Intelligent Recommendation Decision-Making Method and System Based on Graph Neural Networks and Cognitive Architecture

Abstract

This application provides an intelligent recommendation and decision-making method and system based on graph neural networks and cognitive architecture. The method includes: performing situational awareness knowledge reasoning based on the agent's detection data of the current environmental state to obtain a situational awareness structure; performing task planning knowledge reasoning on the situational awareness structure to obtain all matching task planning rules, and generating an operator structure based on each matching task planning rule; the operator structure representing the agent's actions; using a graph neural network-based intelligent recommendation model to perform matching degree reasoning on the situational awareness structure and the operator structure to obtain dynamic preference values ​​for the operator structures; and selecting the optimal operator structure based on the dynamic preference values ​​of each operator structure to determine the decision action. This solution can solve the problem of insufficient robustness in unmanned systems using cognitive architecture for decision-making, improve the reliability of intelligent agent decision-making, and can be widely applied in the field of intelligent decision-making technology for autonomous mobile systems.

Description

Technical Field

[0001] This application relates to the field of intelligent decision-making technology for autonomous mobile systems, and in particular to an intelligent recommendation decision-making method and system based on graph neural networks and cognitive architecture. Background Technology

[0002] In recent years, unmanned systems (i.e., autonomous mobile systems, such as drone systems or unmanned vehicle systems) have been increasingly widely used in military and civilian fields. While the autonomous capabilities of individual unmanned systems have made significant progress, efficient collaboration among multiple unmanned systems remains a technical challenge when facing complex and dynamic tasks. Currently, research on collaborative decision-making for multiple unmanned systems mainly relies on deep learning methods. Although methods based on deep reinforcement learning and swarm intelligence can handle complex problems, their decision-making process is like a "black box," making it difficult to explain why a particular decision is made. In scenarios requiring high safety and reliability, this unexplainable deep learning collaborative decision-making method is clearly unsuitable.

[0003] To address the issue of uninterpretability, related technologies for autonomous decision-making in unmanned systems employ planning and rule-based methods, relying on predefined rule bases or planning strategies. For example, this involves writing logic based on behavior trees or finite state machines for unmanned systems, or using symbolic AI methods such as the Soar cognitive architecture. The Soar architecture simulates human cognitive decision-making processes through production rules, viewing problem-solving as a search within a problem space. Its decision-making process conforms to the OODA (Observe-Adjust-Decision-Action) cycle, exhibiting good logic and interpretability. However, purely planning and rule-based methods heavily rely on expert knowledge. The preferences of rules and operators are statically preset and cannot adaptively adjust to dynamic changes in the battlefield environment and practical experience. When unexpected new situations or tactics arise, system performance deteriorates sharply, exhibiting insufficient scalability and robustness, leading to low reliability of intelligent decision-making in unmanned systems. Summary of the Invention

[0004] The main objective of this application is to propose an intelligent recommendation decision-making method and system based on graph neural networks and cognitive architecture, aiming to solve the problem of insufficient robustness of unmanned systems when using cognitive architecture for decision-making, and to improve the reliability of intelligent agent decision-making.

[0005] To achieve the above objectives, this application proposes an intelligent recommendation decision-making method based on graph neural networks and cognitive architecture, comprising the following steps:

[0006] The situational awareness structure is obtained by reasoning about the situational awareness knowledge based on the agent's detection data of the current environmental state.

[0007] The situational awareness structure is subjected to task planning knowledge reasoning to obtain all matching task planning rules, and an operator structure is generated based on each matching task planning rule; the operator structure represents the agent's actions.

[0008] By using an intelligent recommendation model based on graph neural networks, the matching degree inference of the situational awareness structure and the operator structure is performed to obtain the dynamic preference value of the operator structure.

[0009] The optimal operator structure is selected based on the dynamic preference values ​​of each operator structure, and the decision action is determined.

[0010] In some embodiments, the step of performing situational awareness knowledge reasoning based on the agent's detection data of the current environmental state to obtain the situational awareness structure includes the following steps:

[0011] Acquire the agent's detection data of the current environmental state; the detection data includes the external environment of the system and the system state;

[0012] The detection data is preprocessed to obtain detection memory elements; the preprocessing includes at least feature processing and symbolic processing.

[0013] Based on the cognitive situation rules in the cognitive architecture, situational cognitive knowledge reasoning is performed on the probe memory elements to obtain a situational cognitive structure; the situational cognitive structure is used to represent the current situational information in a structured form.

[0014] In some embodiments, the step of performing matching degree inference on the situational awareness structure and the operator structure using an intelligent recommendation model based on a graph neural network to obtain the dynamic preference value of the operator structure includes the following steps:

[0015] The intelligent recommendation model is used to embed vector representations of the situational awareness structure and the operator structure, respectively, to obtain the first embedding vector and the second embedding vector.

[0016] Perform a dot product operation on the first embedding vector and the second embedding vector to obtain the dynamic preference value of the operator structure.

[0017] In some embodiments, the intelligent recommendation model is trained through the following steps:

[0018] A training dataset is collected during the operation of the unmanned system; each sample data in the training dataset includes the agent's situational awareness structure, operator structure, and decision reward for executing the operator structure at the decision-making time.

[0019] A graph structure is constructed on the training dataset to obtain a situation-operator bipartite graph; the node set of the situation-operator bipartite graph includes a situation node set and an operator node set, and the edge set of the situation-operator bipartite graph represents the historical execution of operator node j by situation node i, and the edge weight represents the corresponding decision reward;

[0020] The situation-operator bipartite graph is input into an intelligent recommendation model based on graph convolutional networks for embedding propagation and neighbor aggregation to obtain the embedding vector of each node; the embedding vector is used to characterize the matching degree between the situation and the operator.

[0021] The parameters of the intelligent recommendation model are updated in reverse based on the embedding vectors of each node to obtain the trained intelligent recommendation model.

[0022] In some embodiments, updating the parameters of the intelligent recommendation model in reverse based on the embedding vectors of each node to obtain the trained intelligent recommendation model includes the following steps:

[0023] For each situation node, positive sample operators that it has interacted with and negative sample operators that it has not interacted with are sampled from the training dataset, and the prediction preference values ​​of the situation node with the positive sample operators and the negative sample operators are calculated respectively.

[0024] The optimization objective is to make the predicted preference value of the positive sample operator greater than that of the predicted preference value of the negative sample operator. The parameters of the intelligent recommendation model are continuously optimized until the optimization objective is achieved, thus obtaining a trained intelligent recommendation model.

[0025] In some embodiments, the step of selecting the optimal operator structure based on the dynamic preference values ​​of each operator structure and determining the decision action includes the following steps:

[0026] According to the dynamic preference value, each of the operator structures is ordered from high to low, and the operator structure with the highest ranking is determined as the first candidate operator and the operator structure with the second ranking is determined as the second candidate operator.

[0027] The confidence level is determined based on the difference between the dynamic preference values ​​of the first candidate operator and the second candidate operator.

[0028] If the confidence level is greater than or equal to the expected threshold, a decision action is determined based on the first candidate operator;

[0029] If the confidence level is less than the expected threshold, simulations are performed on the first candidate operator and the second candidate operator respectively to obtain the corresponding simulated returns. The candidate operator with the higher simulated return is selected from the first candidate operator and the second candidate operator to determine the decision action.

[0030] In some embodiments, the intelligent recommendation decision-making method based on graph neural networks and cognitive architecture further includes the following steps:

[0031] Obtain the decision feedback from the environment after executing the decision action in the current environment state;

[0032] New sample data is generated based on the situational awareness structure of the current environmental state, the decision-making actions, and the decision-making rewards.

[0033] The intelligent recommendation model is periodically batch-trained based on multiple new sample data to optimize the intelligent recommendation model.

[0034] To achieve the above objectives, embodiments of this application also propose an intelligent recommendation decision-making system based on graph neural networks and cognitive architecture, comprising:

[0035] The first module is used to perform situational awareness knowledge reasoning based on the agent's detection data of the current environmental state to obtain the situational awareness structure;

[0036] The second module is used to perform task planning knowledge reasoning on the situational awareness structure to obtain all matching task planning rules, and generate an operator structure based on each matching task planning rule; the operator structure represents the agent's actions.

[0037] The third module is used to perform matching degree inference on the situational awareness structure and the operator structure through an intelligent recommendation model based on graph neural networks, so as to obtain the dynamic preference value of the operator structure.

[0038] The fourth module is used to select the optimal operator structure based on the dynamic preference values ​​of each operator structure and determine the decision action.

[0039] To achieve the above objectives, embodiments of this application also propose an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0040] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0041] The intelligent recommendation decision-making method, system, electronic device, and storage medium based on graph neural networks and cognitive architecture in this application embodiment have at least the following beneficial effects: First, a situational awareness structure is obtained by performing situational awareness knowledge reasoning based on the agent's detection data of the current environmental state. Then, task planning knowledge reasoning is performed on the situational awareness structure to obtain all matching task planning rules. An operator structure is generated based on each matching task planning rule. The operator structure represents the agent's actions. Then, a matching degree reasoning is performed on the situational awareness structure and the operator structure through an intelligent recommendation model based on graph neural networks to obtain the dynamic preference value of the operator structure. The intelligent recommendation model generates dynamic preference values ​​of operators based on the current environment. Compared with the static preference value method of operators based on expert experience preset in the Soar architecture, the preference value is more in line with the current new situation. Then, the optimal operator structure is selected to determine the decision action based on the dynamic preference value of each operator structure. The decision system has stronger scalability and robustness, and improves the reliability of the agent's intelligent decision-making. Attached Figure Description

[0042] Figure 1 This is a flowchart of the intelligent recommendation decision-making method based on graph neural networks and cognitive architecture provided in the embodiments of this application;

[0043] Figure 2 This is a schematic diagram of the situation-operator bipartite diagram provided in the embodiments of this application;

[0044] Figure 3 This is a schematic diagram of the node information aggregation process provided in the embodiments of this application;

[0045] Figure 4 This is a schematic diagram illustrating the overall technical concept of the agent decision-making process provided in the embodiments of this application;

[0046] Figure 5 This is a data flow diagram of the intelligent recommendation decision-making method based on graph neural networks and cognitive architecture provided in the embodiments of this application;

[0047] Figure 6 This is a schematic diagram of a forest environment under a simulated environment provided in the embodiments of this application;

[0048] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0051] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0052] Soar architecture, short for State-Operator-Result, emphasizes changing the state space through operators to achieve a goal. It is a cognitive modeling framework based on symbolism, and its core is to simulate the human problem-solving process through state space search and rule triggering mechanisms.

[0053] LightGCN is a lightweight graph convolutional network optimized for recommender systems. It significantly improves the performance and training efficiency of collaborative filtering tasks by simplifying redundant operations in traditional GCN models.

[0054] Currently, research on collaborative decision-making in multi-unmanned systems mainly focuses on the following aspects:

[0055] (1) Machine learning-based methods: especially deep reinforcement learning (DRL), which designs reward functions to allow agents to learn optimal policies through a large number of simulated interactions. For example, multi-agent reinforcement learning algorithms (such as MADDPG and MAPPO) are used to train UAV swarms for air combat or collaborative reconnaissance. This approach suffers from the problems of being a black box and having poor interpretability. Although methods based on deep reinforcement learning and swarm intelligence can handle complex problems, their decision-making process is like a "black box," making it difficult to explain why a certain decision is made. In scenarios requiring high security and high reliability, this lack of interpretability is a major drawback.

[0056] (2) Heuristic algorithms based on swarm intelligence: These algorithms mimic the behavior of biological groups in nature, such as Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO), and are actively used to solve problems such as path planning and task allocation in multi-unmanned systems. However, when faced with complex autonomous collaborative decision-making tasks in multi-unmanned systems, the algorithms need to iterate multiple times to reach a stable solution, which does not meet the real-time decision-making and action execution requirements of the intelligent agent.

[0057] (3) Knowledge-based recommendation methods: By constructing a knowledge rule base, the input data is matched and reasoned to recommend the required capabilities. This type of method introduces the idea of ​​recommendation systems into the decision-making field, but it is mainly applied to relatively macroscopic and semi-structured problems, and it is difficult to directly apply it to the millisecond-level, dynamic, and adversarial real-time decision-making scenarios of unmanned systems.

[0058] (4) Planning and rule-based methods: Decision-making is carried out through a predefined rule base or planning strategy. For example, logic based on behavior trees or finite state machines is written for unmanned systems, or symbolic AI methods such as the Soar cognitive architecture are adopted. The Soar architecture simulates the human cognitive decision-making process through production rules, treating problem-solving as a search in the problem space. Its decision-making process conforms to the OODA (Observe-Adjust-Decision-Action) cycle, and has good logic, interpretability, and real-time performance. Pure planning and rule-based methods rely heavily on expert knowledge. The "preference values" of rules and operators are usually statically preset and cannot be adaptively adjusted according to the dynamic changes in the battlefield environment and practical experience. When new situations or tactics that experts have not anticipated occur, the system performance will drop sharply, and the scalability and robustness are insufficient, resulting in low reliability of intelligent decision-making.

[0059] Therefore, to address the limitations of the Soar architecture in operator selection, which relies on static and preset preference values, this application provides an intelligent recommendation decision-making method and system based on graph neural networks and a cognitive architecture. This solution, using the Soar cognitive architecture as the main decision-making framework, introduces an intelligent recommendation model based on a graph convolutional network (GCN) to dynamically generate preference values ​​for candidate operators, thereby optimizing action selection. The intelligent recommendation model learns from historical data offline and dynamically scores and ranks the candidate actions (operators) generated by Soar online, thus achieving data-driven adaptive decision-making.

[0060] The intelligent recommendation decision-making method based on graph neural networks and cognitive architecture provided in this application relates to the field of intelligent decision-making technology for autonomous mobile systems. This method can be applied to terminals, servers, or software running on either a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, or intelligent terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the intelligent recommendation decision-making method based on graph neural networks and cognitive architecture, but is not limited to the above forms.

[0061] The intelligent recommendation decision-making method based on graph neural networks and cognitive architecture in this application is applied to an autonomous mobile system, where the autonomous mobile entity is referred to as an intelligent agent. The autonomous mobile system in this embodiment is also called an unmanned system, which can be a drone system or an unmanned vehicle system. In a drone system, the intelligent agent is the drone; in an unmanned vehicle system, the intelligent agent is the unmanned vehicle.

[0062] Please refer to Figure 1 The embodiments of this application provide an intelligent recommendation decision-making method based on graph neural networks and cognitive architecture, including but not limited to S101 to S104.

[0063] S101, based on the agent's detection data of the current environmental state, perform situational awareness knowledge reasoning to obtain the situational awareness structure.

[0064] Soar views the problem-solving process as a series of state-space transformations. Therefore, Soar first acquires the agent's current environmental state and then performs situational awareness reasoning based on the probe data of the current environmental state to obtain a cognitive structure reflecting the agent's perception of the external environment. Probe data refers to the information about the external environment and system state of the unmanned system acquired by the agent at the current decision-making moment. An unmanned system refers to a collaborative or adversarial system composed of multiple agents. The external environment can be information such as terrain and obstacles outside the system. The system state refers to the state information of a specific agent within the system and the state information of other agents (such as the drone's position, speed, and attitude). By performing situational awareness reasoning on the probe data based on the Soar architecture, a structured representation of the current environmental state (i.e., the situational awareness structure) can be obtained.

[0065] In some embodiments, S101 may include, but is not limited to, S201 to S203:

[0066] S201, Acquire the agent's detection data on the current environmental state; the detection data includes the external environment and the system state;

[0067] S202, Preprocess the probe data to obtain probe memory elements; the preprocessing includes at least feature processing and symbolic processing;

[0068] S203, based on the cognitive situation rules in the cognitive architecture, performs situational cognitive knowledge reasoning on the probe memory elements to obtain the situational cognitive structure; the situational cognitive structure is used to represent the current situational information in a structured form.

[0069] The agent acquires probe data about the external environment, such as information about the external environment, itself, allies, and enemies. Preprocessing operations, such as feature generation and symbolization, are performed on the probe data to form short-term memory elements (i.e., probe memory elements). Based on the situational awareness rules in the Soar kernel, these short-term memory elements are integrated and inferred to form a structured situational awareness structure. (Cognitive Structure). For example, It can be represented as {self_health:high,num_enemies_visible:2, terrain:open, formation:loose}.

[0070] S102, perform task planning knowledge reasoning on the situational awareness structure to obtain all matching task planning rules, and generate operator structures based on each matching task planning rule; the operator structures represent the agent's actions.

[0071] The situational awareness structure generated by S101 is matched with the conditional part of the task planning rules in the Soar kernel to find all task planning rules that match the current situation. A corresponding operator structure is generated for each matching rule, forming a candidate operator set. The operator structure includes agent actions and may also include the name and conditions of the matched task planning rule. In related technologies, after the task planning rule matching operation, corresponding preference values ​​are usually matched as well; that is, the generated operator structure also includes the matched preference values. For example, the operator structure {Operator_Attack, Operator_Flank, Operator_TakeCover}. Since the preference values ​​are dynamically generated in this embodiment, the "preference value" field in the operator structure generated after task planning rule matching is empty or ignored.

[0072] S103 uses a graph neural network-based intelligent recommendation model to perform matching degree inference on the situational awareness structure and operator structure, thereby obtaining the dynamic preference value of the operator structure.

[0073] Traditional symbolic methods such as the Soar architecture, while possessing learning mechanisms (e.g., chunking), are inefficient at utilizing large-scale, high-dimensional historical adversarial data. The complex, nonlinear "state-action" latent correlation patterns inherent in vast amounts of simulation and combat data remain largely unexplored to optimize decision preferences. This embodiment, in addition to using an intelligent recommendation model to perform matching degree inference on the situational awareness structure and operator structure to obtain dynamic preference values ​​for the operator structure, also employs an intelligent recommendation model based on a graph neural network architecture (e.g., LightGCN) to process the situational awareness structure and operator structure. This fully leverages the large amounts of data generated by multi-unmanned systems in historical adversarial or collaborative scenarios, thereby learning and mining the optimal action strategies under different battlefield situations to improve the accuracy and foresight of decision-making.

[0074] In some embodiments, S103 may include, but is not limited to, S301 to S302:

[0075] S301, the situational awareness structure and the operator structure are represented by embedding vectors through the intelligent recommendation model, and the first embedding vector and the second embedding vector are obtained respectively.

[0076] S302, perform a dot product operation on the first embedding vector and the second embedding vector to obtain the dynamic preference value of the operator structure.

[0077] The intelligent recommendation model is an embedding vector library containing embedding vectors of all known situational awareness structures and operator structures. The first embedding vector of the situational awareness structure and the second embedding vector of the operator structure can be determined by embedding vector retrieval.

[0078] Specifically, firstly, regarding the current situational awareness structure generated by S101... The first embedding vector is found or calculated in the trained intelligent recommendation model. For each structure operator in the candidate operator set generated by S102 Retrieve the corresponding second embedding vector from the intelligent recommendation model. .

[0079] Next, the recommendation score (i.e., dynamic preference value) is calculated for each generated candidate operator structure. Calculate its relationship with the current situation The matching score (i.e., recommendation score) is calculated as the dot product of the embedding vectors: Score( )= Further, the calculated recommendation score (Score) Enter or update the "Preference Value" field in the corresponding operator structure.

[0080] In some embodiments, the intelligent recommendation model in S103 can be trained through S401 to S404:

[0081] S401, collect the training dataset when the unmanned system is working; each sample data in the training dataset includes the agent's situational awareness structure, operator structure, and decision reward for executing the operator structure at the decision-making time.

[0082] Collect large amounts of adversarial data from multiple agents in unmanned systems, or large amounts of collaborative data from multiple agents in unmanned systems, from simulation environments or actual exercises. Organize these data into sample data, with each sample data record being a triplet. ,in:

[0083] This represents the situational awareness structure of an intelligent agent before making a decision. It is a symbolic and structured information formed after reasoning from raw sensor data (such as enemy and friendly positions, speeds, terrain, and the agent's own state) using Soar's situational awareness rules. For example, It can be represented as {self_health:high,num_enemies_visible:2,terrain:open,formation:loose}.

[0084] This represents the action operator (i.e., the corresponding action in the operator structure) that the agent ultimately chooses and executes in this situation. The operator structure is an encapsulation of an atomic or macroscopic action in the Soar architecture. For example, You can use {op_attack_threat1} (attack threat target #1) or {op_take_cover} (find cover).

[0085] This represents the decision reward obtained from the environmental feedback after executing the operator. This reward is used to evaluate the merits of the decision and can be a quantitative indicator such as task completion rate or survival time. For example, =+0.8 (indicating a significant advantage).

[0086] Furthermore, more complex reward functions can be designed, such as those based on the battle damage exchange ratio and mission completion time.

[0087] S402, construct a graph structure on the training dataset to obtain a situation-operator bipartite graph; the node set of the situation-operator bipartite graph includes the situation node set and the operator node set, and the edge set of the situation-operator bipartite graph represents the historical execution of operator node j by situation node i, and the edge weight represents the corresponding decision reward.

[0088] Please refer to Figure 2 The diagram shown illustrates the situation-operator bipartite graph. Based on the collected training dataset, a situation-operator bipartite graph G=(V,E) is constructed, where:

[0089] Node set V: Contains two types of nodes: the situation node set U and the operator node set I. Each unique situation awareness structure Corresponding to a situation node Each unique operator Corresponding to one operator node .

[0090] Edge set E: If historically in the situation Operators executed below Then at the corresponding node and Establish an edge between them. The weight of the edge can be determined by the decision reward. Decisions are made such that, for example, the more positive outcomes (i.e., the greater the reward of the decision), the higher the weight. Under the simplified idea of ​​LightGCN, only interaction relationships can be retained, that is, only the existence of edges (interactions) are recorded, without using explicit weights.

[0091] Furthermore, to enhance the information carrying capacity of the graph, each edge is assigned a weight W(Sᵢ, Oⱼ). This weight W is a comprehensive evaluation value, which can be calculated using the function W=f(R, C) Calculate, where R is the decision reward. To account for the time decay factor, C represents the confidence level of this historical data point. The weight calculation process takes into account the time decay factor of decision rewards, ensuring that recent and high-reward decision experiences have a higher weight in the training of the intelligent recommendation model. For example... Figure 2 As shown, the situation-operator bipartite graph constructed by the above process is the basis for subsequent LightGCN model learning.

[0092] S403, the situation-operator bipartite graph is input into the intelligent recommendation model based on graph convolutional network for embedding propagation and neighbor aggregation to obtain the embedding vector of each node; the embedding vector is used to characterize the matching degree between situation and operator.

[0093] To improve the processing efficiency of graph structures in intelligent recommendation models based on graph convolutional networks, node encoding can be performed on the situation-operator bipartite graph before inputting it into the intelligent recommendation model, as follows:

[0094] Each node is assigned an initial feature vector (i.e., the initial feature set X), enabling the graph neural network model to learn not only the graph's topology but also the nodes' intrinsic properties. The initial feature set includes state node features. and operator node features :

[0095] Situation node characteristics The symbolic situational awareness structure S is further transformed into a multi-dimensional real vector. For example, for {health:high,num_enemies:2,terrain:open}, it can be mapped to the vector [1.0,0.85,2.0,1,0,0], where the first two values ​​represent the quantified health and weapon status, the third is the number of enemies, and the last three are the one-hot encoding of the terrain.

[0096] Operator node features Encode the intrinsic properties of the operator. For example, the operator {op_attack} can be encoded as a vector [1,0,0,10,50], where the first three bits are the one-hot encoding of the operator type, and the last two bits represent the standard time and energy consumption required to execute the operator, respectively.

[0097] After encoding the situation-operator bipartite graph, the LightGCN model is used for training. LightGCN is a lightweight graph convolutional network that only includes the neighborhood aggregation and propagation parts of graph convolution, removing feature transformation and nonlinear activation functions, making it suitable for collaborative filtering and recommendation tasks. The forward propagation process of the LightGCN model includes embedding layer node initialization, LightGCN information propagation, and final embedding representation, as detailed below:

[0098] Embedding Layer and Initialization: In the embedding layer, for each situation node s∈U_S and operator node o∈ in graph G... Initialize a random low-dimensional embedding vector. and The 0th layer embedding of the model

[0099] (1)

[0100] in, and It is a randomly initialized d-dimensional dense vector.

[0101] LightGCN Information Dissemination: The core of LightGCN lies in its extremely simple neighborhood information aggregation and dissemination mechanism. Please refer to... Figure 3 The diagram illustrates the node information aggregation process. The model updates node embeddings through multi-layer (e.g., K-layer) graph propagation. For any node (e.g., a situation node s), its (k+1)th layer embedding is formed by aggregating the k-th layer embedding and the embeddings of its neighboring nodes. and operator nodes The rules for information dissemination are expressed as follows:

[0102] (2)

[0103] (3)

[0104] Where N(s) represents the set of neighboring operator nodes of state node s, and N(o) represents the set of neighboring state nodes of operator node o. It is a symmetric normalization term used to prevent node degree from having an excessive impact on the aggregation result and to enhance training stability.

[0105] Final Embedded Representation: After K layers of propagation, each node will receive K+1 embedded representations. The final embedded representation is obtained by weighted aggregation (e.g., direct addition or weighted averaging) of these embeddings from different layers, as follows:

[0106] (4)

[0107] (5)

[0108] in, These are the weight coefficients of the k-th layer, which can be set to 1 / (K+1). Thus, the final embedding vector... and It integrates the node's own information and the structural information of its multi-level neighborhood.

[0109] S404. Update the parameters of the intelligent recommendation model in reverse according to the embedding vector of each node to obtain the trained intelligent recommendation model.

[0110] After inputting the situation-operator bipartite graph into the intelligent recommendation model based on graph convolutional network for forward embedding propagation and neighbor aggregation, the embedding vector of each node is obtained. The model is then updated in reverse according to the embedding vector output by the model so that the model can accurately represent the matching degree of the node (i.e., the embedding vector).

[0111] In some embodiments, S404 may include, but is not limited to, S501 to S502:

[0112] S501, For each situation node, sample the positive sample operators it has interacted with and the negative sample operators it has not interacted with from the training dataset, and calculate the prediction preference values ​​of the situation node with the positive sample operators and the negative sample operators respectively.

[0113] S502, with the optimization objective being to make the predicted preference value of the positive sample operator greater than that of the negative sample operator, continuously optimizes the parameters of the intelligent recommendation model until the optimization objective is achieved, thus obtaining a well-trained intelligent recommendation model.

[0114] This embodiment can use Bayesian Personalized Ranking (BPR) loss as the optimization objective. The core idea of ​​BPR loss is that, for a given situation s, the model should predict a higher score (i.e., a predicted preference value) for the positive sample operators o+ that the model has interacted with (i.e., the predicted preference value) than for the negative sample operators o- that the model has not interacted with. The calculation is performed using the dot product of the situation and the operator embedding vector, as follows:

[0115] (6)

[0116] BPR loss function Defined as:

[0117] (7)

[0118] in, It is a set of triples (position, positive sample operator, negative sample operator) sampled from the training dataset, where σ(·) is the sigmoid function. It is minimized using optimization algorithms such as gradient descent (e.g., Adam). The model parameters are continuously adjusted until an embedding representation that accurately reflects the "situation-operator" matching degree is learned. After training, a final embedding vector library containing all known situations and operators is obtained. This embedding vector library is the intelligent recommendation model of this application embodiment.

[0119] S104: Select the optimal operator structure based on the dynamic preference values ​​of each operator structure, and determine the decision action.

[0120] In one example, during action selection, all candidate operator structures can be sorted in descending order of dynamic preference values ​​based on the action selection mechanism in the Soar kernel, and the operator structure with the highest score can be selected as the optimal operator structure. The decision action is determined based on the actions in the optimal operator structure.

[0121] In another example, to improve the reliability of decision-making actions, uncertainty handling and Soar deadlock linkage can be implemented during action selection. A more deterministic operator structure can be selected as the optimal operator structure, and then the decision-making action can be determined based on the actions within this optimal operator structure. Specifically, S104 may include, but is not limited to, S601 to S604:

[0122] S601, according to the dynamic preference value from high to low, the operator structure with the highest ranking is determined as the first candidate operator and the operator structure with the second ranking is determined as the second candidate operator.

[0123] S602, determine the confidence level based on the difference between the dynamic preference values ​​of the first candidate operator and the second candidate operator;

[0124] S603, if the confidence level is greater than or equal to the expected threshold, determine the decision action based on the first candidate operator;

[0125] S604. When the confidence level is less than the expected threshold, simulate and deduce the first candidate operator and the second candidate operator respectively to obtain the corresponding simulated returns. Select the candidate operator with the higher simulated return from the first candidate operator and the second candidate operator to determine the decision action.

[0126] First, the candidate operator set (including the matched operator structures mentioned above) is sorted from high to low according to the dynamic preference value to obtain the optimal candidate operator. (i.e., the first candidate operator) and the second-best candidate operator (Second candidate operator). Calculate the score difference between the two. The score difference represents the confidence level.

[0127] Secondly, With a preset "uncertainty threshold" The comparison is performed against the expected threshold, and the results and actions are as follows:

[0128] Scenario 1 (High Confidence Decision): If ≥ This indicates that the graph neural network model has a high degree of confidence in the optimal choice. At this point, the system directly... Select the optimal operator structure and determine the actions in the optimal operator structure as decision actions.

[0129] Scenario 2 (Low Confidence Decision): If Δ_score < This indicates that the model considers multiple options to be good and cannot make a clear choice. This is tactically dangerous. At this point, the system does not make a direct choice but triggers the Soar kernel's "impasse" mechanism and executes the following sub-steps:

[0130] Sa (Entering a Substate): Soar automatically creates a substate, the goal of which is to... and Make a more reliable choice between them.

[0131] Sb (Lookahead): In this substate, the system invokes one or more "lookahead operators." These operators simulate execution within an internal, lightweight world model. and The potential short-term environmental changes and return assessments (i.e. simulated returns) that may occur later.

[0132] Sc (Solving the Deadlock): Compare the deduction results and select the operator structure with the higher deduction reward as the solution to the deadlock. The selected operator structure is returned to the upper-level state, and then this operator structure is taken as the optimal operator structure. The action in the optimal operator structure is determined as the decision action.

[0133] In some embodiments, the intelligent recommendation decision-making method based on graph neural networks and cognitive architecture in this application may also include, but is not limited to, S701 to S703:

[0134] S701, Obtain the decision feedback from the environment after performing a decision action in the current environmental state;

[0135] S702, new sample data are formed based on the situational awareness structure, decision-making actions, and decision-making rewards of the current environmental state;

[0136] S703 performs periodic batch training on the intelligent recommendation model based on multiple new sample data to optimize the intelligent recommendation model.

[0137] To enable the agent to continuously evolve and become stronger during task execution, after the agent executes the final decision action determined in S104, the environmental state changes, and the agent obtains new state information, i.e., the state transition result. That is, the state transition result is the current environmental state information changed after the agent executes the action corresponding to the optimal operator structure. S101-S104 are repeatedly executed to form a closed-loop OODA decision-making process. The (S,O,R) triples of each decision are recorded for future incremental updates or periodic retraining of the recommendation model, achieving continuous model evolution. Specifically, this embodiment establishes a three-layer learning closed loop including "real-time caching, online fine-tuning, and periodic retraining" to ensure the continuous evolution of the model, as follows:

[0138] Real-time experience cache: The complete record (S, O, R) of each decision is asynchronously sent to a "recent experience cache pool". This cache pool has a fixed size and adopts a first-in, first-out (FIFO) strategy.

[0139] Online Fine-tuning: During periods of low system load or task downtime, an online fine-tuning task is automatically triggered. This task samples a mini-batch of new data from the "recent experience cache pool" and performs gradient updates on some parameters (such as the last layer or fully connected layers) of the loaded graph neural network model (i.e., the intelligent recommendation model) over several rounds. This allows the model to quickly adapt to the latest patterns emerging in the battlefield, offering the advantage of fast response speed.

[0140] Periodic Re-training: To prevent performance degradation (i.e., catastrophic forgetting) caused by shifts in the distribution of new data during continuous fine-tuning, the system periodically (e.g., every N task cycles or every 24 hours) initiates a full background retraining task. This task merges the original offline training dataset with all new data in the "recent experience cache pool," performs a more thorough retraining of all parameters of the graph neural network model from scratch, and then replaces the currently used online model with the newly trained model.

[0141] According to some embodiments of this application, please refer to Figure 4 The schematic diagram shown illustrates the overall technical concept of the agent decision-making process in this application, and introduces the overall technical concept of the agent decision-making process in this embodiment.

[0142] The technical concept of this application mainly includes two stages: an offline model training stage and an online recommendation decision-making stage. The goal of the offline stage is to train an intelligent recommendation model based on historical data; the online stage involves deeply embedding the recommendation model into the Soar decision loop and establishing a collaborative decision-making and continuous evolution mechanism to achieve truly adaptive decision-making.

[0143] Phase 1: Offline Model Training Phase – Building an Intelligent Recommendation Model. The offline model training phase aims to build an intelligent recommendation LightGCN model capable of accurately recommending optimal action operators based on battlefield conditions. This phase includes the following steps:

[0144] Step S11: Historical decision data collection and preprocessing.

[0145] First, large-scale historical decision-making data is collected through multi-unmanned system simulation platforms (such as Unity, VBS, etc.) or actual exercises. To ensure data quality, the data can be filtered, for example, by removing invalid or incomplete adversarial data. Each valid decision data record is abstracted into a triple. .

[0146] Step S12: Construct a situation-operator bipartite graph.

[0147] The historical decision data collected and preprocessed in step S11 is used to construct a situation-operator bipartite graph, such as... Figure 2 The situation-operator bipartite diagram shown has situation nodes on the left and operator nodes on the right. If the situation... Down-execution operator and If all operations are successful, then the corresponding edges are connected. This graph structure can explicitly capture the higher-order relationships between different situations and different operators. The situation-operator bipartite graph is the foundation for subsequent LightGCN model learning.

[0148] Step S13: Training the recommendation model based on LightGCN.

[0149] A lightweight graph convolutional network (LightGCN) is used as a key component of the intelligent recommendation model. It trains the bipartite graph constructed in step S12 to learn low-dimensional embedding vectors for each situation node and operator node. After K layers of propagation, each node receives K+1 embedding representations, resulting in the final embedding vectors. and It integrates the node's own information and the structural information of its multi-level neighborhood.

[0150] The Bayesian Personalized Ranking (BPR) loss is used as the optimization objective. The core idea of ​​BPR loss is that, for a given situation s, the model predicts that the score of the positive sample operator o+ that the user has interacted with should be higher than the score of the negative sample operator o- that the user has not interacted with. This is minimized using optimization algorithms such as gradient descent (e.g., Adam). The model parameters will be continuously adjusted until an embedding representation that can accurately reflect the "situation-operator" matching degree is learned.

[0151] The second stage, the online decision-making and recommendation stage—the Soar decision-making loop of fusion recommendation—represents the decision-making process of the intelligent agent in actual operation, and optimizes and improves the traditional Soar decision-making process. The online decision-making and recommendation stage embeds the trained intelligent recommendation model into the real-time decision-making loop of the multi-unmanned system agent. This loop can be viewed as a combination of the Soar cognitive decision-making cycle and the OODA loop.

[0152] Step S21: Situational awareness reasoning.

[0153] The agent acquires real-time detection data of the external environment through its sensors, which constitutes its short-term working memory. Situational awareness rules in the Soar kernel are triggered, matching, integrating, and reasoning about this raw, fragmented data, ultimately forming a structured, symbolic current cognitive structure. The output of this step is related to the situation during the offline training phase. The format is consistent.

[0154] Step S22: Candidate operator generation.

[0155] Current cognitive structure This is used to match task planning rules in the Soar kernel. The form of a task planning rule is IF{condition}THEN{proposal operator}. All condition parts are related to... Matching rules are activated, and their proposed operators are collected to form a candidate operator set. In this step, the static preference value of the operator is ignored.

[0156] Step S23: Intelligent Recommendation and Operator Optimization. This step replaces the traditional Soar action selection based on static preference values, improving the reliability of agent decision-making under the cognitive structure, as detailed below.

[0157] Embedded vector retrieval: The system loads the offline-trained intelligent recommendation model (i.e., the final embedded vector library for all nodes). For the current cognitive structure... Find its corresponding embedding vector For the candidate operator set Each operator in Find its corresponding embedding vector .

[0158] Recommended score calculation: Each candidate operator in Using equation (6), its relationship with the current situation is calculated. The matching score, which is the dynamic recommendation score. .

[0159] Operator preference setting: The calculated dynamic recommendation score Assign to candidate operators , which serves as its dynamic preference value in the current context.

[0160] Step S24: Action selection.

[0161] The operator structures in the candidate operator set are sorted from high to low according to the dynamic preference values. The optimal operator structure is determined through uncertainty handling and Soar deadlock linkage, thereby determining the decision action.

[0162] Step S25: Action execution and data feedback.

[0163] The agent executes step S24 to determine the optimal operator structure. The corresponding actions interact with the environment, causing changes in the environmental state.

[0164] Step S26: Data feedback and model evolution.

[0165] Repeatedly executing the three-layer learning closed-loop mechanism established in the embodiments of this application, which includes "real-time caching, online fine-tuning, and periodic retraining," enables the intelligent recommendation model to continuously evolve.

[0166] In some embodiments, please refer to Figure 5 This diagram illustrates a data flow diagram of an intelligent recommendation decision-making method based on graph neural networks and a cognitive architecture. Specifically, external environment information is provided by Unity, and the agent obtains the necessary information by interacting with the Unity simulation environment. The pre-processed information, i.e., symbolic environment information, is transmitted to the Java environment via a socket. In the Java environment, the symbolic environment information undergoes secondary processing, being organized into symbolic feature information that conforms to the needs of the Soar kernel. This information is then transmitted to the intelligent recommendation module for recommendation scoring and operator optimization, and finally to the Soar kernel for decision matching and output. The Soar kernel stores instructions that help the agent understand the external environment and tasks, including situational awareness rules, task planning rules, and recommendation-based action selection. The Soar kernel receives information from the external environment, infers decisions through knowledge, and outputs the decision results, i.e., the Soar file, to the Java environment via an SML interface for instruction parsing. The Java environment then packages the decision results and transmits them back to the agent in the Unity simulation environment via a socket. The agent parses the decision and executes the corresponding actions.

[0167] To verify the effectiveness of the intelligent recommendation decision-making method based on graph neural networks and cognitive architecture proposed in this embodiment, a 2v2 forest adversarial simulation environment based on the Unity engine was built. The red team agents (Agent-R1, Agent-R2) adopted the Soar decision-making system integrating intelligent recommendation proposed in this embodiment, while the blue team agents (Agent-B1, Agent-B2) adopted the traditional Soar decision-making system based on static preference value rules.

[0168] The test simulation environment includes an action control model, an equipment model, and environmental data. Specifically, this embodiment establishes a forest environment model, such as... Figure 6As shown, the dark green areas in this environment represent woodland areas. The dense trees in these areas easily obscure visibility, making them ideal for reconnaissance. The gray areas (Attack Area 1 and Attack Area 2) represent designated attack points. If target information is detected at these points, attack agents will be deployed to attack them. The red areas (Red Team Birthplace) and blue areas (Blue Team Birthplace 1 and Blue Team Birthplace 2) represent the spawn points of the red and blue teams, respectively. Furthermore, in this embodiment, the red team's spawn point has a direct attack route to the designated target gray area, and the blue team's spawn point has a direct defense route to contest the designated target gray area. Both the red and blue teams are at a certain distance from the gray areas. Woodland is randomly distributed in any area outside the spawn points, contest points, and mandatory routes.

[0169] Specifically, this embodiment employs a minimum-scale two-vehicle formation applied to a 2V2 multi-distributed unmanned system combat scenario. To verify the feasibility and effectiveness of the technical solution, this embodiment equips each unmanned vehicle with an independent decision-making model, and the rationality of the decision-making model is verified through simulation results. Simultaneously, the execution time of a single decision by the multi-unmanned system is recorded to evaluate the performance of the multi-unmanned system.

[0170] It should be noted that the initial model run involves additional initialization work, such as loading the model and establishing connections. Therefore, the time taken for the first run is significantly longer than for subsequent single decisions. To mitigate this impact, this embodiment calculates the execution time for every 100 subsequent decisions, excluding the initial run. The resulting average single decision time for the model with all decision rules is 1.2–1.5 milliseconds, ensuring that the simulation results fully meet the rapid response requirements of online real-time decision-making.

[0171] Furthermore, the experimental data in this embodiment includes: the number of wins by the red team, the number of wins by the blue team, the number of draws, the total number of matches, and the red team's winning percentage. Task planning rules are established based on task planning knowledge and action prioritization planning decision-making knowledge, including:

[0172] The mission planning rules are as follows: When our forces have not reached the target location: if our condition is good, we advance; if our condition is poor, we rest. When our forces have reached the target location: if our condition is good, we fire; if our condition is poor, we rest.

[0173] The formation change decision rules are as follows: When the terrain is open: advance in a straight line formation, and fire in a straight line formation or a cross formation; when the terrain is not open: advance in a cross formation or a vertical formation, and fire in a vertical formation.

[0174] The formation movement decision rule is as follows: when our unmanned systems are slightly spaced apart, the unmanned systems in front in the direction of movement slow down to wait for their teammates behind, while the unmanned systems behind accelerate to catch up with their teammates in front, thus achieving a tight formation.

[0175] The fire control strike rules based on game theory and adversarial competition specifically compare the threat level of all enemy units with their target value. When a high-threat target exists, it is prioritized for elimination; otherwise, relatively high-value targets are prioritized for elimination. Specifically, an enemy target is classified as a high-threat target only if its threat level exceeds a preset threshold.

[0176] 1. Experimental Environment and Parameter Settings

[0177] Simulation environment: Unity 3D, including various terrain elements such as woodlands, open areas, and bunkers.

[0178] Intelligent agent model: Unmanned vehicle, with basic capabilities such as movement, rotation, firing (with cooldown time), and sensor detection.

[0179] Soar rule base: A basic Soar rule base is pre-configured for all agents, including:

[0180] (1) Situational cognition rules: used to transform sensor data (such as health:0.0-1.0, enemy_distance:0-1000m) into symbolic cognitive structures (such as {health:high / medium / low}, {enemy_contact: close / medium / far}).

[0181] (2) Task planning rules: used to generate candidate operators based on cognitive structure, such as IF{enemy_contact:close}AND{weapon_ready:true}THEN PROPOSE{op_attack}.

[0182] LightGCN model parameters:

[0183] Embedded vector dimension d: 64;

[0184] Number of convolutional layers K: 3;

[0185] Optimizer: Adam;

[0186] Learning rate: 0.001;

[0187] 2. Specific implementation of the offline model training phase

[0188] Step 1: Generation of historical decision dataset.

[0189] First, conduct 1000 adversarial simulations of "blue team vs. blue team," where all four agents use the traditional Soar decision system. In each decision cycle of each adversarial match (e.g., every 500 milliseconds), record the decision triples (S, O, R) for each agent.

[0190] Example of a data entry: Suppose that Agent-B1 makes a decision at the 15th second of a certain match.

[0191] Raw sensor data: {health:0.85,ammo:1,nearest_enemy_dist:123.5,nearest_enemy_angle:-15.2,terrain_under:forest}.

[0192] S (Situational Awareness Structure): After reasoning using the Soar situational awareness rules, a symbolic S_158={health:high,weapon:ready,contact:medium,angle:frontal,terrain:cover} is generated.

[0193] O (Operator to be executed): Agent-B1's traditional Soar system selects the {op_attack} operator based on its static preference value.

[0194] R (Reward): The Blue team won the match, so the reward R for this record is marked as +1.

[0195] After 1,000 simulations, approximately 200,000 such triplet data points were collected, forming a historical decision dataset.

[0196] Step 2: Construction of the situation-operator bipartite graph and model training.

[0197] See Figure 2 The collected dataset is then processed.

[0198] Node extraction: All unique situational awareness structures that have appeared in the statistical data set were collected, creating |US|=35,240 situational nodes; All unique action operators (a total of 12 types, such as attack, movement, finding cover, changing formation, etc.) were counted, creating |UO|=12 operator nodes.

[0199] Edge construction: Traverse all data entries with a reward of +1 and establish edges between the corresponding state nodes and operator nodes. This ultimately constructs a bipartite graph G containing 3525 nodes and approximately 120,000 edges.

[0200] Model Training: The bipartite graph G is input into the LightGCN model for training. The training process follows the steps outlined in the previous embodiment. After approximately 50 epochs of training, the BPR loss of the model converges. Upon completion of training, a 64-dimensional embedding vector containing all 3525 nodes is obtained and saved as the recommendation model file.

[0201] 3. Specific Implementation of the Online Decision Recommendation Stage

[0202] The red team's intelligent agent, equipped with the novel system of this invention, is deployed on the battlefield to compete against the blue team. The following demonstrates the decision-making process of the red team's intelligent agent, Agent-R1, in a specific scenario.

[0203] Scenario: Agent-R1 is currently low on health (0.3), its weapon has cooled down, and it encounters a full-health enemy, Agent-B1, at close range in an open area.

[0204] Step 1: Situational awareness.

[0205] Agent-R1's Soar kernel transforms sensor data {health:0.3,ammo:1,enemy_dist:50.0,terrain:open_field} into a cognitive structure S_current={health:low,weapon:ready,contact:close,terrain:open}.

[0206] Step 2: Candidate operator generation.

[0207] In Soar's task planning rule base, the IF conditions of the following two rules are satisfied by S_current:

[0208] Rule_1:IF{contact:close}AND{weapon:ready} THEN PROPOSE {op_attack}

[0209] Rule 2: IF {health:low} THEN PROPOSE {op_take_cover}. Therefore, the generated candidate operator set CO is {{op_attack}, {op_take_cover}}.

[0210] Step 3: Intelligent Recommendation and Optimization.

[0211] The Agent-R1 system performs the following steps:

[0212] Embedding retrieval: From the loaded recommendation model file, retrieve the embedding vector e_current corresponding to S_current, and the embedding vectors e_attack and e_cover corresponding to the two candidate operators {op_attack} and {op_take_cover}.

[0213] Score Calculation: Calculate the recommended score according to equation (6):

[0214] ;

[0215] ;

[0216] Analysis: In traditional Soar systems, if the static preference value of `op_attack` is set too high by experts, the system is likely to choose to attack head-on, resulting in being destroyed by a fully healed enemy. However, the recommendation model in this embodiment of the invention learns from historical data that in situations of "low health, close-range engagement," the number of cases where the user executes "take cover" (`op_take_cover`) and ultimately wins (i.e., positive samples) far exceeds the number of cases where the user executes "attack" (`op_attack`) and wins. Therefore, the model assigns a higher recommendation score to `op_take_cover`.

[0217] Step 4: Action Selection.

[0218] The system compares the scores, and 0.88 > 0.45, therefore {op_take_cover} is selected as the highest priority operator.

[0219] Step 5: Action execution and feedback.

[0220] Agent-R1 successfully evaded Agent-B1's direct fire by searching for and moving to the nearest cover. Agent-R1 can then use the cover advantage to launch a counterattack. The (S,O,R) data for this decision was recorded and transmitted to the historical database.

[0221] 4. Experimental Results and Verification

[0222] To verify the performance of the invention, 1,000 adversarial experiments were conducted between the red team (the invention) and the blue team (traditional Soar), and key performance indicators (KPIs) were statistically analyzed.

[0223] Table 1 Performance Comparison Experiment Results

[0224] Evaluation indicators Red Square (Method of this Invention) Red vs. Blue (Traditional Soar) Win rate (%) 72.3% 27.7% Average survival time (seconds) 175.6 121.4

[0225] As shown in Table 1, the red side achieved a significant win rate of 72.3% using the method of this invention, which is much higher than the 27.7% of the traditional Soar system. The average survival time of the red side's agent was also significantly improved.

[0226] The results analysis shows that the system of this invention exhibits higher tactical intelligence and adaptability in decision-making. It not only executes basic rules but also makes better trade-offs at critical moments based on massive amounts of data and experience. This data-driven dynamic preference evaluation mechanism effectively compensates for the limitations of static rules, enabling multi-unmanned systems to demonstrate stronger survivability and mission completion capabilities in complex adversarial environments.

[0227] This application also proposes an intelligent recommendation decision-making system based on graph neural networks and cognitive architecture, including:

[0228] The first module is used to perform situational awareness knowledge reasoning based on the agent's detection data of the current environmental state to obtain the situational awareness structure;

[0229] The second module is used to perform task planning knowledge reasoning on the situational awareness structure, obtain all matching task planning rules, and generate operator structures based on each matching task planning rule; the operator structures represent the actions of the intelligent agent.

[0230] The third module is used to perform matching degree inference on the situational awareness structure and operator structure through an intelligent recommendation model based on graph neural networks, and obtain the dynamic preference value of the operator structure.

[0231] The fourth module is used to select the optimal operator structure based on the dynamic preference values ​​of each operator structure and determine the decision action.

[0232] Furthermore, the intelligent recommendation decision-making system based on graph neural networks and cognitive architecture in this application embodiment also includes:

[0233] The fifth module is used to control the agent to perform selected actions.

[0234] The sixth module is used to collect decision data and store it in an external database to support the continuous learning of the model.

[0235] It is understood that the methods described in the above method embodiments are applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0236] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any intelligent terminal, including a computer, an unmanned vehicle, or a drone.

[0237] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0238] Please see Figure 7 , Figure 7 This illustration shows the hardware structure of an electronic device according to one embodiment. The electronic device includes:

[0239] The processor 701 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0240] The memory 702 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 702 and is called and executed by the processor 701.

[0241] The input / output interface 703 is used to implement information input and output;

[0242] The communication interface 704 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0243] Bus 705 transmits information between various components of the device (e.g., processor 701, memory 702, input / output interface 703, and communication interface 704);

[0244] The processor 701, memory 702, input / output interface 703, and communication interface 704 are connected to each other within the device via bus 705.

[0245] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0246] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0247] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0248] According to some embodiments of this application, the intelligent recommendation decision-making method, system, electronic device, and storage medium based on graph neural networks and cognitive architecture of this application have at least one of the following beneficial effects:

[0249] Combining interpretability and adaptability: The decision-making process in this solution follows a clear logic of "situational awareness -> rule matching -> operator selection." The rules themselves are defined by experts, ensuring the interpretability of basic behaviors. Simultaneously, by introducing the LightGCN recommendation model, the priority of operators is learned from historical data, enabling decisions to adapt to the constantly changing battlefield environment and opponent strategies, thus solving the rigidity problem of traditional rule systems.

[0250] Significantly enhanced decision robustness and reliability: Through a unique "collaborative decision-making" mechanism, when the output of the data-driven recommendation model is uncertain, the system can intelligently revert to forward-looking inference based on symbolic reasoning, utilizing Soar's classic planning capabilities to resolve decision deadlocks. This deep integration model of "machine learning suggestions and cognitive architecture adjudication" effectively avoids hasty decisions in critical ambiguity situations, greatly improving the system's reliability and robustness in highly adversarial environments.

[0251] Enhancing the dynamism and accuracy of decision-making: It replaces static, manually set preference values ​​with data-driven dynamic recommendation scores. Furthermore, by introducing node attributes and edge weights, the graph construction becomes more refined. The model can uncover non-linear "situation-action" potential relationships from richer information dimensions that are difficult for experts to detect, thereby making more accurate and better tactical choices in complex situations.

[0252] It achieves true continuous learning and online evolution: a complete data loop including "real-time caching, online fine-tuning, and periodic retraining" has been established. This enables multi-agent systems not only to learn offline using historical big data, but also to quickly adapt to new tactics and environments during task execution, truly achieving "replay" and "getting stronger with each battle".

[0253] The system's scalability and maintainability are improved: when new tactics or actions need to be introduced, the corresponding task planning rules (generating new operators) can be added to Soar, and then the recommendation model can automatically learn its applicable scenarios and priorities by collecting data, without having to manually adjust the preference values ​​of a large number of related rules, which greatly reduces the complexity and cost of system maintenance.

[0254] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0255] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0256] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate; 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.

[0257] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0258] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0259] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0260] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An intelligent recommendation decision-making method based on a graph neural network and a cognitive architecture, characterized in that, The method comprises the following steps: According to the detection data of the agent on the current environment state, the situation awareness knowledge reasoning is carried out to obtain a situation awareness structure; The situation awareness structure is subjected to task planning knowledge reasoning to obtain all matched task planning rules, and an operator structure is generated according to each matched task planning rule; the operator structure represents the action of the agent; Through the intelligent recommendation model based on the graph neural network, the matching degree reasoning is carried out on the situation awareness structure and the operator structure to obtain the dynamic preference value of the operator structure; According to the dynamic preference value of each operator structure, the optimal operator structure is selected to determine the decision action; The intelligent recommendation model based on the graph neural network is used to carry out the matching degree reasoning on the situation awareness structure and the operator structure to obtain the dynamic preference value of the operator structure, which comprises the following steps: The intelligent recommendation model is used to respectively carry out embedding vector representation on the situation awareness structure and the operator structure to obtain a first embedding vector and a second embedding vector; Dot product operation is carried out on the first embedding vector and the second embedding vector to obtain the dynamic preference value of the operator structure; The intelligent recommendation model is trained through the following steps: A training data set is collected when the unmanned system is working; each sample data in the training data set comprises a situation awareness structure of an agent at a decision time, an operator structure and a decision reward of executing the operator structure; The training data set is subjected to graph structure construction to obtain a situation-operator bipartite graph; the node set of the situation-operator bipartite graph comprises a situation node set and an operator node set, and the edge set of the situation-operator bipartite graph represents that the situation node i has executed the operator node j in history, and the edge weight represents the corresponding decision reward; The situation-operator bipartite graph is input into the intelligent recommendation model based on the graph convolution network for embedding propagation and neighbor aggregation to obtain an embedding vector of each node; the embedding vector is used to represent the matching degree between the situation and the operator; The parameters of the intelligent recommendation model are updated in reverse according to the embedding vector of each node to obtain the trained intelligent recommendation model.

2. The intelligent recommendation decision method based on a graph neural network and a cognitive architecture according to claim 1, characterized in that, The situation awareness knowledge reasoning is carried out according to the detection data of the agent on the current environment state to obtain a situation awareness structure, which comprises the following steps: Detection data of an agent on a current environment state is obtained; the detection data comprises a system external environment and a system state; The detection data is preprocessed to obtain a detection memory element; the preprocessing at least comprises feature processing and symbolization processing; Based on the cognitive situation rules in the cognitive architecture, the situation awareness knowledge reasoning is carried out on the detection memory element to obtain a situation awareness structure; the situation awareness structure is used to represent the current situation information in a structured form.

3. The intelligent recommendation decision method based on a graph neural network and a cognitive architecture according to claim 1, characterized in that, The parameters of the intelligent recommendation model are updated in reverse according to the embedding vector of each node to obtain the trained intelligent recommendation model, which comprises the following steps: For each situation node, the positive sample operator that has interacted and the negative sample operator that has not interacted are sampled from the training data set, and the predicted preference value of the situation node with the positive sample operator and the negative sample operator is calculated respectively; The optimization target is that the prediction preference value of the positive sample operator is greater than the prediction preference value of the negative sample operator, and the parameters of the intelligent recommendation model are continuously optimized until the optimization target is reached, and a trained intelligent recommendation model is obtained.

4. The intelligent recommendation decision method based on a graph neural network and a cognitive architecture according to any one of claims 1 to 3, characterized in that, The optimal operator structure selection according to the dynamic preference value of each operator structure determines the decision action, including the following steps: According to the dynamic preference value, the operator structure with the highest ranking is determined as the first candidate operator and the operator structure with the second highest ranking is determined as the second candidate operator; According to the difference between the dynamic preference values of the first candidate operator and the second candidate operator, the confidence is determined; In the case where the confidence is greater than or equal to the expected threshold, the decision action is determined according to the first candidate operator; In the case where the confidence is less than the expected threshold, the first candidate operator and the second candidate operator are simulated and deduced respectively to obtain corresponding simulation returns, and the candidate operator with higher simulation return is selected from the first candidate operator and the second candidate operator to determine the decision action.

5. The intelligent recommendation decision method based on a graph neural network and a cognitive architecture according to claim 1, characterized in that, The intelligent recommendation decision method based on the graph neural network and the cognitive architecture further includes the following steps: Obtain the decision return of the environmental feedback after executing the decision action in the current environmental state; Form new sample data according to the situation cognitive structure of the current environmental state, the decision action and the decision return; Periodically batch train the intelligent recommendation model according to multiple new sample data to optimize the intelligent recommendation model.

6. An intelligent recommendation decision system based on graph neural network and cognitive architecture, characterized in that, It includes: The first module is used for obtaining the situation cognitive structure by performing situation cognitive knowledge reasoning on the detection data of the agent in the current environmental state; The second module is used for performing task planning knowledge reasoning on the situation cognitive structure to obtain all matched task planning rules, and generating an operator structure according to each matched task planning rule; the operator structure represents the action of the agent; The third module is used for performing matching degree reasoning on the situation cognitive structure and the operator structure by the intelligent recommendation model based on the graph neural network to obtain the dynamic preference value of the operator structure; The fourth module is used for selecting the optimal operator structure according to the dynamic preference value of each operator structure to determine the decision action; The third module is specifically used for performing the following steps: The intelligent recommendation model is trained by the following steps: Collect the training data set when the unmanned system is working; each sample data in the training data set includes the situation cognitive structure of the agent at the decision time, the operator structure and the decision return of executing the operator structure; ​ ​ performing graph structure construction on the training data set to obtain a situation-operator bipartite graph; a node set of the situation-operator bipartite graph includes a situation node set and an operator node set, and an edge set of the situation-operator bipartite graph represents that a situation node i has performed an operator node j in history, and an edge weight represents a corresponding decision reward; inputting the situation-operator bipartite graph into a graph convolution network-based intelligent recommendation model to perform embedding propagation and neighbor aggregation, to obtain an embedding vector of each node; the embedding vector is used to represent a matching degree between a situation and an operator; updating parameters of the intelligent recommendation model according to the embedding vector of each node to obtain a trained intelligent recommendation model.

7. An electronic device, comprising: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to implement the method in any one of claims 1 to 5.

Images