Intelligent action instruction generation method

Abstract

The invention belongs to the field of computer application, and relates to an action instruction intelligent generation method, which comprises the following steps of: (10) constructing an action scene: determining an action target, analyzing an action environment, determining a participant and designing diversified action scenes; and (20) finely adjusting a language large model generated by the action instruction: modifying parameters in a pre-training model to enable the model to adapt to a specific task. And (30) scene-oriented action instruction intelligent generation: combining action scenes to organize action verbal skills, and generating action instructions through a fine-tuned language large model. And (40) comprehensively evaluating the action instruction: establishing an evaluation index system, and comprehensively evaluating the generation of the instruction. And (50) continuously optimizing and iterating the instruction generation model: bringing the actual execution result information of the action into model training to form a closed-loop feedback mechanism. The method has the beneficial effects that the action decision-making efficiency is remarkably improved, accurate instruction transmission is ensured, the action resource configuration is optimized, and the flexible response to the conflict area situation is promoted.

Description

Technical Field

[0001] The invention belongs to the field of computer application and discloses an intelligent generation method of action instructions. Background Art

[0002] In today's era of information-based conflicts, the efficiency, accuracy and intelligence of actions have become one of the key factors in determining military victory or defeat. With the rapid development of cutting-edge technologies such as artificial intelligence, big data, and cloud computing, the military field is undergoing unprecedented changes. Traditional action command generation relies on the commander's experience and manual planning. This process is not only time-consuming and labor-intensive, but also difficult to fully consider the complexity and variability of the conflict area and the uncertainty of the enemy's actions, which often limits the full play of the effectiveness of the action.

[0003] In this context, the intelligent generation technology of action instructions came into being. It aims to automatically or semi-automatically generate scientific, reasonable and efficient action instructions by integrating advanced algorithm models and real-time military data. This technology can significantly improve the timeliness of command decisions, enabling commanders to quickly respond to changes in the situation in the conflict area and make more accurate decisions. At the same time, the intelligent generation method can also effectively reduce the burden on commanders, allowing them to focus more on strategic thinking and judgment, thereby further improving the overall effectiveness of operations.

[0004] However, the current development of intelligent generation technology for action instructions still faces many challenges. On the one hand, the high complexity and dynamics of the conflict zone environment require intelligent systems to have strong data processing and analysis capabilities to quickly identify situation changes and accurately predict enemy action trends. On the other hand, the generation of action instructions must also fully consider the mutual influence and constraints between multiple factors such as force deployment, weapon equipment performance, action mission objectives, and geographical environment to ensure the feasibility and optimality of the instructions. In addition, ensuring the security and reliability of the instruction generation process and preventing information leakage or misjudgment and misoperation are also important issues that need to be solved urgently. Summary of the invention

[0005] In view of the above background, the present invention aims to propose a method for intelligently generating action instructions to overcome the shortcomings of the prior art. The method will make full use of the language large model to realize the intelligent generation and instant transmission of action instructions. By generating action instructions, the efficiency and accuracy of action command can be improved, providing strong technical support for modern action command.

[0006] The present invention implements an intelligent generation method of action instructions through the following technical scheme, including the following steps:

[0007] (10) Construct action scenarios: clarify action objectives, analyze the action environment, identify participants, and design a variety of action scenarios.

[0008] (20) Fine-tune the language model for action command generation: Modify the parameters in the pre-trained model to adapt the model to specific tasks.

[0009] (30) Intelligent generation of scenario-oriented action instructions: Organize action words based on action scenarios, and generate action instructions through a fine-tuned language model.

[0010] (40) Comprehensive evaluation of action instructions: Establish an evaluation index system to conduct a comprehensive evaluation of the generation of instructions.

[0011] (50) Continuous optimization and iteration of the instruction generation model: Incorporate information on the actual execution results of the action into model training to form a closed-loop feedback mechanism.

[0012] The beneficial effects of the present invention are remarkable. It not only greatly improves the intelligent level of action command generation, realizes the precise connection from action goal setting to command execution, but also significantly enhances the flexibility and response speed of action decision-making. By constructing action scenarios, action commands are closer to action needs, ensuring the pertinence and effectiveness of commands. At the same time, the fine-tuning and continuous optimization of the large language model make the command generation more in line with the military professional terminology standards, improving the execution efficiency and coordinated action capabilities of the troops. In addition, the establishment of a comprehensive evaluation system and the introduction of a closed-loop feedback mechanism ensure that the command generation model can continuously improve itself and adapt to complex and changing conflict areas, providing strong technical support for efficient command and action in modern conflicts. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The steps of the method of the present invention are shown in FIG.

[0014] Figure 2 Fine-tuning flowchart for large language models

[0015] Figure 3 Receiving irregular instructions for the action simulation system

[0016] Figure 4 Receive correct instructions for the action simulation system

[0017] Figure 5 Execute the correct instructions for the action simulation system DETAILED DESCRIPTION

[0018] The design of the present invention proposes a method for intelligently generating action instructions, such as Figure 1 As shown, the following steps are included:

[0019] (10) Construct action scenarios: clarify action objectives, analyze the action environment, identify participants, and design a variety of action scenarios.

[0020] The steps of constructing the action scenario in (10) above are as follows:

[0021] (11) Clarify the objectives of the operation: determine the specific objectives to be achieved in this operation, such as area control, weakening enemy forces, intelligence collection, etc.;

[0022] (12) Analyze the operational environment: consider the impact of geography, climate, time, enemy force deployment, friendly support and other environmental factors on the operation;

[0023] (13) Identify the parties involved: Identify the parties involved in the operation, including our side, the enemy side, neutral parties, and possible third-party forces.

[0024] (20) Fine-tune the language model for action command generation: Modify the parameters in the pre-trained model to adapt the model to specific tasks.

[0025] like Figure 2 As shown, the steps of (20) for fine-tuning the language model generated by action instructions are specifically as follows:

[0026] (21) Introducing a pre-trained basic model: Introduce a model that has completed preliminary training on a large-scale text dataset as a starting point. This model has a general language representation capability due to its learning on a wide range of text data.

[0027] (22) Construction of a customized layer: In order to meet the needs of specific tasks, a specific task adaptation layer is constructed or added on the basis of the pre-trained model. In this patent, when processing the instruction generation task, a fully connected layer is added to the top of the model. This layer can predict the JSON format of the instruction generation and the specific content of the instruction based on the features extracted by the pre-trained model.

[0028] Add a fully connected layer on top of the Transformer model for the instruction generation task. Mathematically, this can be expressed as:

[0029] Output = softmax(W o Transformer Output +b o )

[0030] Among them: Transformer Output is the output of the last layer of the Transformer model, representing the vector representation of the text, W o and b o are the weights and bias terms of the newly added fully connected layer. The softmax function is used to convert the output into a probability distribution for instruction generation prediction.

[0031] (23) Fine-tuning parameters: The entire model (including its pre-trained part and the newly added task adaptation layer) is trained using a dataset labeled for the task. This fine-tuning process aims to adjust its parameters based on the specific needs of the new task while maintaining the original general knowledge of the model. In fine-tuning practice, the learning rate, optimization algorithm, and training strategy may be flexibly adjusted according to the characteristics of the task and the dataset.

[0032] The learning rate (η) is a hyperparameter that controls the step size of the model at each weight update. During fine-tuning, the learning rate may need to be adjusted depending on the difficulty of the task and the characteristics of the dataset. A common learning rate adjustment strategy is to use learning rate decay (Decay), where the initial learning rate η0 decreases over time or the number of iterations:

[0033] η t =η0·decay_rate t

[0034] Where t is the number of iterations and decay_rate is the decay rate (a constant less than 1).

[0035] An optimization algorithm is used to update the weights of the model based on the gradient of the loss function. A common optimization algorithm is Stochastic Gradient Descent (SGD) and its variants such as SGD with Momentum.

[0036] SGD:

[0037]

[0038] Among them, θ t is the parameter at time t, J is the loss function, x (i) ,y (i) is the tth training sample and label, is the gradient of the loss function with respect to the parameters.

[0039] SGD with momentum:

[0040]

[0041] θ t+1 =θt+v t+1

[0042] Among them, v t is the velocity vector and μ is the momentum term (usually close to 1).

[0043] Training strategies may include early stopping, regularization (such as L1, L2 regularization), and batch size adjustment. Early stopping usually stops training when the performance of the validation set no longer improves to avoid overfitting. Regularization limits model complexity by adding regularization terms to the loss function.

[0044] L2 Regularization:

[0045]

[0046] Among them, J(θ) is the objective function, L(y (i) ,f(x (i) ; θ)) is the loss function, which measures the model prediction value f(x (i) ;θ) and the true value y (i) The difference between , N is the total number of samples, λ is the regularization strength, are model parameters.

[0047] (24) Performance evaluation and optimization: After fine-tuning, use the validation set or an independent test set to evaluate the performance of the model on the new task. If the performance does not meet expectations, further exploration of the hyperparameter adjustment space during fine-tuning can be done, such as optimizing the learning rate setting, adjusting the batch size, etc., in order to obtain better model performance.

[0048] (30) Intelligent generation of scenario-oriented action instructions: Organize action words based on action scenarios, and generate action instructions through a fine-tuned language model.

[0049] The specific steps of the above (30) scenario-oriented action instruction intelligent generation are:

[0050] (31) Scene Understanding and Representation

[0051] When generating action instructions, the model first needs to understand and represent the action scenario. This usually involves converting the scenario description (such as geographic location, enemy and friendly situation, action objectives, etc.) into a numerical form that the model can understand, that is, a vector representation. These vectors may contain features such as location information, entity type, relationship, etc., similar to word vectors in natural language processing.

[0052] (32) Instruction Generation Mechanism

[0053] The generation of action instructions is a complex reasoning and decision-making process, and the model proceeds through the following steps:

[0054] Decoder role: In the Transformer model, the decoder is responsible for generating a text sequence based on the encoder's output (i.e., the scene representation). For action instructions, the decoder considers the scene vector, context information, and the generated instruction part to predict the next most appropriate word or phrase.

[0055] Conditional probability modeling: Language models essentially model the probability distribution of word sequences. When generating action instructions, the model calculates the conditional probability of each possible word in the current context and selects the word with the highest probability as the output. This can be seen as an implicit "formula" that maximizes the conditional probability sequence:

[0056]

[0057] Among them, w i is the i-th word in the instruction, and context represents the action scenario and other contextual information.

[0058] Attention Mechanism: The self-attention and encoder-decoder attention mechanisms in Transformer allow the model to focus on the entire input sequence (i.e., action scenario description) and the generated instruction part when processing the current word, thereby generating more coherent and accurate instructions.

[0059] (40) Comprehensive evaluation of action instructions: Establish an evaluation index system to conduct a comprehensive evaluation of the generation of instructions.

[0060] The steps for the comprehensive evaluation of the above (40) action instructions are as follows:

[0061] (41) Evaluate the perplexity index

[0062] Perplexity is a basic indicator for measuring the quality of text generated by a language model. It reflects the model's ability to predict test data. The lower the perplexity, the stronger the model's ability to predict the next word. The calculation formula is:

[0063]

[0064] Where, is the number of words in N text, P(w i |w1,...,w i-1 ) is the model prediction for word i w i The probability of, given the previous words w1, ..., w i-1 .

[0065] (42) Evaluate the BLEU indicator

[0066] BLEU (Bilingual Evaluation Understudy) is a machine translation evaluation metric based on n-gram accuracy, which is used to evaluate the similarity between candidate instructions and reference instructions. The BLEU score ranges from 0 to 1, and the higher the score, the better the quality of the candidate instruction. The calculation formula is:

[0067]

[0068] Among them, BP is the length penalty factor, which is used to adjust the score when the candidate instruction length is inconsistent with the reference instruction length. n is the weight of each n-gram, usually Where N is the maximum length of an n-gram (e.g. 4). n is the precision of n-gram, that is, the proportion of n-gram in the candidate translation that appears in the reference translation.

[0069] (43) Evaluate the ROUGE index

[0070] ROUGE is a recall-based text summarization evaluation method that can also be used to evaluate the quality of action instructions. ROUGE-N (N-gram) and ROUGE-L (Longest Common Subsequence) are two commonly used variants. ROUGE-N calculates the n-gram overlap between the candidate text and the reference text, while ROUGE-L considers the longest common subsequence between the two texts.

[0071] Taking ROUGE-L as an example, its calculation involves finding the longest common subsequence LCS(C, R) between the candidate text C and the reference text R, and then calculating the F-score (which combines precision and recall):

[0072]

[0073] Where R is the recall rate, P is the precision rate, β is a balance parameter (usually set to 1.0), LCS is the length of the longest common subsequence, and C and R are the lengths of the candidate text and the reference text, respectively.

[0074] (44) Observe the instruction execution process

[0075] Generating instructions without fine-tuning the language model may result in non-standard instructions that cannot be executed, such as Figure 3 shown.

[0076] The fine-tuned language model can receive generated instructions through the action simulation system, such as Figure 4 shown.

[0077] The action simulation system executes the correct action instructions, such as Figure 5 shown.

[0078] (50) Continuous optimization and iteration of the instruction generation model: Incorporate information on the actual execution results of the action into model training to form a closed-loop feedback mechanism.

[0079] The specific steps of continuous optimization and iteration of the above (50) instruction generation model are as follows:

[0080] (51) Data collection and integration: First, collect information on the actual execution results of the action, including successful and failed cases, feedback during the execution process, etc. These data need to be cleaned and integrated to form a standard data set that can be used for model training.

[0081] (52) Model training and optimization: The integrated data is input into the instruction generation model for retraining. By comparing the actual execution results with the model prediction results, the model parameters are continuously adjusted to optimize the model performance. At the same time, new algorithms or technologies can be introduced, such as the automatic optimization prompt technology, to further improve the accuracy and efficiency of the model generation instructions.

[0082] (53) Establishment and implementation of feedback mechanism: Establish a closed-loop feedback mechanism to ensure continuous optimization of the model.

[0083] The above is only a preferred embodiment of the present invention, and the invention is not limited to the above embodiment. Anyone should know that any technical solution that is the same or similar to the present invention made under the inspiration of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for intelligently generating action instructions, characterized in that: The steps of the method are as follows: (10) Constructing action scenarios: clarifying action goals, analyzing the action environment, identifying participants, and designing diverse action scenarios; (20) Fine-tune the language model for action command generation: modify the parameters in the pre-trained model to adapt the model to a specific task; (30) Intelligent generation of scenario-oriented action instructions: Organize action words based on action scenarios and generate action instructions through a fine-tuned language model; (40) Comprehensive evaluation of action instructions: Establish an evaluation index system to conduct a comprehensive evaluation of the generation of instructions; (50) Continuous optimization and iteration of the instruction generation model: Incorporate information on the actual execution results of the action into model training to form a closed-loop feedback mechanism.

2. A method for intelligently generating action instructions according to claim 1, characterized in that: The steps of constructing the action scenario in (10) are specifically as follows: (11) Clarify the objectives of the operation: determine the specific objectives to be achieved in this operation, such as area control, weakening enemy forces, intelligence collection, etc.; (12) Analyze the operational environment: consider the impact of geography, climate, time, enemy force deployment, friendly support and other environmental factors on the operation; (13) Identify the parties involved: Identify the parties involved in the operation, including our side, the enemy side, neutral parties, and possible third-party forces.

3. A method for intelligently generating action instructions according to claim 1, characterized in that: The steps of fine-tuning the language model generated by the action instruction (20) are specifically as follows: (21) Introducing a pre-trained basic model: Introduce a model that has completed preliminary training on a large-scale text dataset as a starting point. This model has a general language representation capability due to its learning on a wide range of text data; (22) Construction of a customized layer: In order to meet the needs of specific tasks, a specific task adaptation layer is constructed or added on the basis of the pre-trained model. In this patent, when processing the instruction generation task, a fully connected layer is added to the top of the model. This layer can predict the JSON format of the instruction generation and the specific content of the instruction based on the features extracted by the pre-trained model; (23) Fine-tuning parameters: Use the dataset labeled for the task to train the entire model (including its pre-trained part and the newly added task adaptation layer). This fine-tuning process aims to adjust its parameters based on the specific needs of the new task while maintaining the original general knowledge of the model. In the fine-tuning practice, the learning rate, optimization algorithm and training strategy may be flexibly adjusted according to the characteristics of the task and the dataset; (24) Performance evaluation and optimization: After fine-tuning, use the validation set or an independent test set to evaluate the performance of the model on the new task. If the performance does not meet expectations, further exploration of the hyperparameter adjustment space during fine-tuning can be done, such as optimizing the learning rate setting, adjusting the batch size, etc., in order to obtain better model performance.

4. The method for intelligently generating action instructions according to claim 1, characterized in that: The specific steps of intelligently generating scene-oriented action instructions in (30) are as follows: (31) Scene Understanding and Representation When generating action instructions, the model first needs to understand and represent the action scenario. This usually involves converting the scenario description (such as geographic location, enemy and friendly situation, action objectives, etc.) into a numerical form that the model can understand, that is, a vector representation. These vectors may contain features such as location information, entity type, relationship, etc., similar to word vectors in natural language processing; (32) Instruction Generation Mechanism The generation of action instructions is a complex reasoning and decision-making process, and the model proceeds through the following steps: Decoder: In the Transformer model, the decoder is responsible for generating a text sequence based on the encoder's output (i.e., the scene representation). For action instructions, the decoder considers the scene vector, context information, and the generated instruction part to predict the next most appropriate word or phrase; Conditional probability modeling: Language models essentially model the probability distribution of word sequences. When generating action instructions, the model calculates the conditional probability of each possible word in the current context and selects the word with the highest probability as the output; Attention Mechanism: The self-attention and encoder-decoder attention mechanisms in Transformer allow the model to focus on the entire input sequence (i.e., action scenario description) and the generated instruction part when processing the current word, thereby generating more coherent and accurate instructions.

5. The method for intelligently generating action instructions according to claim 1, characterized in that: The comprehensive evaluation steps of the action instructions in (40) are specifically as follows: (41) Evaluate the perplexity index Perplexity is a basic indicator for measuring the quality of text generated by a language model. It reflects the model's ability to predict test data. The lower the perplexity, the stronger the model's ability to predict the next word. (42) Evaluate the BLEU indicator BLEU (Bilingual Evaluation Understudy) is a machine translation evaluation metric based on n-gram accuracy, which is used to evaluate the similarity between candidate instructions and reference instructions. The BLEU score ranges from 0 to 1, and the higher the score, the better the quality of the candidate instruction; (43) Evaluate the ROUGE index ROUGE is a recall-based text summarization evaluation method that can also be used to evaluate the quality of action instruction generation. ROUGE-N (N-gram) and ROUGE-L (Longest Common Subsequence) are two commonly used variants. ROUGE-N calculates the n-gram overlap rate between the candidate text and the reference text, while ROUGE-L considers the longest common subsequence between the two texts; (44) Observe the instruction execution process The generated instructions are received by the action simulation system and then executed.

6. The method for intelligently generating action instructions according to claim 1, characterized in that: The continuous optimization and iteration steps of the instruction generation model in (50) are specifically as follows: (51) Data collection and integration: First, collect information on the actual execution results of the action, including successful and failed cases, feedback during the execution process, etc. These data need to be cleaned and integrated to form a standard data set that can be used for model training; (52) Model training and optimization: The integrated data is input into the instruction generation model for retraining. By comparing the actual execution results with the model prediction results, the model parameters are continuously adjusted to optimize the model performance. At the same time, new algorithms or technologies can be introduced, such as the automatic optimization prompt technology, to further improve the accuracy and efficiency of the model generation instructions; (53) Establishment and implementation of feedback mechanism: Establish a closed-loop feedback mechanism to ensure continuous optimization of the model.

Images