A control system for autonomous learning of aircraft meta-model for collaborative tasks

Abstract

The embodiments of the present application relate to the field of intelligent decision-making technology, and in particular to a control system for autonomous learning of an aircraft meta-model for collaborative tasks, the system comprising: a first acquisition unit, used to acquire multiple different tasks of the aircraft, wherein the multiple different tasks are located in a first target state space; a first processing unit, used to project the multiple different tasks into a low-dimensional feature space through an encoder network in a variational autoencoder, generate multiple state primitives, and decode the multiple state primitives through a decoder network in the variational autoencoder; the problem is solved that when an aircraft performs tasks such as planning and decision-making, it is currently achieved through specially programmed and offline trained algorithms, but the specially programmed and offline trained algorithms require the collection of a large number of training samples for intelligent decision-making and a large amount of training time to learn the optimal strategy, resulting in low efficiency and poor generalization of aircraft strategy learning in complex environments.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of intelligent decision-making technology, and in particular to a control system for autonomous learning of an aircraft meta-model for collaborative tasks. Background Art

[0002] The planning and decision-making of aircraft are completed through special programming or offline training. When the external environment changes, there is a possibility of performance degradation and safety impairment, which seriously restricts its ability to perform tasks. Therefore, current aircraft are mainly completed manually through remote control and other methods. The main reason is that they are not adaptable enough to the environment and tasks.

[0003] How to improve the autonomous intelligence of aircraft, enable them to have continuous autonomous learning capabilities like humans, and realize autonomous learning of knowledge and skills is an important means to improve their flexibility, ease of use, adaptability and rapid deployment capabilities. With the rapid development of related disciplines such as artificial intelligence and neuroscience, especially the development of intelligent developmental learning, reinforcement learning, continuous learning, transfer learning and other technologies and their initial applications in robot and aircraft planning, decision-making, etc., the research and application of continuous autonomous learning technology for aircraft has become feasible. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to propose a control system for autonomous learning of an aircraft meta-model for collaborative tasks. The state primitives are obtained by projecting the historical tasks of the aircraft, and the state primitives are decoded to obtain different tasks after decoding. The state primitives are input into the policy learning module to obtain multiple policy networks containing the task parameters and operation trajectories of each historical task. When a new task to be tested is received, there is no need to re-acquire a large number of tasks for learning. It is only necessary to compare the task to be tested with the decoded task to obtain multiple reconstruction errors. The policy network corresponding to the minimum value of the reconstruction error is determined as the policy network of the task to be tested. The aircraft is controlled to make decisions through the policy network of the task to be tested, which improves the efficiency of the aircraft's policy learning in complex environments and has good generalization.

[0005] To achieve the above-mentioned purpose, an embodiment of the present application provides a control system for autonomous learning of an aircraft meta-model for collaborative tasks, the system comprising: a first acquisition unit, used to acquire multiple different tasks of the aircraft, wherein the multiple different tasks are located in a first target state space; a first processing unit, used to project the multiple different tasks into a low-dimensional feature space through an encoder network in a variational autoencoder, generate multiple state primitives, and decode the multiple state primitives through a decoder network in the variational autoencoder to obtain multiple decoded different tasks, wherein the multiple decoded different tasks are located in a second target state space; a second processing unit, used to input multiple state primitives corresponding to the multiple different tasks into a policy network, and obtain multiple policy networks including motion trajectories of each task in different tasks and task parameters of each task; a second acquisition unit, used to acquire a task to be tested of the aircraft; a calculation unit, used to calculate a reconstruction error according to the task to be tested and the multiple decoded different tasks, and obtain multiple reconstruction errors; a determination unit, used to determine the policy network corresponding to the minimum value of the multiple reconstruction errors as the policy network of the task to be tested; and a control unit, used to control the aircraft to make a decision on the task to be tested based on the policy network of the task to be tested.

[0006] To achieve the above-mentioned objectives, an embodiment of the present application also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned control system for autonomous learning of an aircraft meta-model for collaborative tasks.

[0007] To achieve the above objectives, an embodiment of the present application further provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements a control system for autonomous learning of an aircraft meta-model for collaborative tasks as described above.

[0008] An embodiment of the present application proposes a control system for autonomous learning of an aircraft meta-model for collaborative tasks. The system mainly includes three modules. The first module is called a multiple state primitive generation module, which projects multiple different tasks of the aircraft into a low-dimensional feature space, and generates multiple state primitives in this feature space to characterize the original tasks of the aircraft; the second module is called a strategy learning module, which is responsible for learning a separate strategy for each task of the aircraft based on multiple state primitives; the third module is called a task processing module to be tested, which calculates the reconstruction error between the task to be tested and the original task of the aircraft, and determines the policy network corresponding to the minimum reconstruction error as the policy network of the task to be tested; the aircraft is controlled to make decisions on the task to be tested according to the policy network of the task to be tested. When the aircraft performs tasks such as planning and decision-making, it is currently through specially programmed and offline trained algorithms, but specially programmed and offline trained algorithms are often used. The algorithm of online training needs to collect a large number of training samples for intelligent decision-making and a large amount of training time to learn the optimal strategy, which leads to low efficiency and poor generalization of the aircraft in strategy learning in complex environments. By projecting the historical tasks of the aircraft, state primitives are obtained, and the state primitives are decoded to obtain different tasks after decoding. The state primitives are input into the strategy learning module to obtain multiple strategy networks containing the motion trajectory of each task in different tasks and the task parameters of each task. When a new task to be tested is received, it is not necessary to re-acquire a large number of tasks for learning. It is only necessary to compare the task to be tested with the decoded task to obtain multiple reconstruction errors. The strategy network corresponding to the minimum value in the reconstruction error is determined as the strategy network of the task to be tested. The aircraft is controlled to make decisions through the strategy network of the task to be tested, which improves the efficiency of the aircraft's strategy learning in complex environments and has good generalization.

[0009] In some optional embodiments, the system also includes: a fourth processing unit, which is used to obtain the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to the multiple different tasks based on the multiple different tasks and using Bayesian theory before the first processing unit works.

[0010] In some optional embodiments, based on multiple different tasks, using Bayesian theory, the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to the multiple different tasks are obtained, which is implemented by the following formula:

[0011] (1)

[0012] in, is the posterior probability of one of the multiple state primitives, For one of many different tasks, is a state primitive among multiple state primitives, is the probability of one of the multiple different tasks occurring under the condition that one of the multiple state primitives occurs, is the probability of one state primitive among multiple state primitives, is the probability of being one of multiple different tasks.

[0013] In some optional embodiments, the expression in the low-dimensional feature space is:

[0014] (2)

[0015] in, is the mean of the low-dimensional feature space, is the variance of the low-dimensional feature space, is the first target state space, For the encoder network It is a low-dimensional feature space;

[0016] The encoder network in the variational autoencoder projects multiple different tasks into a low-dimensional feature space to generate multiple state primitives, including:

[0017] right Sampling is performed to obtain one state primitive among multiple state primitives; wherein, , is the weight corresponding to the encoder network, For the encoder network, When it happens, The probability of occurrence, is a low-dimensional feature space.

[0018] In some optional embodiments, multiple state primitives are decoded by a decoder network in a variational autoencoder to obtain multiple decoded different tasks, which are implemented by the following formula:

[0019] (3)

[0020] in, is one of the multiple decoded different tasks, is the second target state space, For the decoder network is the weight corresponding to the decoder network, For the decoder network, When it happens, Probability of occurrence.

[0021] In some optional embodiments, a policy network corresponding to a minimum value among multiple reconstruction errors is determined as a policy network for the task to be tested, including: determining a decoded target task corresponding to the minimum value among multiple reconstruction errors based on the minimum value among multiple reconstruction errors; determining a target task among multiple different tasks corresponding to the decoded target task based on the decoded target task; and determining a policy network corresponding to the target task among multiple different tasks as the policy network for the task to be tested. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of a control system for autonomous learning of an aircraft meta-model for collaborative tasks provided by an embodiment of the present application;

[0023] Figure 2 is a schematic diagram of a plurality of state primitive generation modules provided by an embodiment of the present application;

[0024] Figure 3 It is a flow chart of a control method for autonomous learning of an aircraft meta-model for collaborative tasks provided by an embodiment of the present application;

[0025] Figure 4 It is a schematic diagram of the structure of an electronic device provided by another embodiment of the present application. DETAILED DESCRIPTION

[0026] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. However, it will be appreciated by those skilled in the art that in the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed in the present application can also be implemented. The division of the following embodiments is for the convenience of description, and the specific implementation of the present application should not be construed as any limitation, and the various embodiments can be combined and referenced with each other under the premise of no contradiction.

[0027] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] The control system for autonomous learning of the aircraft meta-model for collaborative tasks proposed in this embodiment can be as follows: Figure 1 As shown, including:

[0029] The first acquisition unit is used to acquire a plurality of different tasks of the aircraft, wherein the plurality of different tasks are located in a first target state space.

[0030] In a specific implementation, a plurality of different tasks of the aircraft in the first target state space are obtained, wherein the plurality of different tasks include obstacle avoidance tasks or other tasks, etc., which are merely exemplified here without specific limitation.

[0031] The first processing unit is used to project a plurality of different tasks into a low-dimensional feature space through an encoder network in a variational autoencoder to generate a plurality of state primitives, and decode the plurality of state primitives through a decoder network in the variational autoencoder to obtain a plurality of decoded different tasks, wherein the plurality of decoded different tasks are located in a second target state space.

[0032] In a specific implementation, the encoder network in the variational autoencoder encodes each of a plurality of different tasks to obtain a plurality of state primitives, and then decodes the plurality of state primitives through the decoder network in the variational autoencoder based on the obtained plurality of state primitives to obtain a plurality of decoded different tasks corresponding to the plurality of tasks, wherein one task in the plurality of different tasks corresponds to a state primitive, and one state primitive corresponds to a decoded task.

[0033] The system may also include: a fourth processing unit, which is used to obtain the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to the multiple different tasks based on the multiple different tasks and using Bayesian theory before the first processing unit works.

[0034] In the specific implementation, according to multiple different tasks, using Bayesian theory, we first calculate the posterior probability of each state primitive in the low-dimensional feature space corresponding to each task.

[0035] Based on multiple different tasks, using Bayesian theory, the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to multiple different tasks are obtained, which is achieved through the following formula:

[0036] (1)

[0037] in, is the posterior probability of one of the multiple state primitives, For one of many different tasks, is a state primitive among multiple state primitives, is the probability of one of the multiple different tasks occurring under the condition that one of the multiple state primitives occurs, is the probability of one state primitive among multiple state primitives, is the probability of being one of multiple different tasks.

[0038] In actual implementation, it is difficult to directly calculate the posterior probability in formula (1) ,because The probability density estimation of is difficult, based on latent variables The calculated expression is:

[0039] (4)

[0040] Variational Bayesian methods use a family of distributions The posterior probability Approximate and adopt Gaussian distribution To solve this problem, namely: , where the mean of the Gaussian distribution is , the variance is , by minimizing and The KL distance of , can be approximated by the expression of the posterior probability:

[0041] (5)

[0042] in, for and The KL distance between It is a constant about task data information.

[0043] Corresponding to the first part of formula (5), assuming that the prior probability of one state primitive among multiple state primitives is It obeys a standard Gaussian distribution, that is: , so the first part of formula (5) can be Further expressed as:

[0044] (6)

[0045] For the second part of formula (5), the encoding part and the decoding part of the variational autoencoder are obtained through multiple error calculations.

[0046] Calculation is performed based on the multiple different tasks and the multiple decoded different tasks to obtain multiple errors between the multiple different tasks and the multiple decoded different tasks.

[0047] The multiple different tasks and the multiple decoded different tasks are calculated to obtain the expressions of multiple errors between the multiple different tasks and the multiple decoded different tasks:

[0048] (7)

[0049] in, is one of multiple errors.

[0050] In fact, as shown in formula (7), the multiple errors can be expressed in another form: interpreted as the input in the variational autoencoder The negative number of the expected log-likelihood value is expressed as formula (8):

[0051] (8)

[0052] Among them, formula (8) corresponds to the second part of formula (5). In the process of continuous learning of multiple tasks, the neural network will not forget the previously learned tasks, and it is proposed to restrict the generated multiple state primitives. Specifically, after learning the first task, the neural network weights of the primitive generation module (that is, and parameters) are fixed.

[0053] The expression in the low-dimensional feature space is:

[0054] (2)

[0055] in, is the mean of the low-dimensional feature space, is the variance of the low-dimensional feature space, is the first target state space, is the encoder network, It is a low-dimensional feature space;

[0056] The encoder network in the variational autoencoder projects multiple different tasks into a low-dimensional feature space to generate multiple state primitives, including:

[0057] right Sampling is performed to obtain one state primitive among multiple state primitives; wherein, , is the weight corresponding to the encoder network, For the encoder network, When it happens, The probability of occurrence, is a low-dimensional feature space.

[0058] In a specific implementation, a primitive generation module is provided. Figure 2 is a schematic diagram of a primitive generation module according to an embodiment of the present invention, such as Figure 2 As shown, in Figure 2 In the encoder will be in the state space X Aircraft missions in Projection to low-dimensional feature space The expression in is:

[0059] (2)

[0060] Then, by sampling the Gaussian distribution, a state primitive among multiple state primitives can be obtained, where: or .

[0061] By decoding multiple state primitives through the decoder network in the variational autoencoder, multiple different tasks after decoding are obtained, which are implemented by the following formula:

[0062] (3)

[0063] in, is one of the multiple decoded different tasks, is the second target state space, For the decoder network is the weight corresponding to the decoder network, For the decoder network, When it happens, Probability of occurrence.

[0064] In the specific implementation, according to Figure 2 As shown, using the decoder A second target state space can be reconstructed based on multiple state primitives , the expression of one task among multiple decoded different tasks is:

[0065] (3)

[0066] The second target state space is a state space similar to the first target state space.

[0067] The second processing unit is used to input a plurality of state primitives corresponding to a plurality of different tasks into the strategy network to obtain a plurality of strategy networks including motion trajectories of each task in the different tasks and task parameters of each task.

[0068] In a specific implementation, one of the multiple state primitives is Input to the policy learning module, assuming a state primitive The output control signal of one task corresponding to multiple different tasks is , the policy network is , then one task of multiple different tasks Corresponding policy learning loss The expression is:

[0069] (9)

[0070] When learning multiple different tasks in sequence, it is necessary to use the strategy learning module to train different task parameters for each of the multiple different tasks. , that is, multiple policy networks containing the motion trajectories of each task in different tasks and the task parameters of each task.

[0071] Combining multiple state primitive generation and strategy learning, the proposed algorithm is trained in an end-to-end manner, and the total loss of the neural network is calculated as follows:

[0072] (10)

[0073] in, is the total loss of the neural network, and is the loss of the variational autoencoder in the state primitive generation module, For the task in the strategy learning module The strategy loss, , They are and The corresponding hyperparameters.

[0074] The second acquisition unit is used to acquire the task to be tested of the aircraft.

[0075] The calculation unit is used to calculate the reconstruction error according to the task to be tested and the multiple decoded different tasks to obtain multiple reconstruction errors.

[0076] In a specific implementation, the task to be tested is subtracted from each decoded task to obtain multiple reconstruction errors.

[0077] The determination unit is used to determine the policy network corresponding to the minimum value among the multiple reconstruction errors as the policy network for the task to be tested.

[0078] The policy network corresponding to the minimum value among multiple reconstruction errors is determined as the policy network of the task to be tested, including: based on the minimum value among the multiple reconstruction errors, determining the decoded target task corresponding to the minimum value among the multiple reconstruction errors; based on the decoded target task, determining the target task among multiple different tasks corresponding to the decoded target task; and determining the policy network corresponding to the target task among the multiple different tasks as the policy network of the task to be tested.

[0079] In a specific implementation, a minimum value among multiple reconstruction errors is determined, and based on the minimum value among the multiple reconstruction errors, a decoded target task corresponding to the minimum value among the multiple reconstruction errors is obtained; based on the decoded target task, a target task among multiple different tasks corresponding to the decoded target task is obtained; and a policy network corresponding to the target task among the multiple different tasks is determined as the policy network of the task to be tested.

[0080] The control unit is used to control the aircraft to make decisions on the task to be tested based on the strategy network of the task to be tested.

[0081] In a specific implementation, the aircraft is controlled by a strategy network according to the task to be tested to make a decision on the task to be tested.

[0082] In this embodiment, a first acquisition unit is used to acquire multiple different tasks of the aircraft, wherein the multiple different tasks are located in a first target state space; a first processing unit is used to project the multiple different tasks into a low-dimensional feature space through an encoder network in a variational autoencoder to generate multiple state primitives, and decode the multiple state primitives through a decoder network in the variational autoencoder to obtain multiple decoded different tasks, wherein the multiple decoded different tasks are located in a second target state space; a second processing unit is used to input multiple state primitives corresponding to the multiple different tasks into a policy network to obtain multiple policy networks including motion trajectories of each task in different tasks and task parameters of each task; a second acquisition unit is used to acquire a task to be tested of the aircraft; a calculation unit is used to calculate a reconstruction error based on the task to be tested and the multiple decoded different tasks to obtain multiple reconstruction errors; a determination unit is used to determine the policy network corresponding to the minimum value of the multiple reconstruction errors as the policy network of the task to be tested; a control unit is used to determine the policy network corresponding to the minimum value of the multiple reconstruction errors as the policy network of the task to be tested; The network controls the aircraft to make decisions on the tasks to be tested, which solves the problem that when the aircraft performs tasks such as planning and decision-making, it is currently through specially programmed and offline trained algorithms. However, specially programmed and offline trained algorithms require the collection of a large number of training samples for intelligent decision-making and a large amount of training time to learn the optimal strategy, resulting in low efficiency and poor generalization of aircraft strategy learning in complex environments. The historical tasks of the aircraft are projected to obtain state primitives, the state primitives are decoded to obtain different tasks after decoding, and the state primitives are input into the strategy learning module to obtain the task parameters and operation trajectory of each historical task. When a new task to be tested is received, there is no need to re-acquire a large number of tasks for learning. It is only necessary to compare the task to be tested with the decoded task to obtain multiple reconstruction errors. The strategy network corresponding to the minimum value in the reconstruction error is determined as the strategy network of the task to be tested. The aircraft is controlled to make decisions through the strategy network of the task to be tested, which improves the efficiency of aircraft strategy learning in complex environments and has a good generalization technical effect.

[0083] The specific process of the control method for autonomous learning of the aircraft meta-model for collaborative tasks proposed in this embodiment can be as follows: Figure 3 As shown, including:

[0084] Step 301 : Acquire a plurality of different tasks of an aircraft, wherein the plurality of different tasks are located in a first target state space.

[0085] Step 302, projecting multiple different tasks into a low-dimensional feature space through the encoder network in the variational autoencoder to generate multiple state primitives, and decoding the multiple state primitives through the decoder network in the variational autoencoder to obtain multiple decoded different tasks, wherein the multiple decoded different tasks are located in the second target state space.

[0086] In step 302, the encoder network in the variational autoencoder projects the first historical task state information into a low-dimensional feature space to generate state primitives. The method may also include: based on multiple different tasks, using Bayesian theory to obtain the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to the multiple different tasks.

[0087] Step 3021, based on multiple different tasks, using Bayesian theory, obtain the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to the multiple different tasks, which is achieved by the following formula:

[0088] (1)

[0089] in, is the posterior probability of one of the multiple state primitives, For one of many different tasks, is a state primitive among multiple state primitives, is the probability of one of the multiple different tasks occurring under the condition that one of the multiple state primitives occurs, is the probability of one state primitive among multiple state primitives, is the probability of being one of multiple different tasks.

[0090] Step 3022, the expression in the low-dimensional feature space is:

[0091] (2)

[0092] in, is the mean of the low-dimensional feature space, is the variance of the low-dimensional feature space, is the first target state space, is the encoder network, It is a low-dimensional feature space;

[0093] The encoder network in the variational autoencoder projects multiple different tasks into a low-dimensional feature space to generate multiple state primitives, including:

[0094] right Sampling is performed to obtain one state primitive among multiple state primitives; wherein, , is the weight corresponding to the encoder network, For the encoder network, When it happens, The probability of occurrence, is a low-dimensional feature space.

[0095] Step 3023, decoding multiple state primitives through the decoder network in the variational autoencoder to obtain multiple decoded different tasks, which is implemented by the following formula:

[0096] (3)

[0097] in, is one of the multiple decoded different tasks, is the second target state space, is the decoder network, is the weight corresponding to the decoder network, For the decoder network, When it happens, Probability of occurrence.

[0098] Step 303: input a plurality of state primitives corresponding to a plurality of different tasks into a policy network, and obtain a plurality of policy networks including motion trajectories of each of the different tasks and task parameters of each task.

[0099] Step 304: Obtain the mission to be tested of the aircraft.

[0100] Step 305 , calculating a reconstruction error according to the task to be tested and a plurality of different tasks after decoding to obtain a plurality of reconstruction errors.

[0101] Step 306: determine the policy network corresponding to the minimum value among the multiple reconstruction errors as the policy network for the task to be tested.

[0102] Step 3061, based on the minimum value among the multiple reconstruction errors, determine the decoded target task corresponding to the minimum value among the multiple reconstruction errors.

[0103] Step 3062: Based on the decoded target task, determine a target task among a plurality of different tasks corresponding to the decoded target task.

[0104] Step 3063: determine the policy network corresponding to the target task among the multiple different tasks as the policy network of the task to be tested.

[0105] Step 307 , the aircraft is controlled based on the strategy network of the task to be tested to make a decision on the task to be tested.

[0106] In this embodiment, a plurality of different tasks of the aircraft are obtained, wherein the plurality of different tasks are located in a first target state space; the plurality of different tasks are projected into a low-dimensional feature space through an encoder network in a variational autoencoder to generate a plurality of state primitives, and the plurality of state primitives are decoded through a decoder network in the variational autoencoder to obtain a plurality of decoded different tasks, wherein the plurality of decoded different tasks are located in a second target state space; a plurality of state primitives corresponding to the plurality of different tasks are input into a policy network to obtain a plurality of policy networks including motion trajectories of each task in different tasks and task parameters of each task; a task to be tested of the aircraft is obtained; a reconstruction error is calculated based on the task to be tested and the plurality of decoded different tasks to obtain a plurality of reconstruction errors; a policy network corresponding to the minimum value of the plurality of reconstruction errors is determined as the policy network of the task to be tested; the aircraft is controlled to make decisions on the task to be tested based on the policy network of the task to be tested, thereby solving the problem that when the aircraft executes planning and However, the specially programmed and offline trained algorithms need to collect a large number of training samples for intelligent decision-making and a large amount of training time to learn the optimal strategy, which leads to the technical problems of low efficiency and poor generalization in the strategy learning of the aircraft in a complex environment. The state primitives are projected on the historical tasks of the aircraft to obtain state primitives, which are decoded to obtain different tasks after decoding, and the state primitives are input into the strategy learning module to obtain the task parameters and operation trajectory of each historical task. When a new task to be tested is received, there is no need to re-acquire a large number of tasks for learning. It is only necessary to compare the task to be tested with the decoded task to obtain multiple reconstruction errors. The strategy network corresponding to the minimum value of the reconstruction error is determined as the strategy network of the task to be tested. The aircraft is controlled to make decisions through the strategy network of the task to be tested, which improves the efficiency of the strategy learning of the aircraft in a complex environment and has a good generalization technical effect.

[0107] Another embodiment of the present application relates to an electronic device, such as Figure 4 As shown, it includes: at least one processor 401; and a memory 402 that is communicatively connected to the at least one processor 401; wherein the memory 402 stores instructions that can be executed by the at least one processor 401, and the instructions are executed by the at least one processor 401 so that the at least one processor 401 can execute a control system for autonomous learning of an aircraft meta-model for collaborative tasks in the above-mentioned embodiments.

[0108] Among them, the memory and the processor are connected in a bus manner, and the bus includes any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripherals, voltage regulators, and power management circuits, which are all well known in the art, and therefore, this application will not further describe them. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium through an antenna, and further, the antenna also receives data and transmits the data to the processor.

[0109] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0110] Another embodiment of the present application relates to a computer-readable storage medium storing a computer program, which implements the above method embodiment when executed by a processor.

[0111] That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of each embodiment method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, ROM (Read-Only Memory), RAM (Random Access Memory), disk or optical disk and other media that can store program codes.

[0112] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. A control system for autonomous learning of aircraft meta-model for collaborative tasks, characterized in that: include: A first acquisition unit, configured to acquire a plurality of different tasks of the aircraft, wherein the plurality of different tasks are located in a first target state space; A first processing unit is used to project the multiple different tasks into a low-dimensional feature space through an encoder network in the variational autoencoder to generate multiple state primitives, and decode the multiple state primitives through a decoder network in the variational autoencoder to obtain multiple decoded different tasks, wherein the multiple decoded different tasks are located in a second target state space; A second processing unit is used to input a plurality of state primitives corresponding to a plurality of different tasks into a policy network to obtain a plurality of policy networks including motion trajectories of each of the different tasks and task parameters of each task; A second acquisition unit, used for acquiring a task to be tested of the aircraft; A calculation unit, used for calculating a reconstruction error according to the task to be tested and a plurality of different tasks after decoding, to obtain a plurality of reconstruction errors; A determination unit, used to determine the policy network corresponding to the minimum value among the multiple reconstruction errors as the policy network for the task to be tested; The control unit is used to control the aircraft to make decisions on the task to be tested based on the strategy network of the task to be tested.

2. The control system for autonomous learning of aircraft meta-model for collaborative tasks according to claim 1, characterized in that: The system further comprises: The fourth processing unit is used to obtain the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to the multiple different tasks based on the multiple different tasks and using the Bayesian theory before the first processing unit works.

3. The control system for autonomous learning of aircraft meta-model for collaborative tasks according to claim 2, characterized in that: Based on multiple different tasks, using Bayesian theory, the posterior probabilities of multiple state primitives in the low-dimensional feature space corresponding to multiple different tasks are obtained, which is achieved through the following formula: in, is the posterior probability of one of the multiple state primitives, For one of many different tasks, is a state primitive among multiple state primitives, is the probability of one of the multiple different tasks occurring under the condition that one of the multiple state primitives occurs, is the probability of one state primitive among multiple state primitives, is the probability of being one of multiple different tasks.

4. The control system for autonomous learning of aircraft meta-model for collaborative tasks according to claim 3, characterized in that: The expression in the low-dimensional feature space is: in, is the mean of the low-dimensional feature space, is the variance of the low-dimensional feature space, is the first target state space, is the encoder network, It is a low-dimensional feature space; The encoder network in the variational autoencoder projects multiple different tasks into a low-dimensional feature space to generate multiple state primitives, including: right Sampling is performed to obtain one state primitive among multiple state primitives; wherein, , is the weight corresponding to the encoder network, For the encoder network, When it happens, The probability of occurrence, is a low-dimensional feature space.

5. The control system for autonomous learning of aircraft meta-model for collaborative tasks according to claim 3, characterized in that: By decoding multiple state primitives through the decoder network in the variational autoencoder, multiple different tasks after decoding are obtained, which are implemented by the following formula: in, is one of the multiple decoded different tasks, is the second target state space, is the decoder network, is the weight corresponding to the decoder network, For the decoder network, When it happens, Probability of occurrence.

6. The control system for autonomous learning of aircraft meta-model for collaborative tasks according to claim 1, characterized in that: The policy network corresponding to the minimum value among the multiple reconstruction errors is determined as the policy network of the task to be tested, including: Based on a minimum value among the multiple reconstruction errors, determining a decoded target task corresponding to the minimum value among the multiple reconstruction errors; Based on the decoded target task, determining a target task among a plurality of different tasks corresponding to the decoded target task; The policy network corresponding to the target task among multiple different tasks is determined as the policy network of the task to be tested.

7. A control method for autonomous learning of aircraft meta-model for collaborative tasks, characterized in that: include: Acquire a plurality of different tasks of the aircraft, wherein the plurality of different tasks are located in a first target state space; Projecting a plurality of different tasks into a low-dimensional feature space through an encoder network in a variational autoencoder to generate a plurality of state primitives, and decoding the plurality of state primitives through a decoder network in the variational autoencoder to obtain a plurality of decoded different tasks, wherein the plurality of decoded different tasks are located in a second target state space; Inputting a plurality of state primitives corresponding to a plurality of different tasks into a policy network, and obtaining a plurality of policy networks including motion trajectories of each task in the different tasks and task parameters of each task; Get the mission to be tested of the aircraft; Calculate the reconstruction error according to the task to be tested and the multiple decoded different tasks to obtain multiple reconstruction errors; The policy network corresponding to the minimum value among the multiple reconstruction errors is determined as the policy network of the task to be tested; The strategy network based on the task to be tested controls the aircraft to make decisions on the task to be tested.

8. An electronic device, characterized in that: include: at least one processor; and a memory communicatively coupled to the at least one processor; Wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a control system for autonomous learning of an aircraft metamodel for collaborative tasks as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a control system for autonomous learning of an aircraft meta-model for collaborative tasks as described in any one of claims 1 to 6 is implemented.

Images