Task processing method and device

Abstract

The embodiment of the invention discloses a task processing method and device which are used for improving the processing accuracy of an unknown prediction task. The method comprises the steps that the computing device obtains a machine model and a mechanism model, the machine model is used for predicting mechanism parameters of the mechanism model, and the mechanism model is used for executing an unknown prediction task; and the computing device inputs the machine model as a mechanism parameter of the mechanism model into the mechanism model to obtain a fusion model, and trains the fusion model based on the training data. The computing device updates the mechanism parameters of the mechanism model based on the machine model corresponding to the training result of the fusion model to obtain an updated mechanism model, the training result of the fusion model is used for indicating the target model parameters of the machine model, and the target model parameters are used for determining the mechanism parameters of the updated mechanism model. And the computing device executes an unknown prediction task based on the updated mechanism model.

Description

[0001] This application claims the priority of the Chinese patent application filed with the State Intellectual Property Office on August 15, 2023, with application number 202311028800.X and invention name “A task processing method and device”, the entire contents of which are incorporated by reference in this application. Technical Field

[0002] The embodiments of the present application relate to the field of computers, and in particular to a task processing method and device. Background Art

[0003] With the development of artificial intelligence technology, machine models have gradually become the mainstream solution for unknown task processing problems. Unknown task processing problems refer to the problem of computing devices trying to solve tasks that have never been seen and trained in machine learning. Usually, machine models need to receive specific types of training data during the training phase in order to make accurate predictions on similar types of tasks during the testing phase.

[0004] In order to improve the processing effect of machine models on unknown tasks, the current method of integrating machine models and mechanism models is usually used to process unknown tasks. There are many ways to integrate machine models and mechanism models. For example, the computing device integrates the machine model and the mechanism model to obtain an integrated model, and uses the training data to train the integrated model, and then uses the trained integrated model to predict unknown tasks.

[0005] In the current integration method, when the computing device predicts an unknown task based on the integrated model, the prediction result of the unknown task is obtained by using the sum of the prediction results of the mechanism model and the machine model. Since the mechanism model and the machine model in the integrated model maintain their respective independence, the mechanism model and the machine model jointly play a role in the prediction result. Since the machine model is a black box model, the model has poor interpretability. At the same time, the machine model directly affects the prediction result, which leads to poor interpretability of the integrated model. Summary of the invention

[0006] The embodiment of the present application provides a task processing method, in which the computing device can generate a fusion model using the machine model as the mechanism parameter of the mechanism model, and train the fusion model, and update the mechanism parameter of the mechanism model based on the machine model corresponding to the training result of the fusion model. Since the machine model is used to predict the mechanism parameter and does not directly act on the prediction result of the unknown task, the embodiment of the present application improves the prediction accuracy of the mechanism model for the unknown task while ensuring the interpretability of the prediction result. The embodiment of the present application also provides a task processing device, a computing device, a computing device cluster, a computer-readable storage medium, and a computer program product corresponding to the task processing method.

[0007] In the first aspect, an embodiment of the present application provides a task processing method, which can be executed by a computing device, or by a component of the computing device, such as a processor, chip or chip system of the computing device, or by a logic module or software that can realize all or part of the functions of the computing device. The method provided in the first aspect includes: the computing device acquires a machine model and a mechanism model, the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to perform unknown prediction tasks. The computing device uses the machine model as the mechanism parameter of the mechanism model to obtain a fusion model, and trains the fusion model based on the training data. The computing device updates the mechanism parameters of the mechanism model based on the machine model corresponding to the training result of the fusion model, and obtains an updated mechanism model. The training result of the fusion model is used to indicate the target model parameters of the machine model, and the target model parameters are used to determine the mechanism parameters of the updated mechanism model. The computing device performs unknown prediction tasks based on the updated mechanism model.

[0008] Mechanism parameters of the mechanism model In the embodiment of the present application, the computing device can determine the mechanism parameters of the mechanism model by training the fusion model. Since the machine model in the fusion model does not directly affect the prediction results of the unknown prediction task, that is, the machine model and the mechanism model are decoupled from each other, the mechanism model in the embodiment of the present application supports the fusion of multiple machine models, thereby improving the prediction accuracy of the mechanism model for unknown tasks while ensuring the interpretability of the prediction results of the unknown prediction task.

[0009] In one possible implementation, before the computing device determines the trained fusion model, the computing device obtains training data and model constraint information, and the model constraint information is used to constrain the range of the mechanism parameters in the mechanism model. The computing device trains the fusion model based on the training data to obtain a trained machine model, and the model output of the trained machine model satisfies the model constraint information. For example, the model constraint information can constrain the slope of the mechanism model, and the model constraint information includes an upper limit, a lower limit, a positive correlation, or a negative correlation of the slope.

[0010] In the embodiment of the present application, the computing device can train the fusion model based on the model constraint information, thereby reducing the training range of the machine model and improving the training efficiency of the fusion model. At the same time, the computing device constrains the range of model parameters in the mechanism model through the model constraint information, so that the fusion model obtained by fusion of different machine models has a higher physical consistency.

[0011] In one possible implementation, the computing device performs data augmentation on the training data of the fusion model based on the data augmentation strategy to obtain augmented training data, and the augmented training data is used to train the fusion model. Specifically, the computing device selects a data augmentation strategy for the training data based on the model constraint information, and performs data augmentation on the training data of the fusion model based on the data augmentation strategy.

[0012] In the embodiment of the present application, the computing device can select a data augmentation strategy for training data based on model constraint information, and generate augmented training data based on the data augmentation strategy, thereby improving the data quality of the augmented training data and further improving the training efficiency of the fusion model.

[0013] In one possible implementation, the computing device updates the data augmentation strategy based on the deviation between the machine model output corresponding to the augmented training data and the model constraint information, and the computing device regenerates the augmented training data based on the updated data augmentation strategy. Updating the data augmentation strategy includes reselecting the data augmentation strategy or adjusting the original data augmentation strategy.

[0014] In the embodiment of the present application, the computing device updates the data augmentation strategy based on the deviation between the machine model output corresponding to the augmented training data and the model constraint information, and regenerates the augmented training data based on the updated data augmentation strategy, thereby improving the data quality of the augmented training data and further improving the training efficiency of the fusion model.

[0015] In one possible implementation, the computing device calculates the inference deviation based on the inference results of the trained fusion model on the test data and the inference results of the mechanism model on the test data. The computing device modifies the model parameters of the trained fusion model based on the inference deviation. For example, a domain expert can conduct mechanism research based on the inference deviation to assist in updating the mechanism model.

[0016] In the embodiment of the present application, the computing device calculates the inference deviation based on the inference results of the trained fusion model on the test data and the inference results of the mechanism model on the test data, and corrects the mechanism analysis model parameters of the trained fusion model based on the inference deviation, thereby improving the prediction accuracy of the mechanism model for unknown tasks.

[0017] In one possible implementation, the computing device calculates model parameters corresponding to different test data based on the trained machine model, and the model parameters corresponding to different test data are used to calculate the constraint satisfaction of the model constraint information. The computing device evaluates the reliability of the model constraint information based on the constraint satisfaction.

[0018] In the embodiment of the present application, the computing device calculates the model parameters output by the machine model after training based on different test data, and calculates the satisfaction degree of the model parameters output by the machine model with the model constraint information, thereby evaluating the reliability of the model constraint information and updating the model constraint information, further improving the prediction accuracy of the mechanism model for unknown tasks.

[0019] In a possible implementation, before executing the unknown prediction task, the computing device needs to identify the unknown prediction task for the received task. When the prediction task is identified as an unknown prediction task, the unknown prediction task is executed based on the trained fusion model. When the prediction task is identified as a known task, the unknown prediction task is executed based on the conventional model.

[0020] In the embodiment of the present application, before the computing device executes the unknown prediction task, the computing device needs to identify the unknown prediction task for the received task and use the conventional model to process the known task, thereby improving the task processing efficiency of the computing device.

[0021] In one possible implementation, when the computing device fuses the machine model and the mechanism model to obtain the fusion model, the machine model and the mechanism model can be fused in a differentiable manner and in a non-differentiable manner. Different fusion methods for obtaining fusion models correspond to different training methods. The fusion model corresponding to the non-differentiable fusion is iteratively trained by inverse coefficient solution, and the fusion model corresponding to the differentiable fusion is trained by back propagation.

[0022] In the embodiment of the present application, the computing device adopts a variety of fusion methods based on the characteristics of the mechanism model to obtain the fusion model, thereby improving the feasibility of the solution.

[0023] In a possible implementation, the data augmentation strategy includes one or more of the following data augmentation strategies: an augmentation strategy based on domain knowledge, an adaptive augmentation strategy, a randomized augmentation strategy, and a hybrid augmentation strategy. The above data augmentation strategies can also be divided into image data augmentation strategies or text data augmentation strategies according to the type of training data.

[0024] In the embodiment of the present application, the computing device can generate augmented training data based on a variety of data augmentation strategies, thereby improving the training effect of the augmented training data on the fusion model, and further improving the accuracy of the mechanism model after training.

[0025] In a possible implementation, the computing device may search for a data augmentation strategy based on the fusion model. For example, during the data augmentation process, the computing device may predict and evaluate the augmented training data based on the fusion model, and select a corresponding data augmentation strategy based on the prediction and evaluation results.

[0026] In the embodiment of the present application, a fusion model can be introduced in the computing device data augmentation strategy search process, and the augmented training data is predicted and evaluated based on the fusion model to select the corresponding data augmentation strategy, thereby improving the data quality of the augmented training data.

[0027] In a second aspect, an embodiment of the present application provides a task processing device, which includes an interface unit, a training unit and a processing unit. Among them, the interface unit is used to obtain a machine model and a mechanism model, the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to perform unknown prediction tasks. The training unit is used to use the machine model as the mechanism parameter of the mechanism model, obtain a fusion model, and train the fusion model based on the training data. The processing unit is used to update the mechanism parameters of the mechanism model based on the machine model corresponding to the training result of the fusion model, and obtain an updated mechanism model. The training result of the fusion model is used to indicate the target model parameters of the machine model, and the target model parameters are used to determine the mechanism parameters of the updated mechanism model. The processing unit also performs unknown prediction tasks based on the updated mechanism model.

[0028] In a possible implementation, training data and model constraint information are obtained, and the model constraint information is used to constrain the range of mechanism parameters in the fusion model. The training unit is also used to train the fusion model based on the training data to obtain a trained fusion model, and the mechanism parameters of the trained fusion model meet the model constraint information.

[0029] In a possible implementation, the device further includes a data unit, and the data unit is used to perform data augmentation on the training data of the fusion model based on a data augmentation strategy to obtain augmented training data, and the augmented training data is used to train the fusion model.

[0030] In a possible implementation, the data unit is further used to update the data augmentation strategy based on the deviation between the machine model output corresponding to the augmented training data and the model constraint information, and to regenerate the augmented training data based on the updated data augmentation strategy.

[0031] In a possible implementation, the device further includes an analysis unit, which is used to calculate an inference deviation based on the inference results of the trained fusion model on the test data and the inference results of the mechanism model on the test data, and to correct model parameters of the trained fusion model based on the inference deviation.

[0032] In a possible implementation, the device further includes an analysis unit, which is further used to calculate model parameters corresponding to different test data based on the trained machine model, and the model parameters corresponding to different test data are used to statistically calculate the constraint satisfaction of the model constraint information. The reliability of the model constraint information is evaluated based on the constraint satisfaction.

[0033] In a third aspect, an embodiment of the present application provides a computing device, comprising a processor, the processor being coupled to a memory, the processor being used to store instructions, and when the instructions are executed by the processor, the computing device executes the method described in the first aspect or any possible implementation manner of the first aspect.

[0034] In a fourth aspect, an embodiment of the present application provides a computing device cluster, the computing device cluster includes one or more computing devices, the computing device includes a processor, the processor is coupled to a memory, the processor is used to store instructions, when the instructions are executed by the processor, the computing device cluster executes the method described in the first aspect or any possible implementation method of the first aspect.

[0035] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium having instructions stored thereon. When the instructions are executed, the computer executes the method described in the first aspect or any possible implementation manner of the first aspect.

[0036] In a sixth aspect, an embodiment of the present application provides a computer program product, which includes instructions. When the instructions are executed, the computer implements the method described in the first aspect or any possible implementation method of the first aspect.

[0037] It can be understood that the beneficial effects that can be achieved by any of the task processing devices, computing devices, computing device clusters, computer-readable media or computer program products provided above can refer to the beneficial effects in the corresponding methods and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the system architecture of a task processing system provided in an embodiment of the present application;

[0039] Figure 2 A flowchart of a task processing method provided in an embodiment of the present application;

[0040] Figure 3 A flowchart of another task processing method provided in an embodiment of the present application;

[0041] Figure 4 A schematic diagram of a model training process provided in an embodiment of the present application;

[0042] Figure 5 A schematic diagram of a model reasoning process provided in an embodiment of the present application;

[0043] Figure 6 A schematic diagram of a task processing device provided in an embodiment of the present application;

[0044] Figure 7A schematic diagram of the structure of a computing device provided in an embodiment of the present application;

[0045] Figure 8 A schematic diagram of the structure of a computing device cluster provided in an embodiment of the present application;

[0046] Fig. 9 A schematic diagram of the structure of another computing device cluster provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] Embodiments of the present application provide a task processing method and apparatus for improving the interpretability of prediction results of a computing device for unknown tasks.

[0048] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application 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 data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that 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.

[0049] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0050] First, some terms involved in the embodiments of the present application are introduced to facilitate technical personnel in this field to understand the technical solution.

[0051] Unknown tasks refer to test samples whose prediction performance of the machine model decreases due to large differences in distribution from training samples during model testing. Unknown task processing refers to additional processing of unknown task samples to achieve better prediction results outside the conventional process.

[0052] Data scenarios refer to data with different distributions that are artificially defined, such as data that is close to the distribution of the training set, data that deviates greatly from the distribution of the training set, data that is well predicted by machine models, and data that is well predicted by mechanism models. These all belong to data in different scenarios. A scenario classifier is a classifier that predicts data scenarios.

[0053] In order to make the technical solution of the present application clearer and easier to understand, the system architecture of the present application is introduced below with reference to the accompanying drawings.

[0054] See also Figure 1 , Figure 1 A system architecture diagram of a task processing system provided for the example of this application. Figure 1 In the system architecture shown, the task processing system 10 includes a cloud-side subsystem 100 and an edge-side subsystem 200, wherein the cloud-side subsystem 100 includes a global management node 101, a cloud-side local control node 102, a model training node 103, and a cloud-side computing node 104, and the edge-side subsystem 200 includes an edge-side local control node 201, an edge-side computing node 202, and a terminal node 203. The specific functions of each part of the task processing system 10 are described below.

[0055] The global management node 101 is used to manage and coordinate the activities of various subsystems and modules in the task processing system 10, including resource allocation and scheduling, task coordination and communication, fault detection and recovery, etc. The global management node 101 includes a model and data set management module 1011 and a control module 1012. The model and data set management module 1011 is used to manage the models and data sets in the knowledge base, the models in the knowledge base include the fusion model composed of the machine model and the mechanism model, and the data set includes the training data set of the fusion model.

[0056] The control module 1012 is used to manage the cloud-side local management node 102 and the cloud-side computing node 104, including creating a cloud-side computing node 104 and controlling the cloud-side computing node 104 to process unknown tasks, sending a model training control signal to the cloud-side local management node 102 based on a model training application sent by the cloud-side computing node 104, etc.

[0057] The cloud-side local management node 102 is used to distribute the control instructions of the global management node 101 and configure the model training node 103. Specifically, the cloud-side local management node 102 receives the control instructions sent by the global management node 101, and sends the control instructions to the model training node 103. The control instructions can be model training instructions. The cloud-side local management node 102 can also configure the model constraint information of the model training node 103 for model training.

[0058] The model training node 103 is used to train the fusion model and store the trained fusion model in the cloud storage system. The model training node 103 includes a data preprocessing module 1031, a fusion module 1032 and a training module 1033, wherein the data preprocessing module 1031 is used to augment the training data according to the model constraint information, thereby obtaining a training data set for the fusion model. The fusion module 1032 is used to fuse the machine model and the mechanism model to generate a fusion model. The training module 1033 is used to train the fusion model based on the training data set, and store the trained model in the cloud storage system.

[0059] The cloud-side computing node 104 is used to manage the knowledge base and process unknown tasks based on the fusion model. The cloud-side computing node 104 includes a knowledge base management module 1041 and an unknown task processing module 1042, wherein the knowledge base management module 1041 is used to manage the knowledge base corresponding to different tasks, and managing the knowledge base includes knowledge base initialization, knowledge base update, and sharing of historical knowledge base, etc. The knowledge base includes a variety of machine models and mechanism models. The unknown task processing module 1042 is used to receive unknown tasks, obtain the fusion model corresponding to the unknown tasks, and perform reasoning work on unknown tasks.

[0060] The edge local control node 201 is used to manage edge data and model resources, including controlling the working status and resources of the edge computing node 202. Specifically, the global management node 101 can create an edge local control node 201 on the edge, and the edge local control node 201 receives the control instructions sent by the global management node 101, and starts model deployment and unknown task processing on the edge computing node 202 according to the control instructions.

[0061] The edge computing node 202 is used to identify unknown tasks from the terminal node 203 and upload the unknown tasks to the cloud storage system. The edge computing node 202 notifies the global management node 101 to process the unknown tasks through the edge local control node 201. The edge computing node 202 is also used to receive the mechanism model updated by the user and push the mechanism model to the cloud side subsystem 100, thereby updating the mechanism model of the cloud side subsystem. The edge computing node 202 is also used to obtain the trained fusion model from the cloud storage system and perform unknown task reasoning on the edge based on the fusion model.

[0062] The edge computing node 202 includes a system library module 2021, an unknown task identification module 2022, an edge reasoning module 2023 and a mechanism model configuration module 2024. Among them, the system library module 2021 is used to provide a calling interface for the model training node 103 and the cloud computing node 104, and the terminal node 203 can call the model training and reasoning services provided by the cloud subsystem 100 through the system library module 2021. The unknown task identification module 2022 is used to receive and identify unknown tasks sent by the terminal node 203. The edge reasoning module 2023 is used to perform unknown tasks using a fusion model. The mechanism model configuration module 2024 is used to receive the mechanism model provided by the user and update the mechanism model in the cloud storage system.

[0063] The terminal node 203 includes a data acquisition and instruction forwarding module 2031 and a device module 2032. The data acquisition and instruction forwarding module 1031 is used to receive the prediction results sent by the edge computing node 202, and send control instructions to the device module 2032 according to the prediction results. The data acquisition and instruction forwarding module 1031 is also used to obtain data collected by sensors, such as quality detection data of the device module 2032, and send the data collected by the sensors to the edge computing node 202.

[0064] The device module 2032 is used to receive the control instructions sent by the data acquisition and instruction forwarding module 2031 and execute the control instructions. The device module 2032 can specifically execute the device, such as the controlled device, the input control pipeline and the monitoring instrument.

[0065] based on Figure 1 The task processing system 10 shown in the figure, the present application also provides a task processing method. The task processing method provided by the embodiment of the present application is introduced below in combination with the embodiment.

[0066] See also Figure 2 , Figure 2 A flowchart of a task processing method provided in an embodiment of the present application. Figure 2 In the example shown, the method includes the following steps:

[0067] 201. The task processing system 10 obtains a machine model and a mechanism model, the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to execute unknown prediction tasks.

[0068] The task processing system 10 receives unknown prediction tasks, wherein the unknown prediction tasks may be unknown prediction tasks sent from terminal devices, such as anomaly detection tasks, future trend prediction tasks, risk assessment tasks, and user behavior prediction tasks.

[0069] For example, in an example of step 201, the unknown prediction task received by the task processing system 10 is a coal production quality prediction task, and the coal production quality prediction task is a regression prediction task. In the coal production quality prediction task, the task processing system 10 needs to predict the quality of coke produced based on the detection index of the coal raw material and various production parameters in coal production. Among them, the detection index of the raw coal and the various production parameters in coal production include data of multiple dimensions, such as the elements, batching ratio and moisture content of the raw coal.

[0070] For another example, in step 201, the unknown prediction task received by the task processing system 10 is a human body temperature feeling prediction task. In the human body temperature feeling prediction task, the task processing system 10 needs to predict the human body temperature feeling based on a variety of collected data. Among them, the various collected data include normal human metabolic rate MET, human body extra work WME, air temperature TA, water vapor pressure PA, and mean radiation temperature TR.

[0071] In a possible implementation, before executing the unknown prediction task, the task processing system 10 needs to identify the received task as an unknown prediction task. Specifically, after receiving the prediction task sent by the terminal device, the task processing system 10 identifies the prediction task as an unknown task. When the prediction task is identified as an unknown prediction task, the unknown prediction task is executed based on the trained mechanism model. When the prediction task is identified as a known task, the unknown prediction task is executed based on the conventional model.

[0072] See also Figure 3 , Figure 3 A flowchart of another task processing method provided in an embodiment of the present application. Figure 3 In the example shown, after receiving the task data, the task processing system 10 identifies an unknown task based on the task data. When the task data is identified as a known task, the task processing system 10 executes the task through a conventional model corresponding to the known task. When the task data is identified as an unknown task, the task processing system 10 inputs the task data into a trained mechanism model, and the trained mechanism model executes the unknown task.

[0073] Before the task processing system 10 receives an unknown prediction task, the task processing system 10 needs to obtain a mechanism model and a machine model, and generate a fusion model based on the machine model and the mechanism model, wherein the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to execute the unknown prediction task.

[0074] 202. The task processing system 10 uses the machine model as the mechanism parameter of the mechanism model to obtain a fusion model, and trains the fusion model based on the training data.

[0075] After the task processing system 10 obtains the mechanism model and the machine model, it generates a fusion model based on the machine model and the mechanism model. Specifically, the task processing system 10 inputs the machine model as the mechanism parameter of the mechanism model into the mechanism model to obtain the fusion model. Afterwards, the task processing system 10 trains the fusion model based on the training data. The following is a detailed introduction:

[0076] The task processing system 10 first generates a fusion model, and then trains the fusion model based on the training data to obtain a trained fusion model. The training result of the fusion model includes a trained machine model, the mechanism model is used to perform unknown prediction tasks, and the trained machine model is used to update the mechanism parameters in the mechanism model.

[0077] Specifically, in the process of acquiring the fusion model, the task processing system 10 acquires the mechanism model and the machine model respectively, and inputs the machine model into the mechanism model as the mechanism parameter of the mechanism model to obtain the fusion model. For example, the model training node 103 of the cloud-side subsystem 100 acquires the machine model and the mechanism model from the knowledge base, and generates the fusion model based on the fusion of the mechanism model and the machine model.

[0078] See also Figure 4 , Figure 4 A flow chart of a fusion model training process provided in an embodiment of the present application. Figure 4 In the example shown, the model training node 103 obtains the machine model and mechanism model in the knowledge base. For example, the machine model is a machine learning model selected by the user or by the system default. ml The mechanism model is the formula f obtained by the user to model the actual problem phy The model training node 103 generates a fusion model based on the fusion of the mechanism model and the machine model. The fusion modes of the mechanism model and the machine model include differentiable fusion and non-differentiable fusion. Differentiable fusion means that differential operations can be performed during the fusion process. Non-differentiable fusion means that differential operations cannot be performed during the fusion process.

[0079] For example, in the above-mentioned coal production quality prediction task example, the mechanism model acquired by the task processing system 10 is a linear function, for example, the formula of the mechanism model satisfies f phy (x) = b + a1x1 + a2x3 + a3x4, the task processing system 10 obtains the machine models respectively and The task processing system 10 obtains a fusion model based on the mechanism model and the machine model to satisfy the following:

[0080]

[0081] Among them, x r represents the input of the machine model, and The initial values ​​are 55, -24.3, 29.3 and 5.8.

[0082] For another example, in a human body temperature perception prediction task, the task processing system 10 determines a mechanism model based on a calculation formula of a predicted mean value PMV. The mechanism formula of the mechanism model is, for example, the following formula:

[0083] pmv=(0.03+0.3*exp(-2.1*met))*(-0.08*met*(34-ta)-0.001*met*(5867-pa)+

[0084] 59.4*met+0.003*met-59.4*wme-max(0,24.4met-24.4wme-24.4)-17.5-others).

[0085] Among them, pmv is the predicted mean evaluation value, met is the normal metabolic rate of the human body, ta is the air temperature, pa is the water vapor pressure, and wme is the extra work done by the human body.

[0086] In the above human body temperature perception prediction task processing process, the task processing system 10 uses multiple neural network models a i The constant parameters in the above mechanism formula are predicted respectively, and a fusion model including the mechanism model and the neural network model is obtained. First, the neural network is made to learn a i The original constant parameters are then calculated using a neural network during the training process, and the PMV value is further calculated. The task processing system 10 calculates the loss value based on the label of the human body temperature perception and uses normal back propagation to train the network. The calculation of the loss value satisfies the following formula:

[0087]

[0088] exist Figure 4 In an example shown, if the model coefficients of the mechanism model can be inversely solved, for example, the mechanism model is a linear function, the mechanism model and the machine model are non-differentiably fused. If the model structure of the mechanism model is more complex, for example, the mechanism model is a discrete function, the machine model supports back propagation, and the mechanism model and the machine model are differentially fused.

[0089] It should be noted that both fusion methods use the machine model to predict the mechanism parameters of the mechanism model. The two fusion methods correspond to different training methods. The fusion model corresponding to the non-differentiable fusion uses the inverse coefficient method for iterative training, while the fusion model corresponding to the differentiable fusion uses the back propagation method for training.

[0090] 203. Based on the machine model corresponding to the training result of the fusion model, the mechanism parameters of the mechanism model are updated to obtain an updated mechanism model.

[0091] The task processing system 10 updates the mechanism parameters of the mechanism model based on the machine model corresponding to the training results of the fusion model to obtain an updated mechanism model. The training results of the fusion model are used to indicate the target model parameters of the machine model, and the target model parameters are used to determine the mechanism parameters of the updated mechanism model.

[0092] In the embodiment of the present application, the updated mechanism model can also be called a trained fusion model, because the association between the fusion model and the mechanism model in the present application lies in that when the mechanism parameters in the mechanism model are the machine model, the mechanism model can also be called the fusion model, and when the mechanism parameters in the mechanism model are the deterministic mechanism parameters output by the machine model (the deterministic mechanism parameters are the machine parameters output by the machine model when the model parameters of the machine model are the target model parameters), the mechanism model is the updated mechanism model.

[0093] In a possible implementation, the task processing system 10 obtains training data and model constraint information, wherein the training data includes an augmented training data set obtained by performing data augmentation based on a data augmentation strategy, and the model constraint information is used to constrain the range of model parameters in the mechanism model. For example, when the mechanism model is a linear function, the model constraint information can constrain the slope of the mechanism model, and the model constraint information includes an upper bound, a lower bound, a positive correlation, or a negative correlation of the slope. The task processing system 10 trains the fusion model based on the training data to obtain a trained machine model, and the trained machine model satisfies the model constraint information.

[0094] For example, in an example of a coal production quality prediction task, for the fusion model The model constraint information obtained by the task processing system 10 satisfies the following:

[0095]

[0096] Please continue reading Figure 4 ,exist Figure 4In the example shown, after the model training node 103 obtains the fusion model based on the machine model and the mechanism model, it trains the fusion model based on the training data set. For example, for the training of the fusion model obtained by differentiable fusion, the model training node 103 embeds the machine model as a mechanism parameter into the mechanism model for joint training, and updates the model parameters of the machine model through the back propagation algorithm. Specifically, the model training node 103 defines the loss function of the fusion model, and the loss function can measure the difference between the predicted result and the actual result of the fusion model. During the fusion model training process, the model training node 103 updates the model parameters of the machine model based on the loss function.

[0097] For another example, for the training of the fusion model obtained by non-differentiable fusion, the model training node 103 determines the training data of the machine model based on the inverse solution of the training data of the fusion model. The model training node 103 trains the machine model based on the training data obtained based on the inverse solution coefficients, and then embeds the trained machine model into the mechanism model to obtain the fusion model.

[0098] In a possible implementation, in the process of acquiring training data, the task processing system 10 performs data augmentation on the training data of the fusion model based on the data augmentation strategy to obtain augmented training data, and the augmented training data is used to train the fusion model. Specifically, the model training node 103 searches for the data augmentation strategy based on the training data and the model constraint information, and after determining the data augmentation strategy, the model training node 103 performs data augmentation based on the data augmentation strategy.

[0099] In the embodiments of the present application, the data augmentation strategy includes one or more of the following data augmentation strategies: an augmentation strategy based on domain knowledge, an adaptive augmentation strategy, a randomized augmentation strategy, and a hybrid augmentation strategy. Among them, the augmentation strategy based on domain knowledge refers to targeted data augmentation based on knowledge of a specific domain or task. The adaptive augmentation strategy automatically adjusts the data augmentation strategy based on the feedback information of the fusion model during the training process. The randomized augmentation strategy randomly selects augmentation operations and parameters within a certain augmentation space. The hybrid augmentation strategy is a data augmentation strategy that combines multiple augmentation methods.

[0100] It can be understood that for different types of training data, data augmentation strategies can also be divided into image data augmentation strategies or text data augmentation strategies according to the type of training data. For example, the data augmentation operations in the image data augmentation strategy include data flipping, data trimming or data scaling, and the data augmentation operations in the text data augmentation strategy include synonym replacement, word insertion or word deletion, etc.

[0101] The corresponding augmentation parameters in the data augmentation strategy in the embodiment of the present application include augmentation slope, augmentation amplitude and augmentation times. Among them, the augmentation slope is used to describe the parameter adjustment range of continuous transformation in the augmentation operation. The augmentation amplitude is used to indicate the degree or size of the augmentation operation. The augmentation times refers to the number of times the augmentation operation is repeated.

[0102] Please continue reading Figure 4 ,exist Figure 4 In the example shown, when the model training node 103 performs training data augmentation based on the model constraint information and the training data, the model training node 103 analyzes the data characteristics based on the training data and performs clustering based on the data characteristics of the training data. Then, the model training node 103 selects a data augmentation strategy based on the clustering result and the model constraint information, and performs data augmentation based on the augmentation strategy to obtain augmented training data.

[0103] exist Figure 4 In the example shown, after the model training node 103 performs clustering based on the data characteristics of the training data, it counts the points in the training data whose distance deviation from the cluster center exceeds the threshold, and regards the points in the training data whose distance deviation from the cluster center exceeds the threshold as abnormal points. The model training node 103 can analyze the abnormal points, for example, analyze the distribution characteristics of the abnormal points, and adjust the data augmentation strategy according to the analysis results. At the same time, the model training node 103 can also analyze the abnormal points in the augmented training data after data augmentation, and update the data augmentation strategy according to the analysis results.

[0104] In a possible implementation, when the model training node 103 searches for a data augmentation strategy based on the model constraint information and generates augmented training data, the model training node 103 updates the data augmentation strategy based on the deviation between the machine model output corresponding to the augmented training data and the model constraint information, and regenerates the augmented training data based on the updated data augmentation strategy. Updating the data augmentation strategy includes reselecting the data augmentation strategy or adjusting the original data augmentation strategy.

[0105] Please continue reading Figure 4 ,exist Figure 4 In the example shown, when the model training node 103 searches for a data augmentation strategy based on model constraint information, the model training node 103 may input augmented training data into the machine model, and select a data augmentation strategy based on the deviation between the machine model output and the model constraint information.

[0106] In a possible implementation, the task processing system 10 may also search for data augmentation strategies based on the fusion model. Specifically, the model training node 103 searches for data augmentation strategies based on the fusion model and model constraint information. For example, during the data augmentation process, the model training node 103 may predict and evaluate the augmented training data based on the fusion model, and select a corresponding data augmentation strategy based on the prediction and evaluation results.

[0107] For example, in the above-mentioned coal production quality prediction task example, the training data set acquired by the task processing system 10 satisfies D={(x i ,y i )}, then the augmented training data set satisfies the following formula:

[0108]

[0109] Among them, k is the augmentation slope, that is, the slope that the augmented data is expected to reach, within the slope constraint interval, σ is the augmentation amplitude, and the augmentation multiple is N, which means that N augmentations are performed, so that |D aug |=N|D|.

[0110] In the above example, the task processing system 10 discretizes the slope requirement interval of each dimension, such as the slope constraint interval for dimension 1 is [-30, -20], and then discretizes to obtain 5 values ​​[-30, -27.5, -25, -22.5, -20], reducing the search space. The task processing system 10 obtains the augmentation multiple interval [0, 1, 2, 3] and the augmentation amplitude range [0, 0.5, 1] ​​through hyperparameters. The task processing system 10 traverses the augmented slope value of each dimension, and experimentally verifies that the error of the model trained on the augmented data shows a convex function relationship with the augmentation multiple and the augmentation amplitude on the data before augmentation. The task processing system 10 sets the initial value of the augmentation multiple to 0 and the initial value of the augmentation amplitude to 0. Monotonically increase the augmentation multiple and the augmentation amplitude, record the prediction error and constraint satisfaction of the model on the original data under the current augmentation strategy, and obtain the augmentation penalty value.

[0111] Error=Error mac +C*Error slope

[0112]

[0113]

[0114] Among them, C is a hyperparameter used to measure the slope satisfaction and model prediction accuracy. Represents the estimated slope of the constrained dimension of the model on the current batch of data. For the estimated slope of the model on dimension 1, consider the partial order dependence function For x1 and Do a linear regression and get the estimated slope When the Error is found to increase, the search stops. Save the data augmentation strategy under the current batch augmentation slope value. Return the augmentation slope with the lowest Error and the corresponding data augmentation strategy.

[0115] Please continue reading Figure 4 ,exist Figure 4 In the example shown, the model training node 103 can use the fusion model as the input for the data augmentation strategy search. It can be understood that when the fusion model is used as the input, the model training node 103 will search for the data augmentation strategy for the fusion model, which will increase the search time but can improve the accuracy of the data augmentation strategy. The model training node 103 may not input the fusion model when searching for the data augmentation strategy, and the model training node 103 uses the default algorithm to search for the data augmentation strategy.

[0116] After the task processing system 10 trains the fusion model based on the augmented training data, the trained mechanism model is determined based on the training results of the fusion model. The training results of the fusion model include the trained machine model. The trained machine model is used to output update parameters. The task processing system 10 updates the mechanism model based on the update parameters to obtain the trained mechanism model.

[0117] For example, in the above-mentioned coal production quality prediction task example, during the training process of the fusion model, the task processing system 10 iterates N times. In each iteration, the machine model satisfies the following:

[0118]

[0119]

[0120] The cut(x) function trims out-of-bounds values. For example, if the slope constraint of the first dimension is (-30, -20), if the value of x is not within the interval (-30, -20), the value of x1 is forced to be the nearest boundary value -30 or -20.

[0121] See also Figure 5 , Figure 5 A schematic diagram of a flow chart of reasoning using a fusion model provided in an embodiment of the present application. Figure 5 In the example shown, the task processing system 10 infers the trained mechanism model based on the test data, wherein the test data includes unknown prediction tasks. Specifically, the task processing system 10 inputs the test data into the trained mechanism model to obtain an inference result.

[0122] In a possible implementation, the task processing system 10 calculates the inference deviation based on the inference result of the trained mechanism model on the test data and the inference result of the mechanism model on the test data. The task processing system 10 modifies the model parameters of the trained mechanism model based on the inference deviation.

[0123] Please continue reading Figure 5 ,exist Figure 5 In the example shown, the task processing system 10 infers the trained mechanism model based on the test data. After obtaining the inference result, the task processing system 10 can also input the test data into the pre-trained mechanism model to obtain the inference result of the previous model before training, and calculate the inference deviation based on the inference results of the mechanism model before and after training. The user can conduct a mechanism study of the mechanism model based on the inference deviation, thereby correcting the mechanism parameters of the mechanism model and improving the accuracy of the mechanism model.

[0124] For example, in the above-mentioned coal production quality prediction task example, the task processing system 10 uses the prediction difference between the trained mechanism model and the original machine model as the error estimate of the original mechanism model. When the error exceeds the set threshold, the current sample is considered to be a difficult sample relative to the original mechanism model. The task processing system 10 identifies and corrects the mechanism defects of the original mechanism model based on the prediction error.

[0125] In a possible implementation, the task processing system 10 calculates model parameters corresponding to different test data based on the trained machine model, and the model parameters corresponding to different test data are used to statistically evaluate the constraint satisfaction of the model constraint information. The task processing system 10 updates the reliability of the evaluation model constraint information based on the constraint satisfaction.

[0126] Please continue reading Figure 5 ,exist Figure 5 In the example shown, the task processing system 10 infers the trained machine model based on the test data to obtain the machine model inference result corresponding to the test data, and the inference result is the mechanism parameter of the mechanism model. The task processing system 10 can count the model constraint deviation values ​​corresponding to different test data based on the model constraint information, and evaluate the reliability of the model constraint information based on the model constraint deviation value.

[0127] For example, in the above-mentioned coal production quality prediction task example, the task processing system 10 perturbs the input test data sample x in the domain of x to obtain multiple test samples, uses the trained mechanism model for reasoning, and then estimates the slope of the model in the domain of x based on the partial order dependency function, and determines whether it exceeds the slope constraint corresponding to the model constraint information, as the content of the reliability evaluation of the model constraint information.

[0128] 204.Perform unknown prediction tasks based on the updated mechanism model.

[0129] The task processing system 10 executes unknown prediction tasks based on the updated mechanism model. The unknown prediction tasks may be unknown prediction tasks sent from the terminal device, such as anomaly detection tasks, future trend prediction tasks, risk assessment tasks, and user behavior prediction tasks.

[0130] Please continue reading Figure 3 ,exist Figure 3 In the example shown, before the task processing system 10 processes the task, the task processing system 10 can receive the mechanism model file, model constraint information and machine model file provided by the user. Among them, the mechanism model file formula file includes, for example, the formula form, formula parameters and data dimensions involved in the mechanism model, the model constraint information includes, for example, the data dimensions and slope range involved in the mechanism parameters of the mechanism model, and the machine model file can be called from the provided training and reasoning functions, or selected from the default machine model list.

[0131] exist Figure 3 In the example shown, during the model training process, the task processing system 10 receives the training data set file and fuses the mechanism model with the machine model. During the reasoning process, the task processing system 10 receives the test data set file and outputs the prediction result based on the trained mechanism model, and the prediction result includes the prediction result corresponding to each test sample. The task processing system 10 can also analyze the prediction error of the mechanism model provided by the user based on the prediction result, and provide suggestions for adjusting the mechanism parameters of the mechanism model. Specifically, the task processing system 10 can display the mechanism model error statistics and mechanism model parameter recommendation statistics over a period of time on the interface, and provide the actual slope performance of the training mechanism model on the specified test sample.

[0132] It can be seen from the above embodiments that the computing device in the embodiments of the present application can determine the mechanism parameters of the mechanism model by training the fusion model. Since the machine model in the fusion model does not directly affect the task processing results, that is, the machine model is pluggable, the mechanism model in the embodiments of the present application supports the fusion of multiple machine models, thereby improving the prediction accuracy of the mechanism model for unknown tasks.

[0133] Based on the above method embodiment, the embodiment of the present application also provides a task processing device. The task processing device provided by the embodiment of the present application is specifically introduced below.

[0134] See also Figure 6 , Figure 6 A schematic diagram of the structure of a task processing device provided in an embodiment of the present application. Figure 6In the example shown, the task processing device 600 is used to implement the various steps performed by the task processing system 10 in the above embodiments. The task processing device 600 includes an interface unit 601, a training unit 602, a processing unit 603, a data unit 604 and an analysis unit 605.

[0135] Among them, the interface unit 601 is used to obtain the machine model and the mechanism model, the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to perform unknown prediction tasks. The training unit 602 is used to use the machine model as the mechanism parameter of the mechanism model, obtain the fusion model, and train the fusion model based on the training data. The processing unit 603 is used to update the mechanism parameters of the mechanism model based on the machine model corresponding to the training result of the fusion model to obtain the updated mechanism model, the training result of the fusion model is used to indicate the target model parameters of the machine model, and the target model parameters are used to determine the mechanism parameters of the updated mechanism model. The processing unit 603 also performs unknown prediction tasks based on the updated mechanism model.

[0136] In a possible implementation, training data and model constraint information are obtained, and the model constraint information is used to constrain the range of mechanism parameters in the fusion model. The training unit 602 is also used to train the fusion model based on the training data to obtain a trained fusion model, and the mechanism parameters of the trained fusion model meet the model constraint information.

[0137] In a possible implementation, the device further includes a data unit 604, and the data unit 604 is used to perform data augmentation on the training data of the fusion model based on the data augmentation strategy to obtain augmented training data, and the augmented training data is used to train the fusion model.

[0138] In a possible implementation, the data unit 604 is further configured to update the data augmentation strategy based on the deviation between the machine model output corresponding to the augmented training data and the model constraint information, and regenerate the augmented training data based on the updated data augmentation strategy.

[0139] In a possible implementation, the device further includes an analysis unit 605, which is used to calculate an inference deviation based on the inference results of the trained fusion model on the test data and the inference results of the mechanism model on the test data, and to correct model parameters of the trained fusion model based on the inference deviation.

[0140] In a possible implementation, the device further includes an analysis unit 605, which is further used to calculate model parameters corresponding to different test data based on the trained machine model, and the model parameters corresponding to different test data are used to calculate the constraint satisfaction of the model constraint information. The reliability of the model constraint information is evaluated based on the constraint satisfaction.

[0141] It can be understood that the interface unit 601, the training unit 602, the processing unit 603, the data unit 604 and the analysis unit 605 in the task processing device 600 can be used as functional modules and Figure 1 There is a mapping relationship between one or more modules in each node in the task processing system 10, so as to realize the function of each node in the task processing system 10.

[0142] It should be understood that the division of the units in the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And the units in the device can all be implemented in the form of software calling through processing elements; they can also be all implemented in the form of hardware; some units can also be implemented in the form of software calling through processing elements, and some units can be implemented in the form of hardware. For example, each unit can be a separately established processing element, or it can be integrated in a certain chip of the device. In addition, it can also be stored in the memory in the form of a program, and called and executed by a certain processing element of the device. The function of the unit. In addition, all or part of these units can be integrated together, or they can be implemented independently. The processing element described here can also be a processor, which can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each unit above can be implemented by an integrated logic circuit of hardware in the processor element or in the form of software calling through a processing element.

[0143] It is worth noting that, for the above method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the described order of actions. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required for the present application.

[0144] Other reasonable step combinations that can be thought of by those skilled in the art based on the above description also fall within the scope of protection of this application. Secondly, those skilled in the art should also be familiar with the fact that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by this application.

[0145] See also Figure 7 , Figure 7 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. Figure 7As shown, the computing device 700 includes: a processor 701, a memory 702, a communication interface 703 and a bus 704. The processor 701, the memory 702 and the communication interface 703 are coupled via a bus (not marked in the figure). The memory 702 stores instructions. When the execution instructions in the memory 702 are executed, the computing device 700 executes the method executed by the computing device in the above method embodiment.

[0146] The computing device 700 may be one or more integrated circuits configured to implement the above method, such as one or more application specific integrated circuits (ASIC), or one or more digital signal processors (DSP), or one or more field programmable gate arrays (FPGA), or a combination of at least two of these integrated circuit forms. For another example, when the unit in the device can be implemented in the form of a processing element scheduler, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call a program. For another example, these units can be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0147] The processor 701 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. A general-purpose processor may be a microprocessor or any conventional processor.

[0148] The memory 702 may be a volatile memory or a nonvolatile memory, or may include both volatile and nonvolatile memories. Among them, the nonvolatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0149] The memory 702 stores executable program codes, and the processor 701 executes the executable program codes to respectively implement the functions of the aforementioned units or modules, thereby implementing the aforementioned task processing method. That is, the memory 702 stores instructions for executing the aforementioned task processing method.

[0150] The communication interface 703 uses a transceiver module such as, but not limited to, a network interface card or a transceiver to implement communication between the computing device 700 and other devices or a communication network.

[0151] In addition to the data bus, the bus 704 may also include a power bus, a control bus, a status signal bus, etc. The bus may be a peripheral component interconnect express (PCIe) bus, or an extended industry standard architecture (EISA) bus, a unified bus (Ubus or UB), a compute express link (CXL), a cache coherent interconnect for accelerators (CCIX), etc. The bus may be divided into an address bus, a data bus, a control bus, etc.

[0152] See also Figure 8 , Figure 8 A schematic diagram of a computing device cluster provided in an embodiment of the present application. Figure 8 As shown, the computing device cluster 800 includes at least one computing device 700 .

[0153] like Figure 8 As shown, the computing device cluster 800 includes at least one computing device 700. The memory 702 in one or more computing devices 700 in the computing device cluster 800 may store the same instructions for executing the above-mentioned task processing method.

[0154] In some possible implementations, the memory 702 of one or more computing devices 700 in the computing device cluster 800 may also store partial instructions for executing the above-mentioned task processing method. In other words, the combination of one or more computing devices 700 may jointly execute instructions for executing the above-mentioned task processing method.

[0155] It should be noted that the memory 702 in different computing devices 700 in the computing device cluster 800 may store different instructions, which are respectively used to execute part of the functions of the above-mentioned task processing device. That is, the instructions stored in the memory 702 in different computing devices 700 may implement the functions of one or more modules in the interface unit, training unit, processing unit, data unit and analysis unit.

[0156] In some possible implementations, one or more computing devices 700 in the computing device cluster 800 may be connected via a network, which may be a wide area network or a local area network.

[0157] See also Fig. 9 , Fig. 9A schematic diagram of computer devices in a computer cluster connected via a network provided in an embodiment of the present application. Fig. 9 As shown, two computing devices 700A and 700B are connected via a network. Specifically, they are connected to the network via a communication interface in each computing device.

[0158] In a possible implementation, the memory in the computing device 700A stores instructions for executing the functions of the interface unit, the training unit, and the processing unit, while the memory in the computing device 700B stores instructions for executing the functions of the data unit and the analysis unit.

[0159] It should be understood that Fig. 9 The functions of the computing device 700A shown in FIG. 7 may also be completed by multiple computing devices. Similarly, the functions of the computing device 700B may also be completed by multiple computing devices.

[0160] In another embodiment of the present application, a computer-readable storage medium is provided, in which computer-executable instructions are stored. When the processor of the device executes the computer-executable instructions, the device executes the method executed by the task processing system in the above method embodiment.

[0161] In another embodiment of the present application, a computer program product is provided, the computer program product includes computer executable instructions, the computer executable instructions are stored in a computer readable storage medium. When the processor of the device executes the computer executable instructions, the device executes the method executed by the task processing system in the above method embodiment.

[0162] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0163] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0164] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0165] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0166] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.

Claims

1. A task processing method, characterized in that: Applied to a cloud platform, the cloud platform is used to manage the infrastructure for providing cloud services, the infrastructure is used to carry the machine model and mechanism model deployed by the user, the method includes: Acquire the machine model and the mechanism model, wherein the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to perform an unknown prediction task; Using the machine model as a mechanism parameter of the mechanism model to obtain a fusion model, and training the fusion model based on training data; Based on the machine model corresponding to the training result of the fusion model, the mechanism parameters of the mechanism model are updated to obtain an updated mechanism model, wherein the training result of the fusion model is used to indicate the target model parameters of the machine model, and the target model parameters are used to determine the mechanism parameters of the updated mechanism model; Perform unknown prediction tasks based on the updated mechanism model.

2. The method according to claim 1, characterized in that The method further comprises: Acquiring training data and model constraint information, wherein the model constraint information is used to constrain the range of the mechanism parameters in the fusion model; The training of the fusion model based on the training data includes: The fusion model is trained based on the training data to obtain the trained fusion model, and the mechanism parameters of the trained fusion model satisfy the model constraint information.

3. The method according to claim 2, characterized in that The method further comprises: The training data of the fusion model is augmented based on a data augmentation strategy to obtain augmented training data, and the augmented training data is used to train the fusion model.

4. The method according to claim 3, characterized in that The method further comprises: Update the data augmentation strategy based on a deviation between the machine model output corresponding to the augmented training data and the model constraint information; Regenerate augmented training data based on the updated data augmentation strategy.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises: Calculating the inference deviation based on the inference result of the trained fusion model on the test data and the inference result of the mechanism model on the test data; Model parameters of the trained fusion model are corrected based on the inference deviation.

6. The method according to any one of claims 1 to 5, characterized in that The method further comprises: Calculating model parameters corresponding to different test data based on the machine model corresponding to the training result of the fusion model, wherein the model parameters corresponding to the different test data are used to count the constraint satisfaction of the model constraint information; The reliability of the model constraint information is evaluated based on the constraint satisfaction.

7. A task processing device, characterized in that: include: An interface unit, used to obtain the machine model and the mechanism model, wherein the machine model is used to predict the mechanism parameters of the mechanism model, and the mechanism model is used to perform an unknown prediction task; A training unit, used for taking the machine model as a mechanism parameter of the mechanism model to obtain a fusion model, and training the fusion model based on training data; A processing unit, configured to update the mechanism parameters of the mechanism model based on the machine model corresponding to the training result of the fusion model to obtain an updated mechanism model, wherein the training result of the fusion model is used to indicate the target model parameters of the machine model, and the target model parameters are used to determine the mechanism parameters of the updated mechanism model; The processing unit also performs unknown prediction tasks based on the updated mechanism model.

8. The device according to claim 7, characterized in that The interface unit is also used for: Acquiring training data and model constraint information, wherein the model constraint information is used to constrain the range of the mechanism parameters in the fusion model; The training unit is also used to train the fusion model based on the training data to obtain the trained fusion model, and the mechanism parameters of the trained fusion model satisfy the model constraint information.

9. The device according to claim 8, characterized in that The apparatus further comprises a data unit, wherein the data unit is configured to: The training data of the fusion model is augmented based on a data augmentation strategy to obtain augmented training data, and the augmented training data is used to train the fusion model.

10. The device according to claim 9, characterized in that The data unit is also used for: Update the data augmentation strategy based on a deviation between the machine model output corresponding to the augmented training data and the model constraint information; Regenerate augmented training data based on the updated data augmentation strategy.

11. The device according to any one of claims 7 to 10, characterized in that The device further comprises an analysis unit, which is used for: Calculating the inference deviation based on the inference result of the trained fusion model on the test data and the inference result of the mechanism model on the test data; Model parameters of the trained fusion model are corrected based on the inference deviation.

12. The device according to any one of claims 7 to 11, characterized in that The device further comprises an analysis unit, which is further configured to: Calculating model parameters corresponding to different test data based on the machine model corresponding to the training result of the fusion model, wherein the model parameters corresponding to the different test data are used to count the constraint satisfaction of the model constraint information; The reliability of the model constraint information is evaluated based on the constraint satisfaction.

13. A computing device, characterized in that: The electronic device comprises a processor coupled to a memory, wherein the processor is used to store instructions. When the instructions are executed by the processor, the electronic device executes the method according to any one of claims 1 to 6.

14. A computing device cluster, characterized in that: The system comprises at least one computing device, wherein the computing device comprises a processor, wherein the processor is coupled to a memory, and the processor is used to store instructions. When the instructions are executed by the processor, the computing device cluster executes the method according to any one of claims 1 to 6.

15. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed, the computer is caused to perform the method according to any one of claims 1 to 6.

16. A computer program product, comprising instructions, characterized in that: When the instructions are executed, the computer implements the method according to any one of claims 1 to 6.

Images