Marine multi-platform collaborative task optimization method and system based on deep reinforcement learning

Abstract

The invention discloses an offshore multi-platform cooperative task optimization method and system based on deep reinforcement learning, and relates to the resource optimization and artificial intelligence technology, and the method comprises the steps: generating a dynamic representation of a task scene, and extracting a task target and related features; processing by utilizing a Transform network based on the extracted task targets and related features to obtain priority distribution of the tasks; the method comprises the following steps: modeling a multi-platform cooperation relationship into a graph structure in advance, and calculating cooperation efficiency among platforms by using a graph neural network; according to the priority distribution of the tasks, optimizing a task allocation strategy by using deep reinforcement learning so as to achieve global optimum under multi-target constraint; and generating a multi-platform task allocation scheme based on a modeling result of the graph neural network and a task allocation optimization result of deep reinforcement learning. According to the method, the limitation of the prior art can be overcome, and more efficient and more intelligent task allocation and optimization are realized.

Description

Technical Field

[0001] The present application relates to the fields of resource optimization and artificial intelligence technology, and in particular to a method and system for optimizing marine multi-platform collaborative tasks based on deep reinforcement learning. Background Art

[0002] In modern maritime missions, command and control systems need to coordinate multiple platforms (such as ships, drones, and autonomous vessels) to complete complex mission combinations such as reconnaissance, search and rescue, and strike. This task allocation requires rapid and dynamic decision-making under limited resources to cope with complex and changing battlefield environments. However, existing collaborative task allocation technologies have the following shortcomings:

[0003] Poor dynamic adaptability to complex environments: Traditional task allocation is mostly based on pre-set rules or static optimization algorithms, and lacks the ability to respond to dynamic environments (such as target movement and weather changes) in real time.

[0004] Imperfect modeling of multi-platform collaboration: The communication and collaboration relationships between multiple platforms are complex, and existing systems lack effective modeling mechanisms to fully utilize these relationships.

[0005] Task priority assessment is not intelligent enough: The importance of tasks is usually manually set and lacks the intelligent ability to dynamically balance battlefield needs and resource constraints.

[0006] Single optimization goal: Traditional algorithms usually only optimize a single goal (such as time or resources) and find it difficult to balance the requirements of multiple tasks at the same time.

[0007] The task allocation steps of existing rule-based and heuristic algorithm-based task allocation systems usually include:

[0008] Situational awareness and modeling: Collect environmental data from radars, drones, and other sensors to form a situational picture.

[0009] Task decomposition: Divide the task into multiple subtasks according to the task objectives.

[0010] Platform selection and allocation: Use rules or heuristic algorithms (such as genetic algorithms and ant colony algorithms) to assign tasks to each platform.

[0011] Plan execution and adjustment: Make adjustments based on task completion and platform status.

[0012] System Structure

[0013] Input module: data collection, including environmental data and task requirements.

[0014] Task allocation module: The rule base is combined with the optimization algorithm to generate the initial task allocation plan.

[0015] Feedback and adjustment module: manual adjustment function based on real-time data.

[0016] Output module: generates task allocation results and instructs the execution platform.

[0017] There are the following disadvantages: Technical characteristics and limitations

[0018] Features: It realizes the preliminary automation of task allocation and has a certain degree of adaptability.

[0019] Lack of global optimization capabilities: Optimization algorithms are usually limited to local optimality and cannot balance multiple objectives.

[0020] Slow response speed: Lack of real-time response capabilities to dynamically changing environments.

[0021] Simple modeling of collaborative relationships: Collaboration between platforms is usually based on static rules and is difficult to adapt to complex collaboration needs.

[0022] Task priority assessment relies on manual work: the level of intelligence in the task allocation process is low. Summary of the invention

[0023] The embodiment of the present application constructs a real-time, intelligent collaborative task allocation framework through the integration of deep reinforcement learning, graph neural network and Transformer model, which can overcome the limitations of existing technologies and achieve more efficient and intelligent task allocation and optimization.

[0024] The present application embodiment provides a method for optimizing a multi-platform collaborative task at sea based on deep reinforcement learning, including:

[0025] Obtain input data from multiple sources;

[0026] Using deep learning models, a battlefield situation model is constructed based on the acquired multi-source input data to generate a dynamic representation of the mission scenario and extract mission objectives and related features;

[0027] The Transformer network is used to process the extracted task objectives and related features, and the task priority is evaluated based on the context information to obtain the priority distribution of the task;

[0028] Model the multi-platform collaboration relationship as a graph structure in advance, and use a graph neural network to calculate the collaboration effectiveness between the platforms, where the platforms are nodes and the collaboration relationships are edges in the graph structure;

[0029] According to the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy to achieve the global optimum under multi-objective constraints;

[0030] Based on the modeling results of graph neural networks and the task allocation optimization results of deep reinforcement learning, a multi-platform task allocation solution is generated.

[0031] Optionally, the acquired multi-source input data includes task information, platform status, resource information, and environmental data;

[0032] The battlefield situation model constructed based on the acquired multi-source input data includes: mission scenario modeling and platform status modeling.

[0033] Optionally, the extracted task objectives and related features include:

[0034] Task characteristics, including task type, task time limit, and required resources; Task characteristics, including current status and historical task information; Environmental characteristics: including external environmental influencing factors;

[0035] The Transformer network is used to process the extracted task objectives and related features, and the task priority is evaluated based on the context information to obtain the task priority distribution, including:

[0036] Converting task features, mission characteristics, and environmental characteristics into feature vectors;

[0037] The input feature vector is mapped to a high-dimensional continuous space using an embedding layer and input into a Transformer network. The feature vectors of the task, platform, and environment are encoded in the input embedding layer of the Transformer network, and the time dimension information is added through position encoding.

[0038] Using the multi-head self-attention mechanism, the relationship between tasks and platforms is modeled to identify key tasks and the degree of coupling between key tasks and platforms.

[0039] According to the key tasks and their coupling degree with the platform, priority scores are generated through a nonlinear activation function.

[0040] Optionally, processing using a Transformer network based on the extracted task objectives and related features, and evaluating the task priority based on context information to obtain the task priority distribution also includes:

[0041] Using the self-attention mechanism, the global relationship between tasks, platforms and environments is calculated.

[0042] Optionally, the multi-platform collaboration relationship is modeled as a graph structure in advance, and the collaboration efficiency between the platforms is calculated using a graph neural network, including:

[0043] Node features include the demand features of the task, and edge features include the potential matching relationship between the task and the platform;

[0044] The graph structure is modeled as follows:

[0045] According to the task priority score, each task and platform is represented as a node;

[0046] The edges between nodes are determined based on the key tasks and the coupling degree between the key tasks and the platform as the edge weights;

[0047] Through the message passing mechanism, task nodes and platform nodes receive feature information from neighbor nodes and update their own status to meet the following requirements:

[0048]

[0049] in, represents the feature vector of node v in layer l, ReLU() represents the activation function, W represents the trainable parameter, Aggregate() represents the aggregation of feature information from neighboring nodes to update the comprehensive features of the target node, e uv represents the characteristics of the edge, and N(v) represents the set of neighbor nodes.

[0050] Optionally, modeling the multi-platform collaboration relationship as a graph structure in advance and using a graph neural network to calculate the collaboration efficiency between the platforms also includes:

[0051] The following results are output based on the constructed graph structure:

[0052] A collaborative matrix of matching priorities between each platform and mission;

[0053] The optimized task and platform matching relationship diagram is used as a collaboration diagram.

[0054] Optionally, based on the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy so as to achieve the global optimum under multi-objective constraints, including:

[0055] The collaboration matrix, the collaboration graph and the environment data are defined as a state space to form a state vector; and the task allocation is defined as an action space;

[0056] Set the reward function to satisfy:

[0057] R = w1·task completion rate + w2·resource utilization rate + w3·priority completion rate - w4

[0058] Delay Penalty

[0059] Among them, w1, w2, w3, and w4 are the corresponding weights respectively;

[0060] Use deep reinforcement learning algorithms to optimize strategies;

[0061] Output the optimal task allocation plan, which includes the allocation relationship between tasks and platforms and the execution order.

[0062] Optionally, based on the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy so as to achieve the global optimum under multi-objective constraints, which also includes the following training process:

[0063] Initialize the state space and randomly generate the initial task allocation plan;

[0064] The agent selects actions through the policy network and updates the task allocation;

[0065] According to the execution results, calculate the immediate reward and update the status;

[0066] Using reward and state transition data, optimize the policy network and value network parameters through gradient descent;

[0067] Iterate sampling and learning until the strategy converges or reaches the expected performance.

[0068] An embodiment of the present application also proposes a system for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the aforementioned method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning are implemented.

[0069] The embodiment of the present application constructs a real-time, intelligent collaborative task allocation framework through the integration of deep reinforcement learning, graph neural network and Transformer model, which can overcome the limitations of existing technologies.

[0070] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0072] Figure 1 This is a flowchart of a method for optimizing a multi-platform collaborative task at sea based on deep reinforcement learning according to an embodiment of the present application;

[0073] Figure 2This is a schematic diagram of the task priority evaluation process of the offshore multi-platform collaborative task optimization method based on deep reinforcement learning in an embodiment of the present application;

[0074] Figure 3 It is a schematic diagram of the construction of a collaborative relationship diagram between optimization tasks and platforms in the offshore multi-platform collaborative task optimization method based on deep reinforcement learning in an embodiment of the present application;

[0075] Figure 4 This is an illustration of task allocation using deep reinforcement learning in an offshore multi-platform collaborative task optimization method based on deep reinforcement learning in an embodiment of the present application. DETAILED DESCRIPTION

[0076] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0077] Existing technologies usually rely on preset rules or static optimization algorithms, which make it difficult to handle emergencies in dynamic environments, such as changes in target locations and rapid changes in environmental conditions (such as sea conditions and weather). Collaboration between multiple platforms is usually based on fixed static models or simple logical rules, which makes it difficult to adapt to complex task collaboration requirements (such as dynamic division of labor and resource sharing). Traditional optimization methods usually only focus on a single goal (such as task completion time or resource consumption) and find it difficult to find a balance in multi-goal scenarios (such as task completion rate, response time, and resource utilization). The setting of priorities in task allocation usually relies on manual evaluation, lacks dynamic adjustment capabilities, and cannot scientifically respond to rapid changes in the battlefield environment.

[0078] Based on this, the embodiment of the present application provides a method for optimizing marine multi-platform collaborative tasks based on deep reinforcement learning, such as Figure 1 As shown, the following steps are included:

[0079] In step S101 , multi-source input data is acquired. In some embodiments, the acquired multi-source input data includes task information, platform status, resource information, and environment data.

[0080] In step S102, a deep learning model (such as a convolutional neural network, CNN) is used to build a battlefield situation model based on the acquired multi-source input data to generate a dynamic representation of the mission scenario and extract mission objectives and related features, including target location, platform capabilities, and resource status.

[0081] In step S103, the extracted task objectives and related features are processed using a Transformer network, and the task priorities are evaluated based on the context information to obtain the priority distribution of the tasks.

[0082] In step S104, the multi-platform collaboration relationship is modeled as a graph structure in advance, and the graph neural network GNN is used to calculate the collaboration efficiency between the platforms, where the platforms are nodes and the collaboration relationships are edges in the graph structure.

[0083] In step S105, according to the priority distribution of tasks, deep reinforcement learning (DRL) is used to optimize the task allocation strategy so as to achieve global optimization under multi-objective constraints;

[0084] In step S106, a multi-platform task allocation solution is generated based on the modeling results of the graph neural network and the task allocation optimization results of deep reinforcement learning.

[0085] In some embodiments, the battlefield situation model constructed based on the acquired multi-source input data includes: mission scenario modeling (such as target location, mission type, etc.) and platform status modeling (such as available resources, current mission load, etc.).

[0086] In some embodiments, Figure 2 As shown in Figure 2, the extracted task objectives and related features include:

[0087] Task characteristics, including task type, task time limit, and required resources; task characteristics, including current status and historical task information; environmental characteristics: including external environmental influencing factors.

[0088] The Transformer network is used to process the extracted task objectives and related features, and the task priority is evaluated based on the context information to obtain the task priority distribution, including:

[0089] Converting task features, mission characteristics, and environmental characteristics into feature vectors;

[0090] The input feature vector is mapped to a high-dimensional continuous space using an embedding layer and input into a Transformer network. The feature vectors of the task, platform, and environment are encoded in the input embedding layer of the Transformer network, and the time dimension information is added through positional encoding.

[0091] The multi-head self-attention mechanism is used to model the relationship between tasks and platforms, identify key tasks and the degree of coupling between key tasks and platforms.

[0092] Feedforward Neural Network (FFN): Generates priority scores through nonlinear activation functions based on key tasks and their coupling with the platform.

[0093] In some embodiments, the extracted task objectives and related features are processed using a Transformer network, and the task priority is evaluated based on contextual information to obtain the priority distribution of the task, which also includes: using a self-attention mechanism to calculate the global relationship between the task, platform and environment.

[0094] This application captures the global relationship between tasks, platforms, and environments through the self-attention mechanism. The parallel computing capability of Transformer ensures the efficiency of priority evaluation. It is suitable for dynamic scenarios and can be fine-tuned for different task scenarios to improve the accuracy of evaluation.

[0095] In some embodiments, Figure 3 As shown in the figure, the multi-platform collaboration relationship is modeled as a graph structure in advance, and the graph neural network is used to calculate the collaboration efficiency between each platform, including:

[0096] Node features include the required features of the task (such as resource type, priority, completion deadline, etc.), and edge features include the potential matching relationship between the task and the platform (such as location, current resource amount, adaptability).

[0097] The graph structure is modeled as follows:

[0098] Each task and platform is represented as a node according to the task priority score.

[0099] The edges between nodes are determined based on the key tasks and the degree of coupling between the key tasks and the platform as the weight of the edge.

[0100] Through the message passing mechanism, task nodes and platform nodes receive feature information from neighbor nodes and update their own status to meet the following requirements:

[0101]

[0102] in, represents the feature vector of node v in layer l, ReLU() represents the activation function, W represents the trainable parameter, Aggregate() represents the aggregation of feature information from neighboring nodes to update the comprehensive features of the target node, e uv represents the characteristics of the edge, and N(v) represents the set of neighbor nodes.

[0103] In some embodiments, modeling the multi-platform collaboration relationship as a graph structure in advance, and using a graph neural network to calculate the collaboration efficiency between the platforms also includes outputting the following results based on the constructed graph structure:

[0104] A collaborative matrix of matching priorities between each platform and mission;

[0105] The optimized task and platform matching relationship diagram is used as a collaboration diagram.

[0106] Through dynamic updates of nodes and edges, it supports real-time changes in task and platform status. The aggregation mechanism of GNN can capture the complex relationship between tasks and platforms and optimize collaboration efficiency from a global perspective. The collaboration graph can intuitively display the matching relationship between tasks and platforms and support transparent system decision-making.

[0107] The present application further utilizes deep reinforcement learning (DRL) technology to dynamically optimize the task allocation scheme to ensure that the system performance is maximized under multi-task and multi-platform constraints. In some embodiments, Figure 4 As shown in the figure, according to the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy so as to achieve the global optimum under multi-objective constraints, including:

[0108] The collaboration matrix, collaboration graph, and environment data are defined as a state space to form a state vector; and the task allocation is defined as an action space. Specifically, the collaboration effectiveness matrix: as the initial data of the state space, reflects the matching degree between the task and the platform. The collaboration graph: provides constraints on the action space, such as only allowing task-platform allocation combinations with high weight edges.

[0109] Dynamic environment data: real-time task status (priority, remaining workload), platform status (remaining resources, current load), and environmental changes (such as sudden task changes, weather changes, etc.).

[0110] State Space Definition: Combine the output of collaborative modeling with task, platform, and environment data to form a complete state description. The state vector includes: task priority and remaining workload, platform remaining resources and availability, and current task-platform matching relationship (derived from the collaborative effectiveness matrix).

[0111] Action Space Definition: Actions are defined as task assignments, such as "Task A is assigned to platform 1". Collaboration graphs are used to restrict possible combinations of action spaces, for example, only highly matched task-platform pairs are allowed as candidate actions. The priority weight of actions is based on the score of the collaboration effectiveness matrix.

[0112] Set the reward function to satisfy:

[0113] R = w1·task completion rate + w2·resource utilization rate + w3·priority completion rate - w4

[0114] Delay Penalty

[0115] Among them, w1, w2, w3, and w4 are the corresponding weights. In the specific example, 1. Task completion rate: The more tasks you complete, the higher the reward. Resource utilization: The efficiency of platform resource use (avoiding resource waste or overload). Priority matching: The completion of high-priority tasks will receive higher rewards. Time efficiency: The faster the task is completed, the higher the reward.

[0116] Use deep reinforcement learning algorithms (such as Proximal Policy Optimization, PPO, or Deep Q-Network, DQN) to optimize policies.

[0117] Policy Network: Constructs a policy function π(a|s) based on a neural network and outputs the probability of taking action a in the current state s. Value Network: Estimates the state value V(s) or action value Q(s,a).

[0118] Output the optimal task allocation plan, which includes the allocation relationship between tasks and platforms and the execution order. In a specific example, the output includes: The optimal task allocation plan: includes the allocation relationship between tasks and platforms and the execution order. System performance indicators: such as task completion rate, resource utilization, allocation efficiency, etc.

[0119] In some embodiments, based on the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy so as to achieve global optimality under multi-objective constraints, which also includes the following training process:

[0120] Initialize the state space and randomly generate the initial task allocation plan;

[0121] The agent selects actions through the policy network and updates the task allocation;

[0122] According to the execution results, calculate the immediate reward and update the status;

[0123] Using reward and state transition data, optimize the policy network and value network parameters through gradient descent;

[0124] Iterate sampling and learning until the strategy converges or reaches the expected performance.

[0125] The embodiment of the present application is based on a reinforcement learning algorithm that can dynamically adapt to changes in tasks and environment without the need for preset rules. Through long-term reward optimization, an allocation strategy close to the global optimal is obtained. Combined with parallel computing, it can adapt to complex task scenarios and quickly generate allocation plans.

[0126] This application uses deep reinforcement learning (DRL) and the Transformer model to achieve real-time optimization of task allocation strategies and dynamic intelligent evaluation of task priorities, overcoming the problems of slow response and unscientific priority evaluation in dynamic environments in existing technologies. Through real-time collection and processing of multi-source data, the system can quickly respond to changes in the battlefield environment, greatly improving the efficiency and flexibility of task allocation.

[0127] The graph neural network (GNN) is used to model the multi-platform collaboration relationship, fully tap the collaboration potential between platforms, optimize resource allocation, and solve the technical problems of low multi-platform collaboration efficiency and insufficient resource utilization in existing technologies. At the same time, the system's multi-objective optimization capability achieves a balance between task completion rate, resource consumption, and response time, and improves the global optimization effect of the allocation strategy.

[0128] An embodiment of the present application also proposes a system for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the aforementioned method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning are implemented.

[0129] It should be noted that in the various embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0130] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0131] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0132] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A method for optimizing maritime multi-platform collaborative tasks based on deep reinforcement learning, characterized in that: include: Obtain input data from multiple sources; Using deep learning models, a battlefield situation model is constructed based on the acquired multi-source input data to generate a dynamic representation of the mission scenario and extract mission objectives and related features; The Transformer network is used to process the extracted task objectives and related features, and the task priority is evaluated based on the context information to obtain the priority distribution of the task; Model the multi-platform collaboration relationship as a graph structure in advance, and use a graph neural network to calculate the collaboration effectiveness between the platforms, where the platforms are nodes and the collaboration relationships are edges in the graph structure; According to the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy to achieve the global optimum under multi-objective constraints; Based on the modeling results of graph neural networks and the task allocation optimization results of deep reinforcement learning, a multi-platform task allocation solution is generated.

2. The method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning according to claim 1, characterized in that: The multi-source input data obtained include task information, platform status, resource information, and environmental data; The battlefield situation model constructed based on the acquired multi-source input data includes: mission scenario modeling and platform status modeling.

3. The method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning according to claim 1, characterized in that: The extracted task objectives and related features include: Task characteristics, including task type, task time limit, and required resources; Task characteristics, including current status and historical task information; Environmental characteristics: including external environmental influencing factors; The Transformer network is used to process the extracted task objectives and related features, and the task priority is evaluated based on the context information to obtain the task priority distribution, including: Converting task features, mission characteristics, and environmental characteristics into feature vectors; The input feature vector is mapped to a high-dimensional continuous space using an embedding layer and input into a Transformer network. The feature vectors of the task, platform, and environment are encoded in the input embedding layer of the Transformer network, and the time dimension information is added through position encoding. Using the multi-head self-attention mechanism, the relationship between tasks and platforms is modeled to identify key tasks and the degree of coupling between key tasks and platforms. According to the key tasks and their coupling degree with the platform, priority scores are generated through a nonlinear activation function.

4. The method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning according to claim 3, characterized in that: The extracted task objectives and related features are processed using the Transformer network, and the task priority is evaluated based on the context information to obtain the task priority distribution, which also includes: Using the self-attention mechanism, the global relationship between tasks, platforms and environments is calculated.

5. The method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning according to claim 3, characterized in that: Model the multi-platform collaboration relationship as a graph structure in advance, and use graph neural networks to calculate the collaboration efficiency between platforms include: Node features include the demand features of the task, and edge features include the potential matching relationship between the task and the platform; The graph structure is modeled as follows: According to the task priority score, each task and platform is represented as a node; The edges between nodes are determined based on the key tasks and the coupling degree between the key tasks and the platform as the edge weights; Through the message passing mechanism, task nodes and platform nodes receive feature information from neighbor nodes and update their own status to meet the following requirements: in, represents the feature vector of node v in layer l, ReLU() represents the activation function, W represents the trainable parameter, Aggregate() represents the aggregation of feature information from neighboring nodes to update the comprehensive features of the target node, e uv represents the characteristics of the edge, and N(v) represents the set of neighbor nodes.

6. The method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning according to claim 5, characterized in that: Modeling the multi-platform collaboration relationship as a graph structure in advance and using graph neural networks to calculate the collaboration efficiency between platforms also includes: The following results are output based on the constructed graph structure: A collaborative matrix of matching priorities between each platform and mission; The optimized task and platform matching relationship diagram is used as a collaboration diagram.

7. The method for optimizing offshore multi-platform collaborative tasks based on deep reinforcement learning according to claim 6, characterized in that: According to the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy so as to achieve the global optimum under multi-objective constraints, including: The collaboration matrix, the collaboration graph and the environment data are defined as a state space to form a state vector; and the task allocation is defined as an action space; Set the reward function to satisfy: R = w1·task completion rate + w2·resource utilization rate + w3·priority completion rate - w4 Delay Penalty Among them, w1, w2, w3, and w4 are the corresponding weights respectively; Use deep reinforcement learning algorithms to optimize strategies; Output the optimal task allocation plan, which includes the allocation relationship between tasks and platforms and the execution order.

8. The method for optimizing marine multi-platform collaborative tasks based on deep reinforcement learning according to claim 7, characterized in that: According to the priority distribution of tasks, deep reinforcement learning is used to optimize the task allocation strategy so as to achieve the global optimum under multi-objective constraints. The training process also includes the following: Initialize the state space and randomly generate the initial task allocation plan; The agent selects actions through the policy network and updates the task allocation; According to the execution results, calculate the immediate reward and update the status; Using reward and state transition data, optimize the policy network and value network parameters through gradient descent; Iterate sampling and learning until the strategy converges or reaches the expected performance.

9. A marine multi-platform collaborative task optimization system based on deep reinforcement learning, characterized in that: It comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the offshore multi-platform collaborative task optimization method based on deep reinforcement learning as described in any one of claims 1 to 8 are implemented.

Images