Multi-satellite intelligent collaborative analysis method based on model accelerated reinforcement learning

By combining deep learning and reinforcement learning methods in satellite mission planning, the problems of slow convergence speed and low accuracy in satellite mission planning in the existing technology are solved, and efficient multi-star intelligent collaborative analysis is achieved.

CN119047349AInactive Publication Date: 2024-11-29DAODATIANJI SOFTWARE TECH BEIJING

Patent Information

Application Number
CN202411546224.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-01
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing intelligent optimization algorithms have problems such as slow convergence speed, low accuracy and low efficiency in satellite mission planning, and lack of fast and efficient multi-star intelligent collaborative analysis methods.

Method used

We adopt a reinforcement learning method based on model acceleration, and optimize it by building deep learning models and reinforcement learning algorithms, combining historical experience to reduce the number of interactions between the agent and the environment and improve decision-making speed.

Benefits of technology

It significantly accelerates the decision-making speed of multi-star intelligent collaborative analysis, improves accuracy and efficiency, and can find optimized solutions more quickly.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047349A_ABST
    Figure CN119047349A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-satellite intelligent collaborative analysis method based on model acceleration reinforcement learning, and relates to the technical field of multi-satellite autonomous collaborative detection. According to the invention, a deep learning model is used to obtain a corresponding output value; and carrying out formalized expression on a multi-satellite cooperative task as a Markov decision process, taking an initial decision result of the deep learning model as an initial state, and carrying out reinforcement learning to obtain a final decision result. According to the method, deep learning and reinforcement learning are combined for use, a conventional reinforcement learning algorithm is accelerated by introducing a deep learning model, learning can be continuously carried out in combination with historical experience of the model, and the number of times of interaction between a reinforcement learning agent and the environment is greatly reduced, so that the algorithm decision-making speed is accelerated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of multi-satellite autonomous collaborative detection, and in particular to a multi-satellite intelligent collaborative analysis method based on model accelerated reinforcement learning. Background Art

[0002] In the existing satellite mission planning research, most of the imaging satellite mission planning problems are modeled as optimization problems, and then various intelligent optimization algorithms are used to solve them. The current mainstream intelligent optimization algorithms mainly include ant colony algorithm, simulated annealing algorithm, particle swarm algorithm and genetic algorithm. The ant colony algorithm utilizes the positive feedback characteristics of pheromones and can search for a better solution in a relatively short time, but the algorithm is prone to premature convergence, that is, when the pheromone concentration of a certain path is significantly higher than that of other paths, the algorithm will converge to the path too quickly. Although the simulated annealing algorithm has strong local search capabilities and is not easy to fall into the local optimal solution, it has the disadvantages of slow convergence speed, long execution time, algorithm performance related to the initial value and parameter sensitivity. Although the particle swarm algorithm has the ability to quickly approach the optimal solution. However, since all particles fly in the direction of the optimal solution, all particles tend to be identical (lose diversity), which makes the convergence speed in the later stage significantly slower. At the same time, when the algorithm converges to a certain accuracy, it cannot continue to optimize, and the accuracy that can be achieved is not high. Genetic algorithms are most widely used in multi-satellite collaborative planning. Although they have strong global search capabilities, they also have an obvious disadvantage. The main disadvantage is that the initial solution of the genetic algorithm is usually generated in a random manner, the search space of the solution is large, the search speed is slow, and it often takes a long time to search for the global optimal solution, which affects the efficiency of the calculation and easily misses the observation time window. Therefore, there is currently a lack of multi-satellite intelligent collaborative analysis methods with fast convergence speed, high accuracy and high efficiency. Summary of the invention

[0003] The purpose of the present invention is to provide a multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning to improve the above technical problems.

[0004] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0005] A multi-satellite intelligent collaborative analysis method based on model-accelerated reinforcement learning, comprising:

[0006] S1. Acquire multi-satellite collaborative mission data for maritime target observation;

[0007] S2, build a deep learning model;

[0008] S3, inputting the multi-satellite collaborative mission data into the deep learning model to obtain a corresponding initial decision result;

[0009] S4. Optimizing the initial decision result using a reinforcement learning algorithm to obtain a corresponding final decision result;

[0010] S5. Implement the final decision result through multiple satellites to complete intelligent collaborative analysis of multiple satellites.

[0011] Furthermore, the multi-satellite collaborative mission data includes satellite mission information and window information; the window information includes N time windows; each time window includes window ID, window start time, window end time, target coverage and the longitude and latitude of the upper left corner, lower left corner, upper right corner and lower right corner of its time window; the mission information includes mission type, mission priority, point observation / area observation, weather cloud cover, satellite ID, payload type, resolution, mission identification, coverage polygon area, spatial attributes, and satellite payload angle.

[0012] Furthermore, the deep learning model includes four layers of LSTM neural networks connected in series; each of the LSTM neural networks includes an input layer, multiple hidden layers and an input layer; and the time step of each of the LSTM neural networks is a time window.

[0013] Furthermore, the training process of the deep learning model includes:

[0014] Obtain historical multi-satellite collaborative mission information for maritime target observation;

[0015] Constructing a deep learning model; inputting the historical multi-satellite collaborative mission information into the deep learning model to obtain a corresponding training output value, i.e., an initial training decision result;

[0016] Constructing a cross entropy loss function based on the initial decision result of the training; solving the cross entropy loss function using the Adadelta optimization algorithm to obtain an optimal solution that minimizes the cross entropy loss function;

[0017] Based on the optimal solution, adjusting the parameters of the deep learning model to obtain a deep learning model after initial training;

[0018] Repeat the above training process until the preset number of iterations is met to obtain the trained deep learning model.

[0019] Furthermore, the reinforcement learning algorithm adopts a reinforcement learning method based on a value function;

[0020] Based on the initial decision result, the reinforcement learning algorithm is used to perform reinforcement learning on the multi-satellite collaborative task to obtain the corresponding final decision result, including:

[0021] Step 1: Determine the observation requirements for multi-satellite collaborative missions;

[0022] Step 2: Convert the multi-satellite collaborative task into a Markov decision process; take a single satellite as an intelligent agent, define the environment, state space, transition model and reward function of the intelligent agent, and take the initial decision result as the initial state;

[0023] Step 3: Determine an initial strategy based on the initial state, the conversion model and the observation requirement; calculate a state value function corresponding to the initial state based on the initial strategy and the reward function;

[0024] Step 4: Select the action with the largest state value function as the action of the next state, and calculate the corresponding state-action value function to obtain the corresponding q value;

[0025] Step 5: Implement the action in step 4 through the agent and update the state;

[0026] Step 6: Update the q value according to the updated state;

[0027] Step 7: Repeat steps 3 to 6 until the optimal q value, i.e. the optimal state-action value function, is obtained.

[0028] Step 8: Based on the optimal state-action value function, determine the strategy to obtain the corresponding final decision result.

[0029] Furthermore, the environment includes payload types, mission attributes, and business constraint rules of multiple satellites;

[0030] The business constraint rules are as follows: electronic satellites must be scanned whenever they pass by, optical satellites with large width and low resolution are given priority, cloud cover is not higher than 20%, applicable during the day, and high-resolution satellites are used for tracking and monitoring after the target is found; SAR satellites with large width and low resolution are given priority, and high-resolution is used for tracking and monitoring after the target is found.

[0031] Furthermore, the initial decision result and the final decision result are both time window selection results.

[0032] The reward function is defined; the reward function is a summary of reward and penalty; when the time window is increased or the business constraint rule is met, the reward is obtained; when the business constraint rule is not met, the penalty is obtained.

[0033] The present invention combines deep learning and reinforcement learning. By introducing a deep learning model, conventional reinforcement learning algorithms are accelerated. Combined with the historical experience of the model, continuous learning can be achieved, which greatly reduces the number of interactions between the reinforcement learning agent and the environment, thereby accelerating the algorithm's decision-making speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 is a flow chart of the invention in an embodiment of the present invention;

[0036] Figure 2 is a structural diagram of a deep learning model in an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of a Markov process in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0039] See also Figure 1 This embodiment provides a multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning, which includes:

[0040] S1. Acquire multi-satellite collaborative mission data for maritime target observation;

[0041] Among them, the multi-satellite collaborative mission data includes satellite mission information and window information; the window information includes N time windows; each time window includes window ID, window start time, window end time, target coverage rate and the longitude and latitude of the upper left corner, lower left corner, upper right corner and lower right corner of its time window; the mission information includes mission type, mission priority, point observation / area observation, weather cloud cover, satellite ID, payload type, resolution, mission identification, coverage polygon area, spatial attributes, and satellite payload angle.

[0042] The expression of the window information is:

[0043] ;

[0044] in, Indicates window information. Represents the total number of time windows, Represents the set of positive integers.

[0045] The expression of the time window is:

[0046] ;

[0047] in, Indicates time windows; represents the total number of task features; , , Represent the first task feature, the second task feature, task features; the task features are window ID, window start time, window end time, target coverage rate and the latitude and longitude of the upper left corner, lower left corner, upper right corner and lower right corner of its time window.

[0048] S2, build a deep learning model;

[0049] like Figure 2 As shown, the deep learning model includes four layers of LSTM neural networks connected in series; each of the LSTM neural networks includes an input layer, multiple hidden layers and an input layer; and the time step of each LSTM neural network is a time window. Figure 2 middle, , Represent the initial hidden state and final memory unit of the deep learning model, respectively. , Represent the final hidden state and final memory unit of the deep learning model, respectively. , , , They represent the initial hidden state of the first layer of LSTM neural network, the hidden state after the first time step, the hidden state after the n-1th time step, and the hidden state after the nth time step. , , , They represent the initial memory unit of the first layer of LSTM neural network, the memory unit after the first time step, the memory unit after the n-1th time step, and the memory unit after the nth time step. , , , They represent the initial hidden state of the second layer LSTM neural network, the hidden state after the first time step, the hidden state after the n-1th time step, and the hidden state after the nth time step, respectively. , , , They represent the initial memory unit of the second layer LSTM neural network, the memory unit after the first time step, the memory unit after the n-1th time step, and the memory unit after the nth time step. , , , They represent the initial hidden state of the w-th layer LSTM neural network, the hidden state after the first time step, the hidden state after the n-1th time step, and the hidden state after the nth time step, respectively. , , , They represent the initial memory unit of the w-th layer LSTM neural network, the memory unit after the first time step, the memory unit after the n-1th time step, and the memory unit after the nth time step. The value of w is 4.

[0050] The training process of the deep learning model includes:

[0051] Acquire historical multi-satellite collaborative mission data for marine target observation; the multi-satellite collaborative mission data includes satellite mission information and window information;

[0052] Window Information The corresponding expression of the selection result is:

[0053] ;

[0054] in, Indicates window information The corresponding selection results are: Represents each time window The corresponding selection results.

[0055] ;

[0056] in, Indicates the decision prediction result corresponding to the state, Represents the decision prediction result corresponding to each time window.

[0057] ;

[0058] in, Represents historical task data.

[0059] Construct a deep learning model; input the multi-satellite collaborative mission data into the deep learning model, merge the satellite mission information and the window information and input them into the deep learning model to obtain the corresponding training output value, that is, the initial training decision.

[0060] Constructing a cross entropy loss function based on the training output value; solving the cross entropy loss function using the Adadelta optimization algorithm to obtain an optimal solution that minimizes the cross entropy loss function;

[0061] The formula of the cross entropy loss function is:

[0062] ;

[0063] in, represents the summation function, represents the total number of categories, g Indicates taking a given sample Case Label Equal to category The logarithmic function of probability, represents the conversion model, is a batch of task feature data, Indicates the number of windows, , Represents the time window , The corresponding selection result.

[0064] Based on the optimal solution, adjusting the parameters of the deep learning model to obtain a deep learning model after initial training;

[0065] Repeat the above training process until the preset number of iterations is met to obtain the trained deep learning model.

[0066] Based on the data, the environment is modeled and fitted to a deep learning model. The deep learning model is the crystallization of planning experience, and then the intelligent agent generates results based on the deep learning model. In this process, the intelligent agent makes full use of the data, because once the deep learning model is fitted, the intelligent agent can directly generate decision results based on the deep learning model. Through clever feature design, the deep learning model can be used to model the multi-satellite collaborative scene environment. Through the generalized approximation method based on the value function, the "dimensionality curse problem" of reinforcement learning can be effectively solved.

[0067] S3, inputting the multi-satellite collaborative mission data into the deep learning model to obtain a corresponding initial decision result;

[0068] If reinforcement learning is used directly for multi-star collaborative tasks, the learning efficiency will be too low, and the agent will frequently interact with the environment, and each interaction will bring a lot of overhead in terms of events and space. However, using a deep learning model to obtain a learning result and use the learning result as the initial state of the agent, and then performing reinforcement learning, the number of interactions between the agent and the environment in reinforcement learning can be greatly reduced, thereby reducing the calculation time.

[0069] S4. Optimizing the initial decision result using a reinforcement learning algorithm to obtain a corresponding final decision result;

[0070] The satellite collaborative planning problem is described as a Markov decision process, and the mapping relationship from the system state in the planning to the decision result is regarded as the decision strategy of this Markov decision process, thereby converting the satellite collaborative planning problem into a sequential decision-making problem. In the decision-making process, according to the current state, it is decided whether to accept or reject the current time window observation demand. When making a decision, it is necessary not only to consider the immediate impact of the decision under the current state, but also to consider the impact of the decision on the future. Specifically in the satellite collaborative planning problem, if the algorithm accepts the current time window observation demand, it consumes observation resources (such as power, working time, storage, etc.), which may make it impossible to execute the time window task with greater benefits in the future; if the algorithm rejects the observation demand, the benefit of the current window observation demand is lost; therefore, the algorithm needs to comprehensively consider the attributes of the current observation demand and the system state of the satellite to make a decision.

[0071] The reinforcement learning algorithm adopts a reinforcement learning method based on a value function;

[0072] The method of optimizing the initial decision result by using a reinforcement learning algorithm to obtain a corresponding final decision result includes:

[0073] Step 1: Determine the observation requirements for multi-satellite collaborative missions;

[0074] Step 2: Convert the multi-satellite collaborative task into a Markov decision process; take a single satellite as an intelligent agent, define the environment, state space, transition model and reward function of the intelligent agent, and take the initial decision result as the initial state;

[0075] The environment includes payload types, mission attributes, and business constraints of multiple satellites;

[0076] The business constraint rules are as follows: electronic satellites must be scanned whenever they pass by, optical satellites with large width and low resolution are given priority, cloud cover is not higher than 20%, applicable during the day, and high-resolution satellites are used for tracking and monitoring after the target is found; SAR satellites with large width and low resolution are given priority, and high-resolution is used for tracking and monitoring after the target is found.

[0077] like Figure 3 As shown, the definition of Markov decision process is: A sequential decision problem in a fully observable stochastic environment with a Markov transition model and additional rewards is called a Markov decision process or MDP, which consists of an environment and a set of states.

[0078] The state space includes multiple states; each state includes a set of actions; the state refers to the state of the agent. In the multi-satellite collaborative task scenario, the state refers to the time window selection of the task issued. For example, selecting the first 5 time windows from all time windows is one state, and selecting the first 10 time windows is another state. In MDP, use To express the state, .

[0079] Actions are actions that the agent can take in each state. The set of actions for each state is not necessarily the same. , there are (total number of time windows - state The number of possible actions, and use the parameter To indicate action.

[0080] The reward function refers to the points obtained when the agent enters a state, which is expressed as . From a certain state, the agent will receive a certain reward when performing an action. For example, the current state is to select 5 time windows. When the action of adding a time window is executed, a certain reward will be obtained. This reward is because the newly added time window observes the task one more time, so the benefit is obtained. Cumulative rewards mean that the rewards from different actions are cumulative. The constraint rules are set through the reward function. When the constraint conditions are not met, penalties will be imposed. The final score (total reward) is the sum of the points obtained along the way from not selecting a time window to completing the entire sequence decision process.

[0081] The reward function is defined; the reward function is a summary of reward and penalty; when the time window is increased or the business constraint rule is met, the reward is obtained; when the business constraint rule is not met, the penalty is obtained.

[0082] Step 3: Determine an initial strategy based on the initial state, the conversion model and the observation requirement; calculate a state value function corresponding to the initial state based on the initial strategy and the reward function;

[0083] State value function The calculation formula is:

[0084] ;

[0085] in, represents the reward function for the next state, represents the discount factor, Indicates the next state, represents the value of the next state, Represents the expected value operator, indicating the state The expected return from starting and following a strategy.

[0086] Step 4: Select the action with the largest state value function as the action of the next state, and calculate the corresponding state-action value function to obtain the corresponding q value;

[0087] State-Action Value Function The calculation formula is:

[0088] ;

[0089] in, Indicated in strategy The expected value operator under (current decision) represents the policy The mean or central tendency of a random variable under Indicates the action selected for the next state, The state-action value function that represents the next state. The value of is the q value.

[0090] Step 5: Implement the action in step 4 through the agent and update the state;

[0091] Step 6: Update the q value according to the updated state. The corresponding formula is:

[0092] ;

[0093] in, represents the maximum value function, , Respectively represent the next state and the corresponding action, Represents the reward function corresponding to the current state.

[0094] Step 7: Repeat steps 3 to 6 until the optimal q value, i.e. the optimal state-action value function, is obtained.

[0095] Step 8: Based on the optimal state-action value function, determine the strategy, that is, obtain the corresponding final decision result. The purpose of calculating the state-action value function is to construct a learning algorithm to obtain the optimal strategy from the sequence data. Each strategy corresponds to a state-action value function, and the optimal strategy naturally corresponds to the optimal state-action value function.

[0096] The initial decision result and the final decision result are both time window selection results.

[0097] The states and actions are constructed into a q-table to store the q values. In step 2, the q-table needs to be initialized. The initialized q-table is shown in Table 1:

[0098] Table 1

[0099]

[0100] After reinforcement learning, the q-table is shown in Table 2:

[0101] Table 2

[0102]

[0103] S5. Implement the final decision result through multiple satellites to complete intelligent collaborative analysis of multiple satellites.

[0104] The present invention combines deep learning and reinforcement learning. By introducing a deep learning model, conventional reinforcement learning algorithms are accelerated. Combined with the historical experience of the model, continuous learning can be achieved, which greatly reduces the number of interactions between the reinforcement learning agent and the environment, thereby accelerating the algorithm's decision-making speed.

[0105] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning, characterized in that: include: S1. Acquire multi-satellite collaborative mission data for maritime target observation; S2, build a deep learning model; S3, inputting the multi-satellite collaborative mission data into the deep learning model to obtain a corresponding initial decision result; S4. Optimizing the initial decision result using a reinforcement learning algorithm to obtain a corresponding final decision result; S5. Implement the final decision result through multiple satellites to complete intelligent collaborative analysis of multiple satellites.

2. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 1 is characterized in that: The multi-satellite collaborative mission data includes satellite mission information and window information; the window information includes N time windows; each time window includes window ID, window start time, window end time, target coverage rate and the longitude and latitude of the upper left corner, lower left corner, upper right corner and lower right corner of its time window; the mission information includes mission type, mission priority, point observation / area observation, weather cloud cover, satellite ID, payload type, resolution, mission identification, coverage polygon area, spatial attributes, and satellite payload angle.

3. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 2 is characterized in that: The deep learning model includes four layers of LSTM neural networks connected in series; each of the LSTM neural networks includes an input layer, multiple hidden layers and an input layer; the time step of each LSTM neural network is a time window.

4. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 3 is characterized in that: The training process of the deep learning model includes: Obtain historical multi-satellite collaborative mission information for maritime target observation; Constructing a deep learning model; inputting the historical multi-satellite collaborative mission information into the deep learning model to obtain a corresponding training output value, i.e., an initial training decision result; Constructing a cross entropy loss function based on the initial decision result of the training; solving the cross entropy loss function using the Adadelta optimization algorithm to obtain an optimal solution that minimizes the cross entropy loss function; Based on the optimal solution, adjusting the parameters of the deep learning model to obtain a deep learning model after initial training; Repeat the above training process until the preset number of iterations is met to obtain the trained deep learning model.

5. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 4 is characterized in that: The reinforcement learning algorithm adopts a reinforcement learning method based on a value function; The method of optimizing the initial decision result by using a reinforcement learning algorithm to obtain a corresponding final decision result includes: Step 1: Determine the observation requirements for multi-satellite collaborative missions; Step 2: Convert the multi-satellite collaborative task into a Markov decision process; take a single satellite as an intelligent agent, define the environment, state space, transition model and reward function of the intelligent agent, and take the initial decision result as the initial state; Step 3: Determine an initial strategy based on the initial state, the conversion model and the observation requirement; calculate a state value function corresponding to the initial state based on the initial strategy and the reward function; Step 4: Select the action with the largest state value function as the action of the next state, and calculate the corresponding state-action value function to obtain the corresponding q value; Step 5: Implement the action in step 4 through the agent and update the state; Step 6: Update the q value according to the updated state; Step 7: Repeat steps 3 to 6 until the optimal q value, i.e. the optimal state-action value function, is obtained. Step 8: Based on the optimal state-action value function, determine the strategy to obtain the corresponding final decision result.

6. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 5 is characterized in that: The environment includes payload types, mission attributes, and business constraints of multiple satellites; The business constraint rules are as follows: electronic satellites must be scanned whenever they pass by, optical satellites with large width and low resolution are given priority, cloud cover is not higher than 20%, applicable during the day, and high-resolution satellites are used for tracking and monitoring after the target is found; SAR satellites with large width and low resolution are given priority, and high-resolution is used for tracking and monitoring after the target is found.

7. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 5 is characterized in that: The initial decision result and the final decision result are both time window selection results.

8. The multi-star intelligent collaborative analysis method based on model-accelerated reinforcement learning according to claim 6 is characterized in that: The reward function is defined; the reward function is a summary of reward and penalty; when the time window is increased or the business constraint rule is met, the reward is obtained; when the business constraint rule is not met, the penalty is obtained.

Citation Information

Patent Citations

  • Multi-satellite collaborative planning method based on maritime moving target

    CN113408835A

  • Distributed decision-making method based on multi-agent deep reinforcement learning

    CN116050740A

  • Satellite space target collaborative observation distributed planning method based on multi-agent reinforcement learning

    CN116187160A

  • Flow prediction satellite path selection method and system based on reinforcement learning

    CN116781139A

  • Positioning method and system for autonomous driving through long short-term memory (LSTM)-based deep reinforcement learning (DRL)

    US20240174257A1

Cited By

  • Multi-satellite planning method and device based on deep learning neural network, and medium

    CN119721661A

  • A multi-satellite planning method, device and medium based on a deep learning neural network

    CN119721661B

  • Navigation satellite power enhancement satellite selection method and system based on deep reinforcement learning

    CN119959982A

  • Space target collaborative credible sensing method based on digital encryption

    CN119966690A