A planting strategy generation method and device, electronic equipment and storage medium

Abstract

The application provides a planting strategy generation method and device and electronic equipment. The method comprises: obtaining a set of action samples in a target object planting environment; based on the set of action samples, decoupling reconstruction processing is performed on different action samples in the set of action samples by an action encoder network in a planting strategy generation model to form an action embedding vector; a change trend of different actions in the target object planting environment is determined by a graph neural network in the planting strategy generation model; and a planting strategy matched with the target object planting environment is determined according to the change trend of different actions in the target object planting environment. Thus, the change trend of different actions in a multidimensional space can be effectively predicted, the planting strategy can be timely and accurately adjusted, the dynamic action space has robustness and generalization, and the calculation cost is reduced.

Description

Technical Field

[0001] This invention relates to information processing technology, and more particularly to a method, apparatus, and electronic device for generating planting strategies. Background Technology

[0002] Artificial Intelligence (AI) is a comprehensive technology within computer science that studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities. AI technology is a multidisciplinary field, encompassing a wide range of areas, including natural language processing and machine learning / deep learning. It is believed that with technological advancements, AI will be applied in more fields and play an increasingly important role.

[0003] When predicting plant growth, physical simulators based on physical formulas are typically used. These simulators predict daily plant growth based on factors such as light intensity, temperature, humidity, and carbon dioxide concentration at any given moment in the greenhouse, combined with the plant's metabolic activity at different growth stages. While these simulators derive daily plant growth under various greenhouse conditions using physical formulas, the numerous factors influencing plant conditions in a real greenhouse, along with their interactions, mean that existing physical simulators cannot accurately simulate the impact of different weather conditions. Therefore, the predictions of plant growth by these simulators often differ significantly from those in a real greenhouse. Furthermore, physical simulators based on physical formulas are slow, typically requiring 3-5 seconds to simulate a single day's plant growth. This results in significant time costs when simulating long growing cycles and finding optimal planting strategies. Artificial intelligence (AI) can increase the accuracy and efficiency of planting strategy generation, enabling the handling of complex planting strategies. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a planting strategy generation method, apparatus, electronic device, and storage medium. The technical solution of the embodiments of the present invention is implemented as follows: This invention provides a method for generating planting strategies, including: Obtain a set of action samples in the target object's planting environment; Based on the action sample set, the action encoder network in the planting strategy generation model is used to decouple and reconstruct different action samples in the action sample set to form an action embedding vector. Based on the action sample set, the state encoder network in the planting strategy generation model is used to perform state encoding processing on different action samples in the action sample set to form an observation embedding vector. Using the action embedding vector and the observation embedding vector, the graph neural network in the planting strategy generation model is used to determine the changing trends of different actions in the planting environment of the target object; Based on the changing trends of different actions in the target object's planting environment, a planting strategy that matches the target object's planting environment is determined.

[0005] This invention also provides a planting strategy generation device, comprising: The information transmission module is used to acquire a set of action samples in the target object's planting environment; The information processing module is used to decouple and reconstruct different action samples in the action sample set based on the action sample set by generating an action encoder network in the model through a planting strategy, thereby forming an action embedding vector. The information processing module is used to perform state encoding processing on different action samples in the action sample set based on the action sample set, and form an observation embedding vector by using the state encoder network in the planting strategy generation model. The information processing module is used to determine the changing trends of different actions in the planting environment of the target object by using the action embedding vector and the observation embedding vector through the graph neural network in the planting strategy generation model. The information processing module is used to determine a planting strategy that matches the planting environment of the target object based on the changing trends of different actions in the target object's planting environment.

[0006] In the above scheme, The information processing module is used to determine a sample collection strategy that matches the planting environment of the target object; The information processing module is used to determine, according to the sample collection strategy, that the sample collection method for the target dimension is random collection, and to configure random values ​​for the sample collection of non-target dimensions. The information processing module is used to concatenate the sample collection results of the target dimension and the random values ​​configured for the non-target dimension to form action samples in the target object's planting environment. The information processing module is used to merge the action samples corresponding to all dimensions in the target object's planting environment to form a set of action samples in the target object's planting environment.

[0007] In the above scheme, The information processing module is used to generate the action encoder network in the model through the planting strategy, and map each action sample in the action sample set to the common latent space to form a latent embedding vector. The information processing module is used to decouple and reconstruct different action samples in the action sample set according to the mapping relationship between the implicit embedding vector and each action sample, so as to form an action embedding vector.

[0008] In the above scheme, The information processing module is used to predict the action sample set through the predictor subnetwork configured in the action encoder network of the planting strategy generation model, so as to adjust the action encoder network according to the predicted action sample set.

[0009] In the above scheme, The information processing module is used to determine the network parameters of the state encoder network based on the planting environment of the target object. The information processing module is used to determine the network structure of the state encoder network based on the network parameters of the state encoder network. The information processing module is used to perform state encoding processing on action samples in different dimensions through the network structure of the determined state encoder network, and form observation embedding vectors.

[0010] In the above scheme, The information processing module is used to exchange the action embedding vector and the observation embedding vector through the action state exchange subnetwork of the graph neural network in the planting strategy generation model. The information processing module is used to generate a graph attention subnetwork of the graph neural network in the model through the planting strategy, process the exchanged action embedding vector and the observation embedding vector, and adjust the weights between the different action samples. The information processing module is used to determine the changing trends of different actions in the target object's planting environment based on the weights between the different action samples.

[0011] In the above scheme, The information processing module is used to determine different associated objects that match the target object based on the knowledge graph of the target object; The information processing module is used to determine the nodes of the graph neural network in the planting strategy generation model based on different associated objects that match the target object. The information processing module is used to determine different edges of the graph neural network in the planting strategy generation model based on the trend latent variables that match the target object. The information processing module is used to determine the graph neural network in the planting strategy generation model based on the nodes of the graph neural network and the different edges of the graph neural network.

[0012] In the above scheme, The information processing module is used to determine the feature information corresponding to any node of the graph neural network; The information processing module is used to broadcast the feature information of the current node to other nodes connected to the current node; The information processing module is used to adjust the corresponding edge lines based on the feature information of other nodes connected to the current node and the current node, so as to determine the graph neural network in the planting strategy generation model.

[0013] In the above scheme, The information processing module is used to determine a training sample set and a validation sample set that match the planting strategy generation model based on the historical parameters of the target object, wherein the training sample set includes at least one set of training samples. The information processing module is used to extract a set of training samples that match the training samples through the planting strategy generation model; The information processing module is used to train the planting strategy generation model based on a set of training samples that match the training samples, so as to determine the model parameters that are suitable for the planting strategy generation model.

[0014] In the above scheme, The information processing module is used to determine a multi-task loss function that matches the planting strategy generation model; The information processing module is used to adjust the network parameters of the planting strategy generation model based on the feature vectors in the sample set, the action feature vectors of the target object, and the multi-task loss function, until the loss functions of different dimensions corresponding to the planting strategy generation model reach the corresponding convergence conditions; so as to make the parameters of the planting strategy generation model adapt to the planting environment.

[0015] In the above scheme, The information processing module is used to adjust the actions in the target object's planting environment according to the determined planting strategy that matches the target object's planting environment. The actions in the target object's planting environment include at least one of the following: light intensity, outdoor temperature, outdoor humidity, outdoor wind speed, and atmospheric temperature.

[0016] This invention also provides an electronic device, the electronic device comprising: Memory, used to store executable instructions; The processor, when running the executable instructions stored in the memory, implements the aforementioned planting strategy generation method.

[0017] This invention also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the aforementioned planting strategy generation method.

[0018] The embodiments of the present invention have the following beneficial effects: This invention obtains a set of action samples in the planting environment of a target object; based on the action sample set, it decouples and reconstructs different action samples in the action sample set using an action encoder network in a planting strategy generation model to form action embedding vectors; based on the action sample set, it performs state encoding processing on different action samples in the action sample set using a state encoder network in the planting strategy generation model to form observation embedding vectors; using the action embedding vectors and the observation embedding vectors, it determines the changing trends of different actions in the planting environment of the target object through a graph neural network in the planting strategy generation model; and based on the changing trends of different actions in the planting environment of the target object, it determines a planting strategy that matches the planting environment of the target object. The planting strategy generation model provided by this application can not only effectively predict the changing trends of different actions in a multi-dimensional space and adjust the planting strategy in a timely and accurate manner, but also has robustness and generalization to dynamic action spaces, reducing computational costs, improving the efficiency of planting strategy generation, and enabling the processing of planting strategies with complex dimensions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating a usage scenario of the planting strategy generation method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the composition of the planting strategy generation device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an optional first process for the planting strategy generation method provided in an embodiment of the present invention; Figure 4 This is an optional second process diagram of the planting strategy generation method provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the data structure of the planting strategy generation model in an embodiment of the present invention; Figure 6 This is a schematic diagram of data collection for the planting strategy generation model in an embodiment of the present invention; Figure 7 This is a schematic diagram of the graph attention network data structure of the planting strategy generation model in an embodiment of the present invention; Figure 8 This is a schematic diagram of an optional processing step in the tomato planting strategy generation method of this invention. Figure 9 This is a schematic diagram illustrating the selection of different tomato planting strategies in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the effects of different tomato planting strategies in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0022] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be interpreted as follows.

[0023] 1) In response to, used to indicate the conditions or states on which the operation performed depends. When the conditions or states on which it depends are met, one or more operations performed may be performed in real time or with a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations are performed.

[0024] 2) Based on, used to indicate the conditions or states on which the operation is performed depends. When the conditions or states on which it depends are met, one or more operations can be performed in real time or with a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations are performed.

[0025] 3) Model training involves multi-class classification learning on the image dataset. This model can be built using deep learning frameworks such as TensorFlow and Torch, employing multiple layers of neural networks like CNNs to form a multi-class classification model. The model input is a three-channel or original-channel matrix generated from images read using tools like OpenCV. The model output is the multi-class probability, ultimately outputting the webpage category through algorithms such as softmax. During training, the model approximates the correct trend using objective functions such as cross-entropy.

[0026] 4) Neural Network (NN): Artificial Neural Network (ANN), also known as neural network or neural network-like network, is a mathematical or computational model in the fields of machine learning and cognitive science that imitates the structure and function of biological neural networks (the central nervous system of animals, especially the brain) and is used to estimate or approximate functions.

[0027] 5) Graph Neural Networks (GNNs): A type of neural network that operates directly on graph structures, primarily processing data with non-Euclidean spatial structures (graph structures). It ignores the input order of nodes; during computation, the representation of a node is influenced by its neighboring nodes, while the graph's connections remain unchanged; the graph structure allows for graph-based reasoning. Typically, a GNN consists of two modules: a propagation module and an output module. The propagation module transmits information and updates the state between nodes in the graph, while the output module provides vector representations of the nodes and edges, defining the objective function according to different tasks. Examples of GNNs include Graph Convolutional Networks (GCNs), Gated Graph Neural Networks (GGNNs), and Graph Attention Networks (GATs) based on attention mechanisms.

[0028] 6) Reinforcement Learning: Used to solve sequential decision-making problems, modeling the problem as a Markov decision process with complete or partial observation. It mainly consists of two components: the environment and the agent. The agent makes decisions based on the state of the environment and its current policy. The environment transitions to a new state based on the agent's decisions and provides a scalar reward signal. The agent continuously optimizes its policy based on the reward from the environment.

[0029] 7) Partially Observed Stochastic Games: A direct extension of partially observed Markov decision processes in multi-agent systems. Similar to partially observed Markov decision processes, it consists of a state space, observation space, action space, reward function, state transition function, and emission function. Each agent's goal is to maximize the expected cumulative reward obtained during the game.

[0030] 8) Collaborative Multi-Agent Reinforcement Learning: A direct extension of reinforcement learning to multi-agent systems, typically modeling the problem as a special case of partially observed stochastic games, or collaborative partially observed stochastic games. In this case, each agent no longer maximizes only its own expected cumulative reward during the game, but rather maximizes the shared expected cumulative team reward.

[0031] 9) Graph Attention Network: The combination of attention mechanism and graph convolutional network, replacing the aggregation operation in graph convolutional network with attention layer instead of simple addition, averaging, etc.

[0032] 10) Hierarchical Variational Autoencoder: A variant of the variational autoencoder. In a hierarchical variational autoencoder, instead of a single latent variable, multiple latent variables are included. These latent variables have a hierarchical relationship, with upper-level latent variables generated based on lower-level latent variables.

[0033] 11) Dynamic Action Space: This corresponds to the static action space in general reinforcement learning problems, meaning the action space does not expand or shrink during algorithm training or as the system evolves. Conversely, the dynamic action space expands or shrinks randomly as the algorithm trains or the system evolves.

[0034] 12) Multi-dimensional action space: Different actions in the action space do not have the same or similar semantics, such as the angles of different joints of a robot in robot control; instead, they can be divided into sub-action spaces with different semantics. For example, in the crop planting scenario, the algorithm can control different controllers of a smart greenhouse, such as temperature, humidity, carbon dioxide and light.

[0035] 13) Non-stationary environment problem: In multi-agent reinforcement learning, when each agent is considered independently, each agent treats the others as part of the environment. As the policies of the other agents are constantly updated, the state transition probabilities of the environment change over time, leading to non-stationarity of the environment.

[0036] 14) Multi-agent credit allocation problem: Similar to the credit allocation problem in single-agent reinforcement learning. In multi-agent reinforcement learning, if all agents share a common reward function, it is impossible to determine the correlation between the reward and the degree of contribution of different agents, thus giving rise to the multi-agent credit allocation problem.

[0037] Figure 1 This is a schematic diagram illustrating a usage scenario of the planting strategy generation method provided in this embodiment of the invention. (See attached diagram.) Figure 1In the application of the planting strategy generation method provided in this application embodiment, the terminal includes terminal 10-1 and terminal 10-2, wherein terminal 10-1 is located on the developer side and is used to control the training and use of the planting strategy generation model, and terminal 10-2 is located on the user side and is used to request the screening of planting strategies; the terminal is connected to server 200 through network 300, which can be a wide area network or a local area network, or a combination of the two, and uses wireless or wired links to realize data transmission.

[0038] Terminal 10-2 is located on the user side and is used to issue a planting strategy generation request to obtain a planting strategy that is compatible with the planting environment of the target object, where the target object can be various types of crops.

[0039] As an example, server 200 is used to deploy the planting strategy generation device to implement the planting strategy generation method provided by the present invention. The latter can deploy a trained planting strategy generation model to screen planting strategies in different environments (e.g., adjusting at least one of the following in an environment for screening different planting strategies: light intensity, temperature, humidity, wind speed, virtual sky temperature, and historical temperature, historical humidity, historical carbon dioxide concentration, and historical light intensity in a greenhouse). Specifically, before using the planting strategy generation model, it needs to be trained. The specific process includes: determining a training sample set and a validation sample set that match the planting strategy generation model based on the historical parameters of the target object, wherein the training sample set includes at least one set of training samples; extracting a training sample set that matches the training samples through the planting strategy generation model; and training the planting strategy generation model based on the training sample set that matches the training samples to determine model parameters that are suitable for the planting strategy generation model. Of course, the planting strategy generation device provided by the present invention can be trained based on planting strategy generation models for the same target object in different planting strategy generation environments. Finally, the planting strategy adapted to the planting environment determined by the planting strategy generation model is presented on the user interface. The obtained planting strategy generation model and the planting strategy adapted to the planting environment can also be called by other applications (such as agricultural simulators or agricultural planting simulation games). Of course, the planting strategy generation model that matches the corresponding planting strategy database can also be migrated to different planting strategy generation processes (such as targeted planting strategy generation process, chemical planting strategy generation process or organic planting strategy generation process).

[0040] Of course, after the planting strategy generation model is trained, recommendations can be made using the planting strategy generation model. Specifically, this includes: obtaining a set of action samples in the target object's planting environment; based on the action sample set, decoupling and reconstructing different action samples in the action sample set using the action encoder network in the planting strategy generation model to form action embedding vectors; based on the action sample set, performing state encoding processing on different action samples in the action sample set using the state encoder network in the planting strategy generation model to form observation embedding vectors; using the action embedding vectors and the observation embedding vectors, determining the changing trends of different actions in the target object's planting environment using the graph neural network in the planting strategy generation model; and determining a planting strategy that matches the target object's planting environment based on the changing trends of different actions in the target object's planting environment.

[0041] The structure of the planting strategy generation device according to an embodiment of the present invention will be described in detail below. The planting strategy generation device can be implemented in various forms, such as a dedicated terminal with planting strategy generation device processing function, or a server equipped with planting strategy generation device processing function, such as the preceding one. Figure 1 Server 200. Figure 2 This is a schematic diagram of the composition of the planting strategy generation device provided in an embodiment of the present invention. It can be understood that... Figure 2 Only an exemplary structure of the planting strategy generation device is shown, not the entire structure; it can be implemented as needed. Figure 2 The structure shown may be part or all of the structure.

[0042] The planting strategy generation apparatus provided in this embodiment of the invention includes: at least one processor 201, a memory 202, a user interface 203, and at least one network interface 204. The various components in the planting strategy generation apparatus are coupled together via a bus system 205. It can be understood that the bus system 205 is used to realize the connection and communication between these components. In addition to a data bus, the bus system 205 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 2 The general labeled all buses as Bus System 205.

[0043] The user interface 203 may include a monitor, keyboard, mouse, trackball, click wheel, buttons, touchpad, or touch screen.

[0044] It is understood that memory 202 can be volatile memory or non-volatile memory, or both. In this embodiment of the invention, memory 202 is capable of storing data to support the operation of a terminal (such as 10-1). Examples of this data include any computer programs used to operate on the terminal (such as 10-1), such as operating systems and applications. The operating system includes various system programs, such as the framework layer, core library layer, driver layer, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications.

[0045] In some embodiments, the planting strategy generation device provided in this invention can be implemented using a combination of hardware and software. For example, the planting strategy generation device can be a processor in the form of a hardware decoding processor, programmed to execute the planting strategy generation method provided in this invention. For instance, the hardware decoding processor can employ one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0046] As an example of the planting strategy generation device provided in this embodiment of the invention, which is implemented by combining software and hardware, the planting strategy generation device provided in this embodiment of the invention can be directly embodied as a combination of software modules executed by processor 201. The software modules can be located in a storage medium, which is located in memory 202. Processor 201 reads the executable instructions included in the software modules in memory 202 and combines them with necessary hardware (e.g., including processor 201 and other components connected to bus system 205) to complete the planting strategy generation method provided in this embodiment of the invention.

[0047] As an example, processor 201 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0048] As an example of the hardware implementation of the planting strategy generation device provided in this embodiment of the invention, the device provided in this embodiment of the invention can be directly executed by a processor 201 in the form of a hardware decoding processor. For example, it can be executed by one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components to implement the planting strategy generation method provided in this embodiment of the invention.

[0049] In this embodiment of the invention, the memory 202 is used to store various types of data to support the operation of the planting strategy generation device. Examples of such data include: any executable instructions for operation on the planting strategy generation device, such as executable instructions, in which a program implementing the planting strategy generation method of this embodiment of the invention may be included.

[0050] In other embodiments, the planting strategy generation device provided in this invention can be implemented in software. Figure 2 A planting strategy generation device stored in memory 202 is shown. This device can be software in the form of programs and plugins, and includes a series of modules. As an example of a program stored in memory 202, it may include a planting strategy generation device, which includes the following software modules: The information transmission module 2081 is used to acquire a set of action samples in the planting environment of the target object.

[0051] The information processing module 2082 is used to decouple and reconstruct different action samples in the action sample set based on the action sample set by generating an action encoder network in the model through a planting strategy, thereby forming an action embedding vector.

[0052] The information processing module 2082 is used to perform state encoding processing on different action samples in the action sample set based on the action sample set, and form an observation embedding vector by using the state encoder network in the planting strategy generation model.

[0053] The information processing module 2082 is used to determine the changing trends of different actions in the planting environment of the target object by using the action embedding vector and the observation embedding vector through the graph neural network in the planting strategy generation model.

[0054] The information processing module 2082 is used to determine a planting strategy that matches the planting environment of the target object based on the changing trends of different actions in the planting environment of the target object.

[0055] In some embodiments, server 200 may be a standalone 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, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal (such as terminal 10-1) may be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0056] In practical applications, the planting strategy generation model provided in this application can be applied to the fields of structural biology and medicine, and can realize the discovery, optimization, and combination of planting strategies.

[0057] according to Figure 2 The electronic device shown, in one aspect of this application, also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform various embodiments and combinations of embodiments provided in the various optional implementations of the above-described dot-planting strategy generation method.

[0058] Continue to combine Figure 2 The illustrated planting strategy generation apparatus illustrates the planting strategy generation method provided in this embodiment of the invention. First, the training process of the planting strategy generation model deployed on the server is introduced, wherein, see [link to documentation]. Figure 3 , Figure 3 This is an optional first process diagram of the planting strategy generation method provided in the embodiments of the present invention. It can be understood that... Figure 3 The steps shown can be performed by various electronic devices that run the planting strategy generation device, such as a dedicated terminal with a planting strategy generation device, a planting strategy database server, or a server cluster of an agricultural provider. The dedicated terminal with the planting strategy generation device can be a preceding... Figure 2The illustrated embodiment is an electronic device with a planting strategy generation apparatus. To overcome the inaccuracies and low efficiency of traditional planting strategy generation methods, the technical solution provided by this invention utilizes artificial intelligence (AI) technology. AI is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0059] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0060] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0061] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0062] The following is about Figure 3 The steps shown are explained in detail.

[0063] Step 301: The planting strategy generation device determines a training sample set and a validation sample set that match the planting strategy generation model based on the historical parameters of the target object.

[0064] The training sample set includes at least one set of training samples. Taking tomato cultivation as an example, the training sample set can include various types of plant status data. The plant status data is the plant status data at the predicted start time, and the greenhouse control parameters and weather data are the greenhouse control parameters and weather data within a predetermined time period from the predicted start time. For example, if the predicted start time can be 0:00 on the predicted day, and the predetermined time period can be 24 hours, then the plant status data is the plant status data at 0:00 on the predicted day, and the greenhouse control parameters and weather data are the greenhouse control parameters and weather data within 24 hours from 0:00 on the predicted day. It should be understood that the predicted start time and predetermined time period are not limited to the specific values ​​disclosed here, but can be selected according to actual needs.

[0065] Step 302: The planting strategy generation device extracts a set of training samples that match the training samples through the planting strategy generation model.

[0066] Step 303: The planting strategy generation device determines a multi-task loss function that matches the planting strategy generation model.

[0067] Step 304: The planting strategy generation device trains the planting strategy generation model based on a set of training samples that match the training samples, so as to determine the model parameters that are suitable for the planting strategy generation model.

[0068] Specifically, based on the feature vectors in the sample set, the action feature vectors of the target object, and the multi-task loss function, the network parameters of the planting strategy generation model are adjusted until the loss functions of different dimensions corresponding to the planting strategy generation model reach the corresponding convergence conditions; so as to make the parameters of the planting strategy generation model adapt to the planting environment.

[0069] Furthermore, considering that in practical applications, the solution of this application can be implemented not only through a fixed planting strategy generation server, but also through a planting strategy generation server group due to the large number of protein molecules and target objects in the planting strategy database, the solution of this application can be implemented through the planting strategy generation information blockchain. When different terminals simultaneously obtain the planting strategy generation model, the planting strategy can be quickly generated for protein molecules and target objects through the trained planting strategy generation model deployed in the blockchain network. At the same time, different servers (nodes) in the server group obtain the activity screening results of the target object binding product in the blockchain network.

[0070] In this invention, embodiments can be implemented using cloud technology. Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, storage, processing, and sharing. It can also be understood as a general term for network technologies, information technologies, integration technologies, management platform technologies, and application technologies based on cloud computing business models. The backend services of network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites; therefore, cloud technology needs cloud computing as its support.

[0071] It's important to note that cloud computing is a computing model that distributes computing tasks across a resource pool comprised of numerous computers, enabling various application systems to access computing power, storage space, and information services as needed. The network providing these resources is called the "cloud." From the user's perspective, resources in the "cloud" are infinitely scalable, readily available, and can be used on demand, expanded at any time, and paid for based on usage. As the foundational providers of cloud computing capabilities, they establish cloud resource pool platforms, often referred to as cloud platforms or Infrastructure as a Service (IaaS). These platforms deploy various types of virtual resources within the resource pool for external customers to choose from. The cloud resource pool primarily includes: computing devices (which can be virtualized machines containing operating systems), storage devices, and network devices.

[0072] Once the trained planting strategy generation model is deployed on the appropriate server, it can generate planting strategies that match the target object's planting environment. (Continue referencing...) Figure 4 , Figure 4 This is a schematic diagram of an optional second process for the planting strategy generation method provided in an embodiment of the present invention. Figure 4 The steps shown can be performed by various electronic devices that run the planting strategy generation device, such as a dedicated terminal with a planting strategy generation device, a planting strategy database server, or a server cluster of planting simulators. The dedicated terminal with the planting strategy generation device can be a preceding... Figure 2 The electronic device with a planting strategy generation device shown in the embodiment specifically includes: Step 401: The planting strategy generation device acquires a set of action samples in the planting environment of the target object.

[0073] In some embodiments of the present invention, obtaining a set of action samples in the target object's planting environment can be achieved in the following ways: A sample collection strategy matching the target object's planting environment is determined. Based on this strategy, the sample collection method for the target dimension is determined to be random, and random values ​​are configured for the sample collection of non-target dimensions. The sample collection results for the target dimension and the random values ​​configured for non-target dimensions are concatenated to form action samples in the target object's planting environment. The action samples corresponding to all dimensions in the target object's planting environment are merged to form a set of action samples in the target object's planting environment. During plant cultivation, there are various action types, such as: temperature (24-dimensional continuous value, range [13,25]), carbon dioxide concentration (24-dimensional continuous value, range [400,700]), light (2-dimensional continuous value, range [0,24]), and irrigation (2-dimensional continuous value, range [0,24]). The reward for each simulation day (continuous value, range [0,30]) is obtained based on the accumulated fresh weight of tomatoes. Different actions can be represented as the following multivariate based on Markov game theory:

[0074] Where n represents the number of agents; X represents the agent space; and S represents the finite state space. Represents the finite action space of agent i; Represents a finite space of joint actions; Indicates the state transition probability; Represents the finite observation space of agent i; Represents a finite joint observation space; Represents the emission probability function; Let represent the cost function of agent i. The POSG can be expanded into a finite or infinite sequence of stages (or time steps), the number of which is called the horizon. This paper considers only the finite horizon problem. Each agent's objective in the game is to minimize the expected cumulative cost. The cooperative part-random observation stochastic game is based on the POSG and satisfies the following constraints:

[0075] in as well as Representing the agent space A pair of intelligent agents; as well as Represents the policy space of two agents as well as The corresponding strategy. Intuitively, the above constraints indicate that the interests of any pair of agents are not conflicting. A common example of cooperative POSG is fully cooperative POSG (also known as decentralized partially observed Markov process, Dec-POMDP), where all agents share the same cost function. The single-agent reinforcement learning problem with a dynamic multi-dimensional action space can be modeled as a fully cooperative POSG problem, where each dimension of the sub-action space represents an independent agent. Without loss of generality, the optimization objective of the fully cooperative POSG problem can be defined as follows:

[0076] in This represents the parameters of the policy functions of all agents. This represents the joint policy function, which can be decomposed into the local policy functions of each agent. Indicates joint strategy The corresponding stationary distribution.

[0077] Step 402: The planting strategy generation device, based on the action sample set, decouples and reconstructs different action samples in the action sample set through the action encoder network in the planting strategy generation model to form action embedding vectors.

[0078] In some embodiments of the present invention, based on the action sample set, different action samples in the action sample set are decoupled and reconstructed by the action encoder network in the planting strategy generation model to form action embedding vectors. This can be achieved in the following ways: The action encoder network in the model is generated through a planting strategy. Each action sample in the action sample set is mapped to a common latent space, forming a latent embedding vector. Based on the mapping relationship between the latent embedding vector and each action sample, different action samples in the action sample set are decoupled and reconstructed to form action embedding vectors. Specifically, refer to... Figure 5 , Figure 5 This is a schematic diagram of the data structure of the planting strategy generation model in an embodiment of the present invention. The SCARL (Structured Collaborative Reinforcement Learning) multi-agent algorithm is used to solve the variable-number heterogeneous MARL problem. Specifically, it may include an action space representation learning stage and robust generalization. In the strategy learning phase, SCARL employs a Centralized Critic (CCR) and Decentralized Actor (DA) framework, introducing a Graph Neural Network (GNN) as the CCR to address the dynamic multi-dimensional action space problem. Taking tomato cultivation as an example, in a smart planting scenario, as temperature rises, enzyme activity increases simultaneously; if light intensity also increases, photosynthesis intensifies, ultimately leading to higher crop yields. Explicitly modeling the dependencies between different sub-action spaces can improve multi-agent exploration efficiency and thus enhance algorithm performance. To address this, SCARL first introduces a Hierarchical Variational Autoencoder (HVAE) to map the multi-dimensional sub-action spaces to a common latent space. An attention mechanism is then introduced into the GNN-based CCR to form a Graph Attention Network (GAT)-based CCR. The attention weights on the edges in the GAT are used to capture the dependencies between the projected sub-action spaces projected into the common latent space. Finally, modeling the MARL problem as a variable-number heterogeneous problem also brings up the multi-agent credit allocation problem. SCARL, based on GAT, introduces an implicit credit allocation structure in the centralized commentator-decentralized actor framework to solve this problem.

[0079] In some embodiments of the present invention, the action sample set can also be predicted by the predictor subnetwork configured in the action encoder network of the planting strategy generation model, so as to adjust the action encoder network according to the predicted action sample set.

[0080] Step 403: The planting strategy generation device performs state encoding processing on different action samples in the action sample set based on the action sample set, and forms an observation embedding vector through the state encoder network in the planting strategy generation model.

[0081] In some embodiments of the present invention, based on the action sample set, different action samples in the action sample set are state-encoded by a state encoder network in the planting strategy generation model to form observation embedding vectors. This can be achieved in the following ways: Based on the target object's planting environment, the network parameters of the state encoder network are determined; based on the network parameters of the state encoder network, the network structure of the state encoder network is determined; using the determined network structure of the state encoder network, state encoding processing is performed on action samples in different dimensions to form observation embedding vectors. Wherein, reference... Figure 6 , Figure 6This diagram illustrates data acquisition for the planting strategy generation model in an embodiment of the present invention. In a smart planting scenario, the temperature dimension sub-action space is responsible for adjusting the greenhouse temperature, while the carbon dioxide controller is responsible for regulating the carbon dioxide concentration in the greenhouse. Without incorporating prior knowledge from agricultural experts, it is difficult for the algorithm to directly model the correlation between the temperature and carbon dioxide dimensions based on samples taken from the environment by each agent. However, the impact of sub-actions of different dimensions on different aspects of the environment is ultimately reflected in crop yield. Therefore, we can make a reasonable assumption: sub-actions of different dimensions affect common latent variables related to crop yield, such as enzyme activity. If we project all sub-actions of different dimensions into this common latent space, we can more easily capture the correlation between sub-actions of different dimensions and naturally apply the algorithm to scenarios with new sub-action spaces.

[0082] To achieve the above objectives, the SCARL algorithm first encodes the entire action space. First, the SCARL algorithm uses a default policy (in this invention, a random policy is used; however, any default policy, such as an expert policy, can be used) to collect the corresponding set of observation samples for each dimension of the action space. In other words, since each dimension of the sub-action space is modeled as an independent agent, this process involves the agent sampling in the environment according to the default policy. Specifically, to more comprehensively encode the entire action space, the agent using the default policy needs to collect samples with sufficient diversity. Therefore, for each sub-action space, the agent using the default policy first fixes the actions of other dimensions as random values, then randomly samples actions in the current sub-action space, and finally concatenates the two action values ​​and inputs them into the environment, thus collecting a batch of observation samples. This process is repeated multiple times until the number of collected observation samples reaches a predefined threshold. After sampling for each sub-action space, the set of observation samples corresponding to each sub-action space is obtained.

[0083] The SCARL algorithm introduces a hierarchical variational autoencoder (HAV) to encode each sub-action space. The first layer of the VAE encodes the set of observation samples corresponding to each sub-action space into a sub-action space embedding representation. This sub-action space embedding representation, along with a specific action belonging to this sub-action space, is used as the conditional input to the second layer of the VAE to reconstruct the observation samples corresponding to that specific action. The reason SCARL uses the sub-action space embedding representation along with the specific action as input is to obtain the latent embedding representation of the specific action belonging to this sub-action space in the common latent space. Only by using the embedding representations of sub-actions of different dimensions in the common latent space, rather than in the atomic action space, can we more easily capture the correlation between sub-actions of different dimensions using only independent samples sampled by different agents. To enhance the representational power of the sub-action space embedding representation, the SCARL algorithm also adds a self-supervised auxiliary task. Specifically, in addition to the original reconstruction decoder of HAVE, SCARL adds a predictor to predict the agent's observations in the near future.

[0084] For each dimension of the multi-dimensional action space, the sub-action space We first start with the fixed sub-action spaces of other dimensions. Random sampling A different fixed random value Then, for each fixed random value... We then start from the current sub-action space Random sampling One action Therefore, we sampled a total of... These are different joint actions, corresponding to the sample set. middle Each different observation. HAVE performs a sample of each observation. Mean pooling is performed on the corresponding observation sample set. Encoding as sub-action space embedding representation .

[0085] We refer to the motion space encoder that performs the above calculations as the motion space representation module. Note that during the training phase, we sample a sub-action space embedding representation from this module for subsequent computations, but after training, we take the mean of the module's outputs as the embedding representation for each sub-action space. Next, for each observation sample... HAVE uses embedded representation As a motion encoder and motion decoder The conditional input. The motion encoder is based on the observation. and embedded representation As a conditional input, the sub-action Projecting onto the common latent space yields the implicit embedding representation of the sub-actions within the latent space. In addition to the action decoder based on the implicit embedding representation... Beyond reconstructing observed samples, the predictor Based on embedding representation As conditional input, based on implicit embedding representation To predict consecutive future observations within the next t time steps In summary, action space representation learning comprises three parts of the loss function: 1) Regarding all observation samples in the observation sample set Reconstruction loss; 2) Motion space representation module and motion encoder and their corresponding prior distributions and KL divergence between them; 3) Regarding all observation samples in the observation sample set Corresponding subsequent observations The predicted loss.

[0086] Therefore, the action space representation learning problem can be formalized as maximizing the following ELBO (lower bound of evidence): SCARL uses a Long Short-Term Memory (LSTM) network to implement the predictor. (The predictor can be implemented using any type of deep neural network), while the action space representation module is implemented using a bidirectional long short-term memory network (Bi-LSTM). (It should be noted that the action space representation module can be implemented using any type of deep neural network.)

[0087] Step 404: The planting strategy generation device uses the action embedding vector and the observation embedding vector to determine the changing trends of different actions in the planting environment of the target object through the graph neural network in the planting strategy generation model.

[0088] In some embodiments of the present invention, the changing trends of different actions in the planting environment of the target object are determined by using the action embedding vector and the observation embedding vector through a graph neural network in the planting strategy generation model. This can be achieved in the following ways: The action state exchange subnetwork of the graph neural network in the planting strategy generation model swaps the action embedding vector and the observation embedding vector. The graph attention subnetwork of the graph neural network in the planting strategy generation model then processes the swapped action embedding vector and the observation embedding vector, adjusting the weights between the different action samples. Based on the weights between the different action samples, the changing trends of different actions in the target object's planting environment are determined. (Referring to...) Figure 7 , Figure 7 This is a schematic diagram of the graph attention network data structure of the planting strategy generation model in an embodiment of the present invention. The centralized commentator comprises four sub-modules: a state encoder... Motion encoder Action-state interaction module p w and graph attention module The state encoder and action encoder are the same as the two encoders used in the action space representation learning phase. At each time step of each episode... State encoder Motion encoder Receive local observations from each agent separately and the actions output by the policy network As input. Simultaneously, the sub-action space embedding representation for each agent. This information is also simultaneously input into the action encoder (consistent with the action space representation learning phase) to obtain the local observation embedding representation for each agent. and action embedding representation :

[0089] Next, the local observation embedding representation for each agent. and action embedding representation They will be input together into the action-state interaction module. In this process, we obtain the agent embedding representation of agent i at time t. :

[0090] Action-State Pair Interaction Module This can be intuitively understood as the impact of different dimensional sub-actions on the environment. Since each dimensional sub-action has been mapped to a common latent space through the action encoder, their impact on the environment also acts on a common latent variable, such as enzyme activity. Therefore, we can reasonably use a graph attention module to model the dependencies between agents. Specifically, we will calculate the attention weights between any two agents i and j:

[0091] The graph attention module consists of parameters. Perform parameterization. This represents the concatenation operation. Then, the agent's embedded representation is updated using the aggregation function of the graph neural network:

[0092] The above graph convolution operation performed by the SCARL algorithm totals [number] times. Next, we record the first The agent embedding representation of the layer output is as follows Finally, we perform average pooling on the embedding representations of all agents in the graph to obtain the joint Q-value of all agents:

[0093] In the process of determining the graph neural network, different associated objects matching the target object can be identified based on the knowledge graph of the target object; nodes of the graph neural network in the planting strategy generation model can be determined based on the different associated objects matching the target object; different edges of the graph neural network in the planting strategy generation model can be determined based on the trend latent variables matching the target object; and the graph neural network in the planting strategy generation model can be determined based on the nodes and different edges of the graph neural network. Further, the feature information corresponding to any node of the graph neural network is determined; the feature information of the current node is broadcast to other nodes connected to the current node; and the corresponding edges are adjusted based on the feature information of the other nodes connected to the current node and the current node to determine the graph neural network in the planting strategy generation model.

[0094] Step 405: The planting strategy generation device determines a planting strategy that matches the planting environment of the target object based on the changing trends of different actions in the target object's planting environment.

[0095] Continue to refer to Figure 8 , Figure 9 , Figure 10 , Figure 8This is a schematic diagram of an optional processing step in the tomato planting strategy generation method according to an embodiment of the present invention. It can be understood that... Figure 8 The steps shown can be performed by various electronic devices that run the planting strategy generation device, such as a dedicated terminal with a planting strategy generation device, a planting strategy database server, or a server cluster of planting simulators. The dedicated terminal with the planting strategy generation device can be a preceding... Figure 2 The illustrated embodiment is an electronic device with a planting strategy generation apparatus. Figure 9 This is a schematic diagram illustrating the selection of different tomato planting strategies in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the effects of different tomato planting strategies in an embodiment of the present invention. Specifically, the following steps are included: Step 801: Obtain a set of action samples in the tomato growing environment; Step 802: Based on the action sample set, the action encoder network in the model is generated by planting strategy to decouple and reconstruct different action samples in the action sample set to form action embedding vectors; Step 803: Based on the action sample set, the state encoder network in the planting strategy generation model is used to perform state encoding processing on different action samples in the action sample set to form observation embedding vectors; Step 804: Using the action embedding vector and the observation embedding vector, determine the changing trends of different actions in the planting environment of the target object through the graph neural network in the planting strategy generation model; Step 805: Based on the changing trends of different actions in the target object's planting environment, determine a planting strategy that matches the target object's planting environment, and adjust the temperature, carbon dioxide concentration, light, and irrigation amount.

[0096] Beneficial technical effects: This invention obtains a set of action samples in the planting environment of a target object; based on the action sample set, it decouples and reconstructs different action samples in the action sample set using an action encoder network in a planting strategy generation model to form action embedding vectors; based on the action sample set, it performs state encoding processing on different action samples in the action sample set using a state encoder network in the planting strategy generation model to form observation embedding vectors; using the action embedding vectors and the observation embedding vectors, it determines the changing trends of different actions in the planting environment of the target object through a graph neural network in the planting strategy generation model; and based on the changing trends of different actions in the planting environment of the target object, it determines a planting strategy that matches the planting environment of the target object. The planting strategy generation model provided by this application can not only effectively predict the changing trends of different actions in a multi-dimensional space and adjust the planting strategy in a timely and accurate manner, but also has robustness and generalization to dynamic action spaces, reducing computational costs, improving the efficiency of planting strategy generation, and enabling the processing of planting strategies with complex dimensions.

[0097] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating a planting strategy, characterized in that, The method includes: Obtain a set of action samples in the planting environment of the target object, wherein the set of action samples includes temperature dimension, carbon dioxide concentration dimension, light dimension and irrigation dimension; wherein, each dimension's sub-action space represents an independent intelligent agent; Based on the action sample set, the action encoder network in the planting strategy generation model is used to decouple and reconstruct different action samples in the action sample set to form action embedding vectors. In each round at each time step, the action output by the policy network of each agent and the sub-action space embedding vector corresponding to each agent are input into the action encoder network to obtain the action embedding vector corresponding to each agent. Based on the action sample set, the state encoder network in the planting strategy generation model performs state encoding processing on different action samples in the action sample set to form observation embedding vectors. In each time step of each round, the local observations of each agent are input into the state encoder network to obtain the observation embedding vectors corresponding to each agent. The action state exchange subnetwork of the graph neural network in the planting strategy generation model is used to exchange the action embedding vector and the observation embedding vector. Specifically, the observation embedding vector and the action embedding vector corresponding to each agent are input into the action state exchange subnetwork to obtain the agent embedding vector corresponding to the agent at the corresponding time step. The graph attention subnetwork of the graph neural network in the planting strategy generation model is used to process the exchanged action embedding vector and observation embedding vector to adjust the weights between different action samples. In this process, the graph attention subnetwork is used to model the dependencies between agents using the agent embedding vectors and to calculate the attention weights between any two agents as the weights between different action samples. Based on the weights among the different action samples, the changing trends of different actions in the target object's planting environment are determined; Based on the changing trends of different actions in the target object's planting environment, a planting strategy that matches the target object's planting environment is determined.

2. The method according to claim 1, characterized in that, The process of obtaining the action sample set in the target object's planting environment includes: Determine a sample collection strategy that matches the planting environment of the target object; Based on the sample collection strategy, the sample collection method for the target dimension is determined to be random collection, and random values ​​are configured for the sample collection of non-target dimensions; The sample collection results of the target dimension and the random values ​​configured for the non-target dimension are concatenated to form the action samples in the planting environment of the target object. The action samples corresponding to all dimensions in the target object's planting environment are merged to form a set of action samples in the target object's planting environment.

3. The method according to claim 1, characterized in that, The process of decoupling and reconstructing different action samples in the action sample set based on the action sample set, through the action encoder network in the planting strategy generation model, to form action embedding vectors includes: The action encoder network in the model generated by the planting strategy maps each action sample in the action sample set to a common latent space to form a latent embedding vector. Based on the mapping relationship between the implicit embedding vector and each action sample, the different action samples in the action sample set are decoupled and reconstructed to form an action embedding vector.

4. The method according to claim 3, characterized in that, The method further includes: The predictor subnetwork configured in the action encoder network of the planting strategy generation model predicts the action sample set, so as to adjust the action encoder network according to the predicted action sample set.

5. The method according to claim 1, characterized in that, The step of generating observation embedding vectors by using the state encoder network in the planting strategy generation model to perform state encoding processing on different action samples in the action sample set based on the action sample set includes: The network parameters of the state encoder network are determined based on the planting environment of the target object. The network structure of the state encoder network is determined based on the network parameters of the state encoder network. Using the determined network structure of the state encoder network, state encoding is performed on action samples in different dimensions to form observation embedding vectors.

6. The method according to claim 1, characterized in that, The method further includes: Based on the knowledge graph of the target object, different associated objects that match the target object are identified; The nodes of the graph neural network in the planting strategy generation model are determined based on the different associated objects that match the target object; Based on the latent trend variables that match the target object, the different edges of the graph neural network in the planting strategy generation model are determined; Based on the nodes of the graph neural network and the different edges of the graph neural network, the graph neural network in the planting strategy generation model is determined.

7. The method according to claim 6, characterized in that, The process of determining the graph neural network in the planting strategy generation model based on the nodes and different edges of the graph neural network includes: Determine the feature information corresponding to any node of the graph neural network; Broadcast the current node's feature information to other nodes connected to the current node; Based on the feature information of other nodes connected to the current node and the current node, the corresponding edge lines are adjusted to determine the graph neural network in the planting strategy generation model.

8. The method according to claim 1, characterized in that, The method further includes: Based on the historical parameters of the target object, a training sample set and a validation sample set that match the planting strategy generation model are determined, wherein the training sample set includes at least one set of training samples; The planting strategy generates a model that extracts a set of training samples that match the training samples. The planting strategy generation model is trained based on a set of training samples that match the training samples, so as to determine the model parameters that are suitable for the planting strategy generation model.

9. The method according to claim 8, characterized in that, The method further includes: Determine a multi-task loss function that matches the planting strategy generation model; Based on the feature vectors in the training sample set, the action feature vectors of the target object, and the multi-task loss function, the network parameters of the planting strategy generation model are adjusted until the loss functions of different dimensions corresponding to the planting strategy generation model reach the corresponding convergence conditions; so as to make the parameters of the planting strategy generation model adapt to the planting environment.

10. The method according to claim 1, characterized in that, The method further includes: Based on the determined planting strategy that matches the planting environment of the target object, the actions in the planting environment of the target object are adjusted, wherein the actions in the planting environment of the target object include at least one of the following: light intensity, outdoor temperature, outdoor humidity, outdoor wind speed, and atmospheric temperature.

11. A planting strategy generation device, characterized in that, The device includes: The information transmission module is used to acquire a set of action samples in the planting environment of the target object, wherein the set of action samples includes temperature dimension, carbon dioxide concentration dimension, light dimension and irrigation dimension; wherein each dimension's sub-action space represents an independent intelligent agent; The information processing module is used to decouple and reconstruct different action samples in the action sample set based on the action sample set by the action encoder network in the planting strategy generation model to form action embedding vectors. In each time step of each round, the action output by the policy network of each agent and the sub-action space embedding vector corresponding to each agent are input into the action encoder network to obtain the action embedding vector corresponding to each agent. The information processing module is further configured to perform state encoding processing on different action samples in the action sample set based on the action sample set, through the state encoder network in the planting strategy generation model, to form an observation embedding vector. In each time step of each round, the local observation of each agent is input into the state encoder network to obtain the observation embedding vector corresponding to each agent. The information processing module is further configured to exchange the action embedding vector and the observation embedding vector through the action state exchange subnetwork of the graph neural network in the planting strategy generation model. The observation embedding vector and the action embedding vector corresponding to each agent are input into the action state exchange subnetwork to obtain the agent embedding vector corresponding to the agent at the corresponding time step. The information processing module is further configured to generate a graph attention sub-network of the graph neural network in the model through the planting strategy, process the exchanged action embedding vector and observation embedding vector, and adjust the weights between the different action samples. The graph attention sub-network is used to model the dependency relationship between the agents using the agent embedding vector, and calculates the attention weight between any two agents as the weight between the different action samples. The information processing module is also used to determine the changing trend of different actions in the target object's planting environment based on the weights between the different action samples. The information processing module is also used to determine a planting strategy that matches the planting environment of the target object based on the changing trends of different actions in the target object's planting environment.

12. The apparatus according to claim 11, characterized in that, The information processing module is also used to determine a sample collection strategy that matches the planting environment of the target object; The information processing module is also used to determine, according to the sample collection strategy, that the sample collection method for the target dimension is random collection, and to configure random values ​​for the sample collection of non-target dimensions; The information processing module is also used to concatenate the sample collection results of the target dimension and the random values ​​configured for the non-target dimension to form action samples in the target object planting environment. The information processing module is also used to merge the action samples corresponding to all dimensions in the target object's planting environment to form a set of action samples in the target object's planting environment.

13. The apparatus according to claim 11, characterized in that, The information processing module is also used to generate an action encoder network in the model through the planting strategy, and map each action sample in the action sample set to the common latent space to form a latent embedding vector. The information processing module is further configured to decouple and reconstruct different action samples in the action sample set according to the mapping relationship between the implicit embedding vector and each action sample, thereby forming an action embedding vector.

14. The apparatus according to claim 13, characterized in that, The information processing module is further configured to predict the action sample set by using the predictor subnetwork configured in the action encoder network of the planting strategy generation model, so as to adjust the action encoder network according to the predicted action sample set.

15. The apparatus according to claim 11, characterized in that, The information processing module is also used to determine the network parameters of the state encoder network based on the planting environment of the target object; The information processing module is also used to determine the network structure of the state encoder network based on the network parameters of the state encoder network. The information processing module is also used to perform state encoding processing on action samples in different dimensions through the network structure of the determined state encoder network to form observation embedding vectors.

16. The apparatus according to claim 11, characterized in that, The information processing module is also used to determine different associated objects that match the target object based on the knowledge graph of the target object; The information processing module is further configured to determine the nodes of the graph neural network in the planting strategy generation model based on different associated objects that match the target object. The information processing module is also used to determine different edges of the graph neural network in the planting strategy generation model based on the trend latent variables that match the target object. The information processing module is also used to determine the graph neural network in the planting strategy generation model based on the nodes of the graph neural network and the different edges of the graph neural network.

17. The apparatus according to claim 16, characterized in that, The information processing module is also used to determine the feature information corresponding to any node of the graph neural network; The information processing module is also used to broadcast the feature information of the current node to other nodes connected to the current node; The information processing module is also used to adjust the corresponding edge lines based on the feature information of other nodes connected to the current node and the current node, so as to determine the graph neural network in the planting strategy generation model.

18. The apparatus according to claim 11, characterized in that, The information processing module is further configured to determine a training sample set and a validation sample set that match the planting strategy generation model based on the historical parameters of the target object, wherein the training sample set includes at least one set of training samples. The information processing module is also used to extract a set of training samples that match the training samples through the planting strategy generation model; The information processing module is further configured to train the planting strategy generation model based on a set of training samples that match the training samples, so as to determine model parameters that are suitable for the planting strategy generation model.

19. The apparatus according to claim 18, characterized in that, The information processing module is also used to determine a multi-task loss function that matches the planting strategy generation model; The information processing module is further configured to adjust the network parameters of the planting strategy generation model based on the feature vectors in the training sample set, the action feature vectors of the target object, and the multi-task loss function, until the loss functions of different dimensions corresponding to the planting strategy generation model reach the corresponding convergence conditions; so as to make the parameters of the planting strategy generation model compatible with the planting environment.

20. The apparatus according to claim 11, characterized in that, The information processing module is further configured to adjust the actions in the target object's planting environment according to the determined planting strategy that matches the target object's planting environment, wherein the actions in the target object's planting environment include at least one of the following: light intensity, outdoor temperature, outdoor humidity, outdoor wind speed, and atmospheric temperature.

21. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the planting strategy generation method according to any one of claims 1 to 10.

22. A computer-readable storage medium storing executable instructions, characterized in that, When the executable instructions are executed by the processor, they implement the planting strategy generation method according to any one of claims 1 to 10.

23. A computer program product, comprising a computer program or computer-executable instructions, characterized in that, When the computer program or computer-executable instructions are executed by a processor, the planting strategy generation method according to any one of claims 1 to 10 is implemented.

Images