135, repeating S3132-S3134 until the set of fuzzy texts after deletion is an empty set;

[0038] S3136, determining the initial set of confirmed texts as the set of user confirmed texts.

[0039] As an optional implementation, in the first aspect of the embodiment of the present application, the simulation evaluation processing on the set of MSDL simulation texts to obtain an optimal evaluation value and an optimal MSDL text comprises:

[0040] S41, using a preset simulation engine to respectively simulate the M MSDL texts in the set of MSDL simulation texts to obtain a set of simulation information;

[0041] The set of simulation information comprises M simulation information; the simulation information comprises simulation time and a plurality of target completion information; the target completion information comprises a target completion flag and completion time; the value of the target completion flag is 0 or 1;

[0042] S42, using a simulation evaluation model to process the set of simulation information to obtain a set of simulation evaluation values; the set of simulation evaluation values comprises M simulation evaluation values;

[0043] S43, setting the optimal evaluation value as the largest simulation evaluation value in the set of simulation evaluation values;

[0044] S44, setting the optimal sequence number p as the sequence number of the optimal evaluation value in the set of simulation evaluation values;

[0045] S45, setting the optimal MSDL text as the pth MSDL text in the set of MSDL simulation texts.

[0046] The second aspect of the embodiment of the present application discloses a simulation scenario generation device based on a large language model, which comprises a model construction module, a text acquisition module, a simulation text generation module and a simulation evaluation module;

[0047] The model construction module is configured to construct a fine-tuning model based on a preset first large language model and a second large language model;

[0048] The text acquisition module is configured to acquire a user scenario description text;

[0049] The simulation text generation module is configured to process the user scenario description text using the fine-tuning model to obtain a set of MSDL simulation texts;

[0050] The simulation evaluation module is configured to perform simulation evaluation processing on the MSDL simulation text set to obtain an optimal evaluation value and an optimal MSDL text.

[0051] A third aspect of the embodiment of the present application discloses another simulation scenario generation device based on a large language model, and the device comprises:

[0052] a memory storing executable program codes;

[0053] a processor coupled to the memory;

[0054] The processor invokes the executable program codes stored in the memory to perform part or all of the steps of the simulation scenario generation method based on a large language model disclosed in the first aspect of the embodiment of the present application.

[0055] The fourth aspect of the present application discloses a computer storage medium storing computer instructions, which are invoked to perform part or all of the steps of the simulation scenario generation method based on a large language model disclosed in the first aspect of the embodiment of the present application.

[0056] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0057] The fine-tuning model obtained by the present application processes the user scenario description text to obtain a MSDL simulation text set, and further processes the simulation evaluation to obtain an optimal evaluation value and an optimal MSDL text, which can automatically generate a standardized and executable optimal simulation scenario and a corresponding evaluation value by deep reasoning of high-level natural language intention combined with domain knowledge. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0059] Figure 1 is a flowchart of a simulation scenario generation method based on a large language model disclosed in the embodiment of the present application.

[0060] Figure 2 is a structural diagram of a simulation scenario generation device based on a large language model disclosed in the embodiment of the present application.

[0061] Figure 3 is a structural diagram of another simulation scenario generation device based on a large language model disclosed in the embodiment of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0063] In the description of the present application, it should be noted that the orientations or positional relationships indicated by the terms 'center', 'upper', 'lower', 'left', 'right','vertical', 'horizontal', 'inner', 'outer' and the like are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms 'first','second', 'third' are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0064] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms'mounting', 'connecting', 'connection' should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0065] Embodiment one

[0066] Please refer to Figure 1 . Figure 1 is a flowchart of a simulation scene generation method based on a large language model disclosed by the embodiments of the present application. Among them, Figure 1 The simulation scene generation method based on a large language model described is applied to the generation of simulation scenes in the field of scenario simulation, such as military deduction, unmanned system cluster control, city traffic simulation, etc. The embodiments of the present application are not limited. As shown in Figure 1 The method comprises the following steps:

[0067] S1, based on the first large language model and the second large language model, a fine-tuning model is constructed.

[0068] It should be noted that the first large language model can be qwen-max-2025-08, Doubao-pro-2408, Hunyuanpro-2408, or ERNIE 4.0Turbo, and the like, which are commonly used large language models, and the embodiments of the present application are not limited. The second large language model is any commonly used large language model different from the first large language model.

[0069] S2, obtaining a user scenario description text.

[0070] It should be noted that the user scenario description text describes the specific content of the simulation task in a natural language manner, including entities, event timing and trigger conditions, and scoring methods of simulation results, and the like.

[0071] S3, processing the user scenario description text by using the fine-tuning model to obtain a MSDL simulation text set; the MSDL simulation text set includes M MSDL texts; M is an integer greater than 1.

[0072] S4, performing simulation evaluation processing on the MSDL simulation text set to obtain an optimal evaluation value and an optimal MSDL text.

[0073] In an optional embodiment, the fine-tuning model is constructed based on the preset first large language model and the second large language model, including:

[0074] S11, presetting a MSDL text training set; the MSDL text training set includes a plurality of MSDL texts.

[0075] S12, generating processing the MSDL text training set by using the first large language model to obtain a scenario description text training set; the scenario description text training set includes a scenario description text corresponding to each MSDL text in the MSDL text training set.

[0076] It should be noted that the generation processing is to convert all MSDL texts in the MSDL text training set into corresponding scenario description texts in natural language form by using the first large language model. Since the first large language model and the second large language model use different models, the diversity of the generated scenario description texts can be enhanced.

[0077] S13, fine-tuning training the second large language model based on the MSDL text training set and the scenario description text training set to obtain the fine-tuning model.

[0078] Optionally, the fine-tuning training uses a LoRA algorithm, wherein each MSDL text and the corresponding scene description text in the MSDL text training set form a sample pair for training, and the objective function expression of the training is:

[0079]

[0080] In the formula, Loss is the loss function value; NN is the number of sample pairs, that is, the number of MSDL texts in the MSDL text training set; sim is a text similarity calculation function; ta i is the scene description text corresponding to the i-th MSDL text; tb i is the target text obtained by converting the i-th MSDL text through the second large language model, wherein the conversion processing refers to converting the MSDL text into a natural language description target text. It can be seen that the above loss function value measures the average similarity between the target text obtained by converting all MSDL texts through the second large language model and the corresponding scene description text.

[0081] In another optional embodiment, the above-mentioned MSDL text includes an entity layer, a timing layer, a trigger layer, and a target layer.

[0082] The above-mentioned entity layer includes a plurality of entities; the above-mentioned entity includes an entity ID field, a main type field, a sub-type field, an attribute sequence, and an initial state sequence; the above-mentioned attribute sequence includes a plurality of attribute fields; and the above-mentioned initial state sequence includes a plurality of initial state fields.

[0083] The above-mentioned timing layer includes a plurality of simulation events; the above-mentioned simulation event includes an event ID field, the above-mentioned entity ID field, a start time field, an entity action, and / or a dependent event field; and the above-mentioned entity action includes an action type field and / or a plurality of action attribute fields.

[0084] The above-mentioned trigger layer includes a plurality of triggers; the above-mentioned trigger includes a trigger ID field, a trigger condition, and an action sequence; the above-mentioned trigger condition includes a condition type field and a plurality of trigger attribute fields; and the above-mentioned action sequence includes a plurality of the above-mentioned entity actions.

[0085] The above-mentioned target layer includes a plurality of simulation targets; the above-mentioned simulation target includes a target ID field, a target type field, a readable description field, a completion condition, a score type, and a target score; the above-mentioned completion condition includes a condition type field, the above-mentioned entity ID field, a detection attribute field, an operator field, and a target value field; and the value of the above-mentioned score type is 0 or 1.

[0086] It should be noted that the above MSDL (Mission Scenario Description Language) text, i.e., the simulation scenario file, is designed in accordance with the following principles: (1) Layering and decoupling: The scenario description is divided into entity layer (Entities), timeline layer (Timeline), event / trigger layer (Triggers) and goal layer (Goals), realizing the decoupling of static resources and dynamic behavior. (2) Event-driven: The core logic is driven by events and triggers, which can describe linear logic that occurs sequentially in time, as well as complex nonlinear logic that can be flexibly described as "when [condition] is met, perform [action]." (3) Extensibility: There are open definitions of entity and action types, which can easily introduce new equipment, sensors or tactical actions without modifying the language core. (4) Human-readable: Using formats similar to JSON or YAML, it is easy for machines to parse and has good human readability, making it easy to debug and verify.

[0087] It should be noted that when the value of the above score type is 0, the corresponding simulation goal is a positive goal, and when the simulation goal is achieved, the total simulation score is added to the corresponding goal score; when the value of the above score type is 1, the corresponding simulation goal is a negative goal, and when the simulation goal fails, the total simulation score is subtracted from the corresponding goal score.

[0088] The following gives a complete example of the MSDL text of an air defense suppression task, which includes the complete entity layer, timeline layer, trigger layer and goal layer:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] In yet another optional embodiment, the above utilizes the above fine-tuning model to process the above user scenario description text to obtain a set of MSDL simulation texts, including:

[0095] S31, utilizing the above fine-tuning model to clarify the above user scenario description text to obtain a clarified scenario description text.

[0096] S32, utilizing the above fine-tuning model to perform initial conversion processing on the above clarified scenario description text to obtain an initial MSDL text.

[0097] Optionally, the fine-tuning model is used to perform initial conversion processing on the clarified scene description text to obtain initial MSDL text, the fine-tuning model is set as an initial conversion agent by a preset initial conversion prompt, and then the initial conversion agent is used to process the clarified scene description text to obtain the initial MSDL text. The initial conversion prompt can be in the following form, where “[example text]” is the MSDL text of the air defense suppression task:

[0098] “You are now a converter, please convert [clarified scene description text] into initial MSDL text, and the reference example of the initial MSDL text is as follows:

[0099] [example text]”

[0100] S33, the fine-tuning model is used to perform random replacement processing on the initial MSDL text to obtain an MSDL simulation text set.

[0101] Optionally, the fine-tuning model is used to perform random replacement processing on the initial MSDL text to obtain an MSDL simulation text set, the fine-tuning model is set as a generation agent by a preset generation prompt, and then the generation agent is used to process the initial MSDL text to obtain M MSDL texts. The generation prompt can be in the following form:

[0102] “You are now a generator, please randomly replace the values of some fields in [initial MSDL text] to obtain 5 MSDL texts. The fields to be replaced are the start_time fields in the timeline layer, and the value of the start_time field located in front of the text is required to be always less than or equal to the value of the start_time field located behind the text.”

[0103] In yet another optional embodiment, the fine-tuning model is used to perform clarification processing on the user scene description text to obtain a clarified scene description text, comprising:

[0104] S311, the fine-tuning model is used to perform fuzzy detection on the user scene description text to obtain a fuzzy text set; the fuzzy text set includes a plurality of fuzzy texts.

[0105] Optionally, the fine-tuning model is used to perform fuzzy detection on the user scene description text to obtain a fuzzy text set, the fine-tuning model is set as a fuzzy detection agent by a preset fuzzy detection prompt, and then the fuzzy detection agent is used to process the user scene description text to obtain the fuzzy text set. The fuzzy detection prompt can be in the following form:

[0106] "You are now a fuzzy detector, please detect and extract the parts with ambiguous meaning in the user input text, get the fuzzy text, and then combine all the fuzzy texts into a fuzzy text set. The parts with ambiguous meaning include the following:

[0107] a. Action parameters are ambiguous (for example, fly to B point, without specifying the location of B point, without specifying the flight speed, height, and path);

[0108] b. Entity usage is ambiguous (for example, let the UAV reconnaissance, without specifying the specific UAV model);

[0109] c. Temporal and spatial constraints are ambiguous (for example, attack after 10 minutes, without specifying whether 10 minutes is simulation time or physical time);

[0110] d. Action conflict (for example, XX plane radar detection and electronic interference at T time, there is a tactical action conflict, which violates the use order);

[0111] S312, based on the above fuzzy text set, query the preset knowledge base to get the associated text sequence set; the above associated text sequence set includes the associated text sequence corresponding to each of the above fuzzy text in the above fuzzy text set; the above associated text sequence includes N associated texts; N is an integer greater than 1.

[0112] Preferably, N is 3.

[0113] It should be noted that the above knowledge base includes the description text of common entities, which explains the main type, sub-type and attribute sequence of each entity, as well as the executable entity action of each entity.

[0114] It should be noted that the above query of the preset knowledge base based on the above fuzzy text set is to query N pieces of text most similar to each fuzzy text in the knowledge base to get the corresponding N associated texts. For example, the fuzzy text "UAV formation" is queried to get the associated text sequence including the following three associated texts: "single aircraft formation", "formation for double aircraft cooperation" and "four aircraft diamond aircraft formation".

[0115] S313, using the above fine-tuning model, the above fuzzy text set and the above associated text sequence set are subjected to multiple rounds of confirmation processing to get a user confirmation text set.

[0116] S314, using the above fine-tuning model, based on the above user confirmation text set, the above user scene description text is subjected to correction processing to get the above clarified scene description text.

[0117] It should be noted that the above modification processing is to set the fine-tuning model as a modification agent by using a preset modification prompt word, and use the modification agent to process the user confirmation text set and the user scene description text to obtain the clarified scene description text. The modification prompt word can be in the following form:

[0118] "Please modify the ambiguous part in [user scene description text] based on [user confirmation text set] and output the modified text."

[0119] For example, the user confirmation text set is {“double machine cooperation”, “fly at an altitude of 2000 meters”, “cruise mode uses economic mode”}, and the user scene description text is “use a UAV formation to conduct reconnaissance on target A 100 km away, and the score after completing the target is 100”, and the following clarified scene description text is obtained: “use a UAV formation with double machine cooperation to conduct reconnaissance on target A 100 km away, fly at an altitude of 2000 meters, and use economic mode for cruise, and the score after completing the target is 100”.

[0120] In yet another optional embodiment, the above multi-round confirmation processing of the above ambiguous text set and the above associated text sequence set by using the above fine-tuning model obtains the user confirmation text set, comprising:

[0121] S3131, set the initial confirmation text set to an empty set.

[0122] S3132, set the current ambiguous text to any of the ambiguous texts in the ambiguous text set; set the current associated text sequence to the associated text sequence corresponding to the current ambiguous text in the associated text sequence set.

[0123] S3133, use the fine-tuning model to perform user confirmation processing on the current ambiguous text and the current associated text sequence to obtain a user confirmation text; add the user confirmation text to the user confirmation text set.

[0124] It should be noted that the above user confirmation processing is to set the fine-tuning model as a user confirmation agent by using a preset user confirmation prompt word, and use the user confirmation agent to process the current ambiguous text and the current associated text sequence to obtain a user question, and then obtain a user confirmation text in response to the user's answer. The user confirmation prompt word can be in the following form:

[0125] "Please generate a question according to [current ambiguous text] and [current associated text sequence]. Wherein [current ambiguous text] is the main body of the stem, and each part of [current associated text sequence] is the main body of an option of the question. Then according to the input option of the user, the corresponding option is output as the user confirmation text. When the user does not select the given option, the answer of the user is taken as the user confirmation text."

[0126] Taking the above ambiguous text "unmanned aerial vehicle formation" and the corresponding associated text sequence ["single machine formation", "formation of double machine cooperation", "four machine diamond aircraft formation"] as an example, the fine-tuning model outputs the following question: "How is the unmanned aerial vehicle formation composed? (1) The unmanned aerial vehicle formation is single; (2) The unmanned aerial vehicle formation is double machine cooperation; (3) The unmanned aerial vehicle formation is a four-machine diamond formation". If the user selects answers 1, 2 or 3, the user confirmation text is "unmanned aerial vehicle formation is single", "unmanned aerial vehicle formation is double machine cooperation" or "unmanned aerial vehicle formation is a four-machine diamond formation" respectively; but if the user inputs the answer "unmanned aerial vehicle formation is a three-machine triangular formation", the answer text is directly taken as the user confirmation text.

[0127] S3134, delete the above-mentioned current ambiguous text from the above-mentioned ambiguous text set to obtain the deleted ambiguous text set;

[0128] S3135, repeat S3132-S3134 until the deleted ambiguous text set is empty;

[0129] S3136, determine the initial confirmation text set as the user confirmation text set.

[0130] In yet another optional embodiment, the simulation evaluation processing of the MSDL simulation text set to obtain the optimal evaluation value and the optimal MSDL text comprises:

[0131] S41, using a preset simulation engine, respectively simulating the M MSDL texts in the MSDL simulation text set to obtain a simulation information set.

[0132] The simulation information set comprises M simulation information; the simulation information comprises simulation time and a plurality of target completion information; the target completion information comprises target completion flag and completion time; the value of the target completion flag is 0 or 1.

[0133] It should be noted that the M simulation information respectively corresponds to the M MSDL texts; the target completion information in the simulation information corresponds one-to-one to the simulation target of the target layer in the MSDL text.

[0134] Optionally, the simulation engine can be HAMERITT, SST (Structural Simulation Toolkit), etc., and the present application is not limited thereto.

[0135] It should be noted that after the simulation engine is started, the corresponding MSDL text is automatically loaded and parsed, and the simulation is started according to the configuration of each field in the MSDL. The simulation engine detects each simulation target in real time during running, sets the corresponding target completion flag to 1 when a certain simulation target is completed, and outputs the corresponding completion time; further, if a certain simulation target is not completed at the end of the entire simulation, the corresponding target completion flag is set to 0.

[0136] S42, using a simulation evaluation model, processing the simulation information set to obtain a simulation evaluation value set; the simulation evaluation value set includes M simulation evaluation values.

[0137] S43, setting the optimal evaluation value as the largest simulation evaluation value in the simulation evaluation value set.

[0138] S44, setting the optimal sequence number p as the sequence number of the optimal evaluation value in the simulation evaluation value set.

[0139] S45, setting the optimal MSDL text as the pth MSDL text in the MSDL simulation text set.

[0140] In another optional embodiment, the expression of the simulation evaluation model is:

[0141]

[0142] In the formula, G j is the jth simulation evaluation value, j is an integer from 1 to M; K j is the number of target completion information in the jth simulation information; f j,k and t j,k are the target completion flag and the completion time of the kth target completion information of the jth simulation information, respectively; T j is the simulation time of the jth simulation information; s j,k and u j,k are the score type and the target score of the kth simulation target of the jth MSDL text, respectively.

[0143] It can be seen that through the simulation evaluation model, the simulation effect can be evaluated in combination with the score type and target score of each simulation target in the target layer of the MSDL text and the simulation result of the MSDL text. The earlier the positive target is completed, the higher the positive contribution to the simulation evaluation value is.

[0144] It can be seen that the simulation scenario generation method based on a large language model described in the embodiments of the present application can perform deep reasoning on high-level natural language intentions in combination with domain knowledge to automatically generate an optimal simulation scenario and a corresponding evaluation value that are standardized and executable.

[0145] Embodiment Two

[0146] Please refer to Figure 2 , Figure 2 is a structural schematic diagram of a simulation scenario generation device based on a large language model disclosed by the embodiments of the present application. Wherein, Figure 2 The simulation scenario generation device based on a large language model described in the embodiments of the present application can be applied to the generation of simulation scenarios in the field of scenario simulation, such as military deduction, unmanned system cluster control, city traffic simulation, etc. The embodiments of the present application are not limited. For example Figure 2 As shown in the figure, the device can include a model construction module 201, a text acquisition module 202, a simulation text generation module 203 and a simulation evaluation module 204.

[0147] The above-mentioned model construction module 201 is used to construct a fine-tuning model based on a preset first large language model and a second large language model;

[0148] The above-mentioned text acquisition module 202 is used to acquire a user scenario description text;

[0149] The above-mentioned simulation text generation module 203 is used to process the above-mentioned user scenario description text by using the above-mentioned fine-tuning model to obtain a MSDL simulation text set;

[0150] The above-mentioned simulation evaluation module 204 is used to perform simulation evaluation processing on the above-mentioned MSDL simulation text set to obtain an optimal evaluation value and an optimal MSDL text.

[0151] It can be seen that the simulation scenario generation device based on a large language model described in the embodiments of the present application can perform deep reasoning on high-level natural language intentions in combination with domain knowledge to automatically generate an optimal simulation scenario and a corresponding evaluation value that are standardized and executable.

[0152] Embodiment Three

[0153] Please refer to Figure 3 , Figure 3is a structural schematic diagram of another large language model-based simulation scene generation device disclosed by the embodiment of the present application. Among them, Figure 3 The large language model-based simulation scene generation device described can be applied to the generation of simulation scenes in the field of scenario simulation, such as military deduction, unmanned system cluster control, urban traffic simulation, etc. The embodiment of the present application is not limited. For example, Figure 3 As shown, the large language model-based simulation scene generation device can include the following parts:

[0154] The memory 301 stores executable program codes;

[0155] The processor 302 is coupled to the memory 301;

[0156] The processor 302 calls the executable program codes stored in the memory 301, which are used to execute the steps in the large language model-based simulation scene generation method described in embodiment one.

[0157] Embodiment four

[0158] The embodiment of the present application discloses a computer readable storage medium, which stores a computer program for electronic data exchange, wherein the computer program makes the computer execute the steps in the large language model-based simulation scene generation method described in embodiment one.

[0159] The device embodiments described above are only schematic, wherein the modules illustrated as separate components can or can not be physically separated, and the components illustrated as modules can or can not be physical modules, that is, they can be located in one place, or they can be distributed on multiple network modules. According to actual needs, part or all of the modules can be selected to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement it without creative labor.

[0160] Those skilled in the art can clearly understand the implementation of the various embodiments by means of software and the necessary general hardware platform through the specific description of the above embodiments, and of course, the various embodiments can also be implemented by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, and the computer software product can be stored in a computer readable storage medium, including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, a magnetic disk storage, a magnetic tape storage, or any other computer readable medium that can be used to carry or store data.

[0161] Finally, it should be noted that: the simulation scene generation method and device based on a large language model disclosed by the embodiments of the present application are only the preferred embodiments of the present application, and are used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that; it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating simulation scenes based on a large-scale language model, characterized in that, include: S1. Based on the preset first and second language models, a fine-tuned model is constructed. S2. Obtain the user scenario description text; S3. Using the fine-tuning model, process the user scenario description text to obtain an MSDL simulation text set; the MSDL simulation text set includes M MSDL texts; M is an integer greater than 1; S4. Perform simulation evaluation processing on the MSDL simulation text set to obtain the optimal evaluation value and the optimal MSDL text.

2. The simulation scene generation method based on a large-scale language model according to claim 1, characterized in that, The fine-tuning model, constructed based on the preset first and second language models, includes: S11. A preset MSDL text training set is provided; the MSDL text training set includes several MSDL texts. S12. Using the first large language model, the MSDL text training set is generated to obtain a scene description text training set; the scene description text training set includes the scene description text corresponding to each MSDL text in the MSDL text training set. S13. Based on the MSDL text training set and the scene description text training set, fine-tune the second large language model to obtain the fine-tuned model.

3. The simulation scene generation method based on a large-scale language model according to claim 1, characterized in that, The MSDL text includes an entity layer, a timing layer, a trigger layer, and a target layer; The entity layer includes several entities; each entity includes an entity ID field, a main type field, a subtype field, an attribute sequence, and an initial state sequence; the attribute sequence includes several attribute fields; the initial state sequence includes several initial state fields. The timing layer includes several simulation events; each simulation event includes an event ID field, an entity ID field, a start time field, an entity action, and / or a dependent event field. The entity action includes an action type field and / or several action attribute fields; The trigger layer includes several triggers; each trigger includes a trigger ID field, a trigger condition, and an action sequence; the trigger condition includes a condition type field and several trigger attribute fields; the action sequence includes several entity actions. The target layer includes several simulation targets; each simulation target includes a target ID field, a target type field, a readable description field, a completion condition, a score type, and a target score. The completion conditions include a condition type field, the entity ID field, a detection attribute field, an operator field, and a target value field; the value of the score type is 0 or 1.

4. The simulation scene generation method based on a large-scale language model according to claim 1, characterized in that, The process of using the fine-tuning model to process the user scenario description text yields an MSDL simulation text set, including: S31. Using the fine-tuning model, the user scenario description text is clarified to obtain the clarified scenario description text; S32. Using the fine-tuning model, perform initial transformation processing on the clarified scene description text to obtain the initial MSDL text; S33. Using the fine-tuning model, the initial MSDL text is randomly replaced to obtain the MSDL simulation text set.

5. The simulation scene generation method based on a large-scale language model according to claim 4, characterized in that, The step of using the fine-tuning model to clarify the user scenario description text to obtain clarified scenario description text includes: S311. Using the fine-tuning model, perform fuzzy detection on the user scenario description text to obtain a fuzzy text set; the fuzzy text set includes several fuzzy texts. S312. Based on the fuzzy text set, query a preset knowledge base to obtain a set of associated text sequences; the set of associated text sequences includes the associated text sequence corresponding to each fuzzy text in the fuzzy text set; the associated text sequence includes N associated texts; N is an integer greater than 1; S313. Using the fine-tuning model, perform multiple rounds of confirmation processing on the fuzzy text set and the associated text sequence set to obtain the user confirmation text set. S314. Using the fine-tuning model, based on the user confirmation text set, the user scenario description text is corrected to obtain the clarified scenario description text.

6. The simulation scene generation method based on a large-scale language model according to claim 4, characterized in that, The process of using the fine-tuning model to perform multiple rounds of confirmation processing on the fuzzy text set and the associated text sequence set to obtain the user confirmation text set includes: S3131. Set the initial confirmation text set to an empty set; S3132. Set the current fuzzy text to any of the fuzzy texts in the fuzzy text set; set the current associated text sequence to the associated text sequence corresponding to the current fuzzy text in the associated text sequence set; S3133. Using the fine-tuning model, perform user confirmation processing on the current fuzzy text and the current associated text sequence to obtain user confirmation text; add the user confirmation text to the user confirmation text set. S3134. Delete the current fuzzy text from the fuzzy text set to obtain the fuzzy text set after deletion; S3135. Repeat S3132 to S3134 until the set of fuzzy texts after deletion is an empty set; S3136. The initial confirmation text set is determined as the user confirmation text set.

7. The simulation scene generation method based on a large language model according to claim 3, characterized in that, The process of performing simulation evaluation on the MSDL simulation text set to obtain the optimal evaluation value and the optimal MSDL text includes: S41. Using a preset simulation engine, simulate and process each of the M MSDL texts in the MSDL simulation text set to obtain a simulation information set. The simulation information set includes M simulation information items; the simulation information includes simulation time and several target completion information items; the target completion information includes a target completion flag and a completion time; the value of the target completion flag is 0 or 1; S42. Using a simulation evaluation model, the simulation information set is processed to obtain a simulation evaluation value set; the simulation evaluation value set includes M simulation evaluation values; S43. Set the optimal evaluation value as the largest simulation evaluation value in the set of simulation evaluation values; S44. Set the optimal index p to the index of the optimal evaluation value in the set of simulation evaluation values; S45. Set the optimal MSDL text as the p-th MSDL text in the MSDL simulation text set.

8. A simulation scene generation device based on a large-scale language model, characterized in that, The device includes a model construction module, a text acquisition module, a simulation text generation module, and a simulation evaluation module; The model construction module is used to construct a fine-tuned model based on a preset first and second language models. The text acquisition module is used to acquire user scenario description text; The simulation text generation module is used to process the user scenario description text using the fine-tuning model to obtain an MSDL simulation text set. The simulation evaluation module is used to perform simulation evaluation processing on the MSDL simulation text set to obtain the optimal evaluation value and the optimal MSDL text.

9. A simulation scene generation device based on a large-scale language model, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the simulation scene generation method based on a large language model as described in any one of claims 1-7.

10. A computer-storable medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the simulation scene generation method based on a large language model as described in any one of claims 1-7.

Images