Methods for determining the amount of training data and rewards for ensemble learning, as well as related equipment.

Abstract

This application discloses a method for determining the training data volume and reward in ensemble learning, an ensemble server, and a computer-readable storage medium. This method determines the training data volume and reward for each terminal device in ensemble learning while maximizing the accuracy of the ensemble learning model and minimizing the reward paid to each terminal device. It also aims to adaptively decide and incentivize terminal devices with different computational and communication costs to participate in ensemble learning. The method includes: determining the initial training data volume and initial reward for each terminal device participating in ensemble learning; using an incentive algorithm, iterating N rounds of decision-making on the initial training data volume and initial reward for each terminal device according to the decision order, obtaining the decision result for each round of iteration; and using the training data volume and reward corresponding to each terminal device whose decision result satisfies a preset convergence condition as the target training data volume and target reward for each terminal device.

Description

Technical Field

[0001] This application relates to the field of determining the amount of training data and rewards for ensemble learning, and more specifically, to a method for determining the amount of training data and rewards for ensemble learning, an ensemble server, and a computer-readable storage medium. Background Technology

[0002] Ensemble learning is a novel swarm intelligence technology for artificial intelligence and machine learning, widely applied in finance, healthcare, transportation, and other fields. Typical ensemble learning trains multiple machine learning models and combines their predictions to determine the final prediction, effectively improving the performance and robustness of machine learning models. However, training multiple neural network models typically consumes significant time and computational resources, especially when the training dataset and models are large, leading to excessively long training times for ensemble learning models. An effective solution is distributed ensemble learning, which distributes the training tasks of individual models across different computing devices. By training multiple models in parallel, it efficiently utilizes the computational resources of distributed computing devices to achieve rapid model training.

[0003] Existing methods for determining the training data volume and reward in ensemble learning involve an ensemble learning server (ensemble server) randomly determining the amount of training data for each participating terminal device and setting a uniform reward per unit of resource for each terminal device. For each distributed terminal device, based on its training data volume and the uniform reward per unit of resource, the final reward to be paid to the terminal device is determined, enabling each terminal device to train the neural network model in parallel using its given training data volume. Furthermore, since the performance of a neural network model is closely related to the amount of training data—generally, a larger amount of training data results in better model performance, but also requires higher rewards to be paid to the distributed terminal devices—the ensemble server often provides a relatively large amount of training data to each terminal device.

[0004] However, different distributed terminal devices typically have different computing and communication costs. Simply calculating the reward for different terminal devices (different computing device owners) using a uniform unit resource reward may lead to device owners with higher computing and communication costs being unwilling to participate in ensemble learning model training. Therefore, without considering the rationality of the reward for each terminal device, it is impossible to maximize the accuracy of the ensemble learning model while minimizing the reward paid to each terminal device. It is also impossible to make adaptive decisions and incentivize terminal devices with different computing and communication costs to participate in ensemble learning. Summary of the Invention

[0005] This application provides a method for determining the amount of training data and rewards for ensemble learning, an ensemble server, and a computer-readable storage medium. It can determine the amount of training data and rewards for each terminal device in ensemble learning while maximizing the accuracy of the ensemble learning model and minimizing the rewards for each terminal device. It can also adaptively make decisions and incentivize terminal devices with different computing and communication costs to participate in ensemble learning.

[0006] In a first aspect, embodiments of this application provide a method for determining the amount of training data and rewards for ensemble learning, including:

[0007] Determine the initial amount of training data and the initial reward for each terminal device participating in ensemble learning;

[0008] Using an incentive algorithm, the initial training data volume and initial reward for each terminal device are iteratively decided in N rounds according to the decision order, yielding the decision result for each iteration. The incentive algorithm is determined based on the computational and communication costs of each terminal device and is used to decide the training data volume and reward for each terminal device, aiming to maximize the accuracy of the ensemble learning model and minimize the rewards for each terminal device after N iterations. The decision result for each iteration includes the training data volume and reward for each terminal device; N is an integer greater than or equal to 1.

[0009] The amount of training data and the reward corresponding to each terminal device whose decision results meet the preset convergence conditions are taken as the target training data and target reward corresponding to each terminal device. The target training data for each terminal device is used to train the local model corresponding to each terminal device. The target reward for each terminal device is the reward obtained by each terminal device training the local model based on the corresponding target training data. The trained local model for each terminal device is obtained, and the trained local models for each terminal device are integrated to obtain the ensemble learning model. The data to be predicted is predicted based on the ensemble learning model.

[0010] Secondly, embodiments of this application provide an integrated server, including:

[0011] The determination unit is used to determine the initial amount of training data and the initial reward for each terminal device participating in ensemble learning;

[0012] A decision-making unit is used to perform N rounds of iterative decision-making on the initial training data volume and initial reward of each terminal device according to the decision-making order using an incentive algorithm, to obtain the decision result corresponding to each round of iteration; wherein, the incentive algorithm is determined based on the computing and communication costs of each terminal device, and is used to decide the training data volume and reward corresponding to each terminal device, so as to maximize the accuracy of the ensemble learning model and minimize the reward of each terminal device after N rounds of iteration; the decision result corresponding to each iteration includes the training data volume and reward corresponding to each terminal device; N is an integer greater than or equal to 1;

[0013] The determining unit is further configured to use the amount of training data and the reward corresponding to each terminal device whose decision result satisfies the preset convergence condition as the target amount of training data and the target reward corresponding to each terminal device; wherein, the target amount of training data corresponding to each terminal device is used to train the local model corresponding to each terminal device; and the target reward corresponding to each terminal device is the reward obtained by each terminal device training the local model based on the corresponding target amount of training data; to obtain the trained local model corresponding to each terminal device, and to integrate the trained local models corresponding to each terminal device to obtain the ensemble learning model, and to predict the data to be predicted based on the ensemble learning model.

[0014] Thirdly, embodiments of this application provide an integrated server, including:

[0015] Central processing unit, memory, input / output interfaces, wired or wireless network interfaces, and power supply;

[0016] The memory is either a short-term storage memory or a persistent storage memory;

[0017] The central processing unit is configured to communicate with the memory and execute instructions in the memory to perform the aforementioned method for determining the amount of training data and reward for ensemble learning.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the aforementioned method for determining the amount of training data and reward for ensemble learning.

[0019] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute the aforementioned method for determining the amount of training data and reward for ensemble learning.

[0020] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: The initial training data volume and initial reward for each terminal device participating in ensemble learning can be determined. Using an incentive algorithm, the initial training data volume and initial reward for each terminal device are iteratively decided in N rounds according to the decision order, obtaining the decision result corresponding to each round of iteration. The incentive algorithm is determined based on the computational and communication costs of each terminal device and is used to decide the training data volume and reward corresponding to each terminal device, aiming to maximize the accuracy of the ensemble learning model and minimize the rewards for each terminal device after N rounds of iteration. The training data volume and reward corresponding to each terminal device whose decision result satisfies the preset convergence condition are taken as the target training data volume and target reward for each terminal device. The incentive mechanism (incentive algorithm) on the ensemble learning server can be used to simultaneously consider the accuracy of the ensemble learning model and the rationality of the rewards for each terminal device, maximizing the accuracy of the ensemble learning model while minimizing the rewards paid to each terminal device. This allows for adaptive decision-making and incentives for terminal devices with different computational and communication costs to participate in ensemble learning. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the architecture of a system for determining the amount of training data and rewards for ensemble learning, as disclosed in an embodiment of this application.

[0022] Figure 2 This is a flowchart illustrating a method for determining the amount of training data and rewards in ensemble learning, as disclosed in an embodiment of this application.

[0023] Figure 3 This is a schematic diagram of the overall process of a distributed ensemble learning incentive mechanism disclosed in an embodiment of this application;

[0024] Figure 4 This is a flowchart illustrating another method for determining the amount of training data and the reward in ensemble learning, as disclosed in an embodiment of this application.

[0025] Figure 5 This is a schematic diagram of the structure of an integrated server disclosed in an embodiment of this application;

[0026] Figure 6 This is a schematic diagram of another integrated server structure disclosed in an embodiment of this application;

[0027] Figure 7 This is a schematic diagram of the structure of another integrated server disclosed in an embodiment of this application. Detailed Implementation

[0028] This application provides a method for determining the amount of training data and rewards for ensemble learning, an ensemble server, and a computer-readable storage medium. These methods are used to determine the amount of training data and rewards for each terminal device in ensemble learning while maximizing the accuracy of the ensemble learning model and minimizing the rewards for each terminal device. The methods also enable adaptive decision-making and incentives for terminal devices with different computational and communication costs to participate in ensemble learning.

[0029] Please see Figure 1 The architecture of the system for determining the amount of training data and the reward for ensemble learning in this application embodiment includes:

[0030] An integration server 101 and a terminal device 102 are configured. When determining the training data volume and reward for ensemble learning, the integration server 101 can connect to the terminal device 102. The integration server 101 can determine the initial training data volume and initial reward for each terminal device participating in the ensemble learning. Using an incentive algorithm, it iterates the initial training data volume and initial reward for each terminal device in N rounds according to the decision order, obtaining the decision result for each round of iteration. The training data volume and reward corresponding to each terminal device whose decision result satisfies the preset convergence condition are taken as the target training data volume and target reward for each terminal device. The target training data volume and target reward for each terminal device are then sent to the corresponding terminal device 102, enabling the terminal device 102 to train a local model based on the target training data volume and obtain the target reward.

[0031] based on Figure 1 For the training data volume and reward determination system of the ensemble learning shown, please refer to [link / reference]. Figure 2 , Figure 2 This is a flowchart illustrating a method for determining the amount of training data and rewards in ensemble learning, as disclosed in an embodiment of this application. The method includes:

[0032] 201. Determine the initial amount of training data and the initial reward for each terminal device participating in ensemble learning.

[0033] In this embodiment, when determining the amount of training data and the reward for ensemble learning, the initial amount of training data and the initial reward for each terminal device participating in ensemble learning can be determined.

[0034] 202. Using an incentive algorithm, the initial training data volume and initial reward for each terminal device are iterated N times according to the decision order to obtain the decision result corresponding to each iteration. The incentive algorithm is determined based on the computation and communication cost of each terminal device and is used to decide the training data volume and reward corresponding to each terminal device so that the accuracy of the ensemble learning model reaches the maximum value and the reward of each terminal device reaches the minimum value after N iterations. The decision result corresponding to the iteration includes the training data volume and reward corresponding to each terminal device. N is an integer greater than or equal to 1.

[0035] After determining the initial training data volume and initial reward for each terminal device participating in ensemble learning, an incentive algorithm can be used to iterate the decision on the initial training data volume and initial reward for each terminal device in N rounds according to the decision order, obtaining the decision result corresponding to each round of iteration. The incentive algorithm is used to decide the training data volume and reward corresponding to each terminal device so that the accuracy of the ensemble learning model reaches the maximum value and the reward of each terminal device reaches the minimum value after N rounds of iteration. The decision result corresponding to the iteration includes the training data volume and reward corresponding to each terminal device; N is an integer greater than or equal to 1.

[0036] Understandably, the decision-making order can be random, or it can be based on the total cost of each terminal device from low to high, or it can be any other reasonable order; no specific restrictions are imposed here.

[0037] 203. The amount of training data and the reward corresponding to each terminal device whose decision results meet the preset convergence conditions are taken as the target training data and target reward corresponding to each terminal device; wherein, the target training data corresponding to each terminal device is used to train the local model corresponding to each terminal device; and the target reward corresponding to each terminal device is the reward obtained by each terminal device training the local model based on the corresponding target training data; to obtain the trained local model corresponding to each terminal device, and to integrate the trained local models corresponding to each terminal device to obtain the ensemble learning model, and to predict the data to be predicted based on the ensemble learning model.

[0038] Using an incentive algorithm, N rounds of iterative decision-making are performed on the initial training data volume and initial reward of each terminal device according to the decision order. After obtaining the decision result corresponding to each round of iteration, the training data volume and reward corresponding to each terminal device whose decision result satisfies the preset convergence condition can be used as the target training data volume and target reward corresponding to each terminal device. The target training data volume corresponding to each terminal device is used to train the local model corresponding to each terminal device. The target reward corresponding to each terminal device is the reward obtained by each terminal device training its local model based on the corresponding target training data volume. The trained local models corresponding to each terminal device are obtained, and the trained local models corresponding to each terminal device are integrated to obtain an ensemble learning model. The data to be predicted is then predicted based on the ensemble learning model.

[0039] In this embodiment, the initial training data volume and initial reward for each terminal device participating in ensemble learning can be determined. Using an incentive algorithm, the initial training data volume and initial reward for each terminal device are iteratively decided in N rounds according to the decision order, yielding the decision result for each iteration. The incentive algorithm is determined based on the computational and communication costs of each terminal device and is used to decide the training data volume and reward for each terminal device, aiming to maximize the accuracy of the ensemble learning model and minimize the rewards for each terminal device after N iterations. The training data volume and reward for each terminal device whose decision result satisfies a preset convergence condition are taken as the target training data volume and target reward for each terminal device. By utilizing the incentive mechanism (incentive algorithm) on the ensemble learning server, considering both the accuracy of the ensemble learning model and the reasonableness of the rewards for each terminal device, the accuracy of the ensemble learning model can be maximized while minimizing the rewards paid to each terminal device. This allows for adaptive decision-making and incentivizes terminal devices with different computational and communication costs to participate in ensemble learning.

[0040] In this embodiment, an incentive algorithm is used to perform N rounds of iterative decision-making on the initial training data volume and initial reward of each terminal device according to the decision-making order. There are various methods to obtain the decision result corresponding to each round of iteration. Figure 2 The methods for determining the amount of training data and rewards in ensemble learning are shown below, and one of these methods is described below.

[0041] In this embodiment, when determining the amount of training data and the reward for ensemble learning, the initial amount of training data and the initial reward for each terminal device participating in ensemble learning can be determined.

[0042] It's important to understand that determining the amount of training data and rewards for ensemble learning can be the same scenario as determining the amount of training data and rewards for distributed ensemble learning. For details, please refer to [link / reference needed]. Figure 3 , Figure 3 This is a schematic diagram of the overall process of a distributed ensemble learning incentive mechanism disclosed in an embodiment of this application. Figure 3 As can be seen, computing devices 1 to N are each terminal device participating in ensemble learning. The ensemble server can obtain the local computing and communication resource costs (computing cost and communication cost) uploaded by computing devices 1 to N, decide the training data size (training data volume) and reward of each computing device, and send the corresponding training data of each computing device to the corresponding computing device so that each computing device can train a local model based on the obtained training data and upload the trained local model to the ensemble server. The ensemble server can perform ensemble processing on the trained local models sent by each computing device to obtain an ensemble learning model, and predict the data to be predicted based on the ensemble learning model.

[0043] Specifically, for example, each distributed computing device provider i reports its local unit computing resource cost α. i and unit communication resource cost β i Secondly, the ensemble learning server (ensemble server) runs an incentive algorithm to determine the participation of each distributed device provider in the ensemble learning variable d. i (d i =1 indicates that distributed computing device provider i participates in ensemble learning; otherwise, it does not participate. Simultaneously, the ensemble learning server also determines the training data size D for the distributed computing devices (terminal devices) participating in ensemble learning. i Reward R per unit of training data i Then, the ensemble learning server samples the data using sampling with replacement, with each user i sampling a data set of size D. i The server sends the sampled training data to the distributed computing devices participating in the ensemble learning. Using the received training data, each distributed computing device trains a local neural network model until convergence and then sends the trained model to the ensemble learning server. Finally, using the models trained by all the devices, the ensemble learning server pays a reward R. i Provide the corresponding computing device provider i. Utilizing all collected models, the ensemble learning server can achieve high-accuracy predictions on the test data according to a majority consensus rule.

[0044] For details, please refer to Figure 4 , Figure 4 This is a flowchart illustrating another method for determining the amount of training data and the reward in ensemble learning, as disclosed in an embodiment of this application. Figure 4 It is known that the training data size (training data volume) and unit resource reward (reward) of all computing device providers (terminal devices) can be randomly initialized. It is also understood that they can be initialized to any arbitrarily determined value, but no specific limitation is made here.

[0045] After determining the initial training data volume and initial reward for each terminal device participating in ensemble learning, an incentive algorithm can be used to iterate the initial training data volume and initial reward for each terminal device in N rounds according to the decision order, obtaining the decision result corresponding to each round of iteration. The incentive algorithm is determined based on the computation and communication costs of each terminal device and is used to decide the training data volume and reward corresponding to each terminal device, so that the accuracy of the ensemble learning model reaches the maximum value and the reward of each terminal device reaches the minimum value after N rounds of iteration. The decision result corresponding to the iteration includes the training data volume and reward corresponding to each terminal device; N is an integer greater than or equal to 1.

[0046] The process involves using an incentive algorithm to iterate N rounds of decisions regarding the initial training data volume and initial reward for each terminal device, following a decision-making order. Before obtaining the decision result for each iteration, the total cost for each terminal device is determined, including computational and communication costs. The total costs of each terminal device are then sorted in ascending order, and this ascending cost order is used to define the decision order. It's understood that a random or other reasonable order could also be used; no specific limitation is made here.

[0047] For details, please continue reading. Figure 4 ,Depend on Figure 4 It can be seen that the computing and communication resource costs of all devices and (the total cost of each terminal device) can be sorted in ascending order, and this order can be used as the decision order.

[0048] It's worth noting that obtaining the total cost for each terminal device and then sorting the total costs of all terminal devices in ascending order, using this ascending cost order as the decision order, can improve the efficiency of determining the amount of training data and rewards for ensemble learning, while ensuring that the accuracy of the ensemble learning model reaches its maximum and the rewards for each terminal device reach their minimum after N iterations. It's also worth mentioning that the incentive algorithm can be used to determine the amount of training data and rewards for each terminal device to achieve other desired goals, such as maximizing the accuracy of the ensemble learning model and minimizing the rewards for each terminal device after N iterations; these are not limited here.

[0049] The method for obtaining the decision result for each iteration by performing N rounds of iterative decision-making on the initial training data volume and initial reward of each terminal device using an incentive algorithm in accordance with the decision-making order can be as follows: For the first iteration, determine the target incentive algorithm corresponding to the first iteration based on the initial training data volume and initial reward of each terminal device; and use the target incentive algorithm corresponding to the first iteration to perform the first iteration on the initial training data volume and initial reward of each terminal device in accordance with the decision-making order to obtain the decision result corresponding to the first iteration; the decision result corresponding to the first iteration includes the training data volume and reward of each terminal device after the first iteration; For the nth iteration, determine the target incentive algorithm corresponding to the nth iteration based on the training data volume and reward of each terminal device after the (n-1)th iteration; and use the target incentive algorithm corresponding to the nth iteration to perform the nth iteration on the training data volume and reward of each terminal device after the (n-1)th iteration in accordance with the decision-making order to obtain the decision result corresponding to the nth iteration; the decision result corresponding to the nth iteration includes the training data volume and reward of each terminal device after the nth iteration; where 2≤n≤N.

[0050] In the first iteration, the target incentive algorithm corresponding to the first iteration is determined based on the initial training data volume and initial reward of each terminal device. Using this target incentive algorithm, the first iteration is performed on each terminal device in turn according to the decision order, taking into account the initial training data volume and initial reward, to obtain the decision result for the first iteration. This can be achieved by: for the terminal device making the first decision in the first iteration, determining the target incentive algorithm corresponding to that terminal device based on its initial training data volume and initial reward; and using this target incentive algorithm, performing the first iteration on that terminal device to obtain the decision result for the first decision in the first iteration. For the terminal device making the m-th decision in the first iteration, based on the amount of training data and rewards corresponding to the first iteration for each of the terminal devices making the first to the (m-1)-th decisions, and the amount of initial training data and initial rewards corresponding to each of the terminal devices making the m-th to the M-th decisions, the target incentive algorithm corresponding to the terminal device making the m-th decision in the first iteration is determined. Then, using the target incentive algorithm corresponding to the terminal device making the m-th decision in the first iteration, the terminal device making the m-th decision makes the decision in the first iteration, and the decision result of the terminal device making the m-th decision in the first iteration is obtained. Where 2≤m≤M, and the target incentive algorithm corresponding to the first iteration includes the target incentive algorithms corresponding to each of the terminal devices making the first to the M-th decisions in the first iteration.

[0051] Specifically, for the nth iteration, the target incentive algorithm corresponding to the nth iteration is determined based on the amount of training data and reward after the (n-1)th iteration for each terminal device. Using the target incentive algorithm corresponding to the nth iteration, the nth iteration is performed on the amount of training data and reward after the (n-1)th iteration for each terminal device according to the decision order to obtain the decision result corresponding to the nth iteration. This can be achieved by: for the terminal device making the first decision in the nth iteration, determining the target incentive algorithm corresponding to the terminal device making the first decision in the nth iteration based on the amount of training data and reward after the (n-1)th iteration for each terminal device; and using the target incentive algorithm corresponding to the terminal device making the first decision in the nth iteration, performing the nth iteration on the terminal device making the first decision to obtain the terminal device making the first decision in the nth iteration. The decision result is obtained; for the terminal device making the m-th decision in the n-th iteration, based on the amount of training data and reward corresponding to the n-th iteration for each of the terminal devices making the 1st to the (m-1)th decisions, and the amount of training data and reward corresponding to the (n-1)th iteration for each of the terminal devices making the m-th decisions, the target incentive algorithm corresponding to the terminal device making the m-th decision in the n-th iteration is determined; and using the target incentive algorithm corresponding to the terminal device making the m-th decision in the n-th iteration, the decision of the terminal device making the m-th decision in the n-th iteration is performed, and the decision result of the terminal device making the m-th decision in the n-th iteration is obtained; where 2≤m≤M, and the target incentive algorithm corresponding to the n-th iteration includes the target incentive algorithms corresponding to each of the terminal devices making the 1st to the Mth decisions in the n-th iteration.

[0052] In this process, an incentive algorithm is used to make N rounds of iterative decisions on the initial training data volume and initial reward of each terminal device according to the decision order. Before obtaining the decision result corresponding to each round of iteration, the incentive algorithm can be determined. Specifically, the method to determine the incentive algorithm can be to first obtain the target accuracy function of each local model of the ensemble learning, then obtain the sum function of the rewards of each terminal device of the ensemble learning, and finally take the maximum value of the difference between the target accuracy function and the sum function of rewards as the incentive algorithm.

[0053] One method for obtaining the target accuracy function of each local model in ensemble learning is as follows: First, determine the total number of terminal devices participating in ensemble learning, the total amount of training data, and the number of local models with prediction errors corresponding to each training data; then, obtain the diversity function among the local models based on the total number of terminal devices participating in ensemble learning, the total amount of training data, and the number of models with prediction errors corresponding to each training data; next, determine the average accuracy of each local model; and then, determine the accuracy function of each local model based on the total number of terminal devices participating in ensemble learning and the average accuracy of each local model; finally, use the sum of the diversity function among the local models and the accuracy function of each local model as the target accuracy function of each local model in ensemble learning.

[0054] Before using the maximum difference between the target accuracy function and the sum of rewards function as the incentive algorithm, the weight value of the target accuracy function can be determined. The method of using the maximum difference between the target accuracy function and the sum of rewards function as the incentive algorithm can be to first multiply the weight value with the target accuracy function to obtain a product, and then use the maximum difference between the product and the sum of rewards function as the incentive algorithm.

[0055] It is worth mentioning that the weight values ​​of the target accuracy function can be determined, and the weight values ​​can be multiplied by the target accuracy function to obtain the product. The maximum value of the difference between the product and the sum of the rewards function is used as the activation algorithm. Based on the accuracy of the ensemble learning model and the importance between the rewards of each terminal device, the amount of training data and the rewards of ensemble learning can be determined, which improves the adjustability of the decision objective.

[0056] Using an incentive algorithm, N rounds of iterative decision-making are performed on the initial training data volume and initial reward of each terminal device according to the decision order. After obtaining the decision result corresponding to each round of iteration, the training data volume and reward corresponding to each terminal device whose decision result satisfies the preset convergence condition can be used as the target training data volume and target reward corresponding to each terminal device. The target training data volume corresponding to each terminal device is used to train the local model corresponding to each terminal device. The target reward corresponding to each terminal device is the reward obtained by each terminal device training its local model based on the corresponding target training data volume. The trained local models corresponding to each terminal device are obtained, and the trained local models corresponding to each terminal device are integrated to obtain an ensemble learning model. The data to be predicted is then predicted based on the ensemble learning model.

[0057] Specifically, one method to determine if a decision result meets a preset convergence condition is to check whether the decision result of the current iteration is consistent with the decision result of the previous iteration. If they are consistent, the decision result meets the preset convergence condition; if they are inconsistent, the decision result does not meet the preset convergence condition. Another method is to check whether the difference between the decision result of the current iteration and the decision result of the previous iteration is less than or equal to a preset difference threshold. If the difference is less than or equal to the threshold, the decision result meets the preset convergence condition; if it is greater, the decision result does not meet the preset convergence condition. It is understood that other reasonable methods for determining if a decision result meets the preset convergence condition may also be used, and these are not limited here.

[0058] For details, please continue reading. Figure 4 ,Depend on Figure 4 It can be seen that the amount of training data Di for computing device i can be determined, as well as the unit resource reward Ri for computing device i. If a device fails to complete its decision in this round, it is moved to the next sorted device i = i+1, and the amount of training data Di for computing device i is determined. If all devices complete their decisions in this round and the decision results of all devices converge, the process ends. If all devices complete their decisions in this round but the decision results of all devices fail to converge, the next iteration continues until all devices complete their decisions in this round and the decision results of all devices converge.

[0059] To facilitate understanding of the embodiments of this application, a specific example is given below:

[0060] Specifically, the target accuracy function for each local model in ensemble learning can be expressed as Equation 1:

[0061]

[0062] In formula 1, N P D represents the number of computing devices (terminal devices) participating in ensemble learning. T Represents the total amount of training data. The value represents the average accuracy of all local models, d represents the data sample (a certain training data), and l represents the average accuracy of all local models. d This represents the number of local models that made incorrect predictions.

[0063] It's important to understand that R and D represent the sum of all rewards and the total amount of training data, respectively. Specifically, R represents the sum of rewards for all terminal devices, and D represents the total amount of training data. Since the reward per unit of resource for computing device provider i is R... i =(α i +β i )D iSince R and D are linearly related, the target accuracy function can be expressed as F(R,D). It's worth noting that the local models for each terminal device involve diverse features and each has its own characteristics. For example, the local models for each terminal device might be face recognition models 1, 2, and 3. These three face recognition models are all used to recognize faces, and each has its own characteristics in preprocessing, feature point extraction, and classification. Therefore, training each face recognition model and integrating the trained face recognition models results in an ensemble learning model (face recognition model) with high overall diversity and accuracy. Thus, the first term on the right-hand side of Equation 1 represents the diversity among the individual local models, and the second term represents the accuracy of all local models. It's also worth mentioning that the final performance of ensemble learning is not only related to the accuracy of each model but also depends on the differences and diversity between different models. Given sufficient performance of individual models, higher complexity between models leads to better performance of the final ensemble learning model. Therefore, analyzing the coupling relationship between the performance of individual models and the diversity between models in ensemble learning, fully considering the differences in resource costs among owners of different computing devices, and designing a distributed ensemble learning incentive mechanism to maximize the overall performance of ensemble learning and minimize the total incentive cost are extremely important. It is understandable that the target accuracy function of ensemble learning can be positively correlated with Equation 1. In addition to the expression in Equation 1, it can also be other reasonable formulas, which are not limited here.

[0064] The activation algorithm can be expressed as Formula 2:

[0065]

[0066] Formula 2 represents the overall objective of the ensemble server (ensemble learning server) as maximizing the accuracy of the ensemble learning model and minimizing the rewards for each terminal device; and γ is the weight coefficient.

[0067] The target incentive algorithm for the terminal device in the m-th decision of the n-th iteration can be expressed as Equation 3:

[0068]

[0069] It is important to understand that Formula 3 is derived from Formula 2; please refer to the following documentation. Figure 4 To maximize the accuracy of the ensemble learning model and minimize the rewards for each terminal device, the ensemble learning server can first randomly initialize the training data size D of all computing device providers. i and unit resource return R i And the computing and communication resource costs of all terminal devices (users) (α) i +β iThe sums of these data are sorted in ascending order. Then, based on the sorting results and given the training data sizes and unit resource rewards of other computing device providers (terminal devices), the training data size D for the current computing device provider i (terminal device i) is first determined. i That is, to solve Formula 3 through a linear search. Determine D. i Then, the reward per unit of resource for computing device provider i can be determined as Ri = (αi + βi) i )D i It is important to understand that if D... i =0 indicates that the computing device (terminal device) does not participate in ensemble learning. The above process is iterated until the training data size (training data volume) and unit resource reward (reward) of all terminal devices (users) no longer change.

[0070] It is also important to understand that, in each decision made by the terminal device, the target incentive algorithm (Formula 3) corresponding to that decision in that iteration can be determined, since F(R,D) is relative to R. i and D i All are concave functions, and the computational and communication resource costs of the terminal device (α) i +β i Since the value is fixed, the target incentive algorithm corresponding to the decision of the terminal device in this round of iteration has a maximum value corresponding to the amount of training data, so that the accuracy of the ensemble learning model reaches the maximum value and the reward of each terminal device reaches the minimum value after N rounds of iteration. Therefore, it is of great significance to determine the target incentive algorithm corresponding to the terminal device for each decision in each round of iteration.

[0071] It is worth mentioning that, according to the total cost of each terminal device, different model structures can be allocated to different terminal devices (users). For example, terminal devices with higher total costs can be allocated models with simpler structures, so that terminal devices with higher total costs can use models with simpler structures as local models and train them; terminal devices with lower total costs can be allocated models with more complex structures, so that terminal devices with lower total costs can use models with more complex structures as local models and train them. This further improves the efficiency of determining the amount of training data and the reward for ensemble learning while maximizing the accuracy of the ensemble learning model and minimizing the reward for each terminal device, as well as the accuracy and efficiency of the final ensemble learning model.

[0072] It is understandable that, besides the method described above which uses an incentive algorithm to iterate N rounds of decisions based on the initial training data and initial reward of each terminal device according to the decision order, and obtains the decision result corresponding to each round of iteration; besides the method described above which, for the first round of iteration, the target incentive algorithm corresponding to the first round of iteration is determined based on the initial training data and initial reward of each terminal device; and using the target incentive algorithm corresponding to the first round of iteration, the first round of iteration is iterated round by round based on the initial training data and initial reward of each terminal device according to the decision order, and obtains the decision result corresponding to the first round of iteration; besides the method described above which, for the nth round of iteration, the target incentive algorithm corresponding to the first round of iteration is determined based on the initial training data and initial reward of each terminal device; and using the target incentive algorithm corresponding to the first round of iteration, the first round of iteration is iterated round by round based on the initial training data and initial reward of each terminal device ... initial training data and initial reward of each terminal device is determined; and so on; The method of determining the target incentive algorithm for the nth iteration based on the amount of training data and reward after the (n-1)th iteration; and using the target incentive algorithm corresponding to the nth iteration, performing the nth iteration on the amount of training data and reward after the (n-1)th iteration for each terminal device according to the decision order to obtain the decision result corresponding to the nth iteration; other reasonable methods besides those described above; other methods besides those described above for obtaining the target accuracy function of each local model of ensemble learning; other methods besides those described above for using the maximum value of the difference between the target accuracy function and the sum of the reward functions as the incentive algorithm; are not limited here.

[0073] In this embodiment, the initial training data volume and initial reward for each terminal device participating in ensemble learning can be determined. Using an incentive algorithm, the initial training data volume and initial reward for each terminal device are iteratively decided in N rounds according to the decision order, yielding the decision result for each iteration. The incentive algorithm is determined based on the computational and communication costs of each terminal device and is used to decide the corresponding training data volume and reward for each terminal device. The goal is to maximize the accuracy of the ensemble learning model and minimize the rewards paid to each terminal device after N iterations. The training data volume and reward for each terminal device whose decision result satisfies a preset convergence condition are taken as the target training data volume and target reward for each terminal device. By utilizing the incentive mechanism (incentive algorithm) on the ensemble learning server, considering both the accuracy of the ensemble learning model and the reasonableness of the rewards for each terminal device, the accuracy of the ensemble learning model can be maximized while minimizing the rewards for each terminal device. This allows for adaptive decision-making and incentives for terminal devices with different computational and communication costs to participate in ensemble learning. Secondly, the accuracy of an ensemble learning model can be maximized by simultaneously considering the accuracy of individual models and the diversity of decisions among different models. This involves analyzing the coupling relationship between the performance of individual models and the diversity among them, fully considering the differences in resource costs among owners of different computing devices, and designing a distributed ensemble learning incentive mechanism. This maximizes the performance of the ensemble learning model while minimizing the total incentive cost (minimizing the rewards paid to each terminal device), improving the feasibility of adaptive decision-making and incentivizing terminal devices with different computing and communication costs to participate in ensemble learning. Furthermore, the total cost for each terminal device can be obtained, and then the total costs of each terminal device can be sorted in ascending order. This ascending cost order serves as the decision order, improving the efficiency of determining the amount of training data and rewards for ensemble learning while achieving the goal of maximizing the accuracy of the ensemble learning model and minimizing the rewards paid to each terminal device after N iterations. Finally, the weight values ​​of the target accuracy function can be determined, and the weight values ​​are multiplied by the target accuracy function to obtain the product. The maximum value of the difference between the product and the sum of the rewards function is used as the activation algorithm. The amount of training data and the rewards for ensemble learning can be determined according to the accuracy of the ensemble learning model and the importance between the rewards of each terminal device, thereby improving the adjustability of the decision objective.

[0074] The above describes the training data volume and reward determination method for ensemble learning in the embodiments of this application. The following describes the ensemble server in the embodiments of this application. Please refer to [link / reference]. Figure 5 One embodiment of the integrated server in this application includes:

[0075] The determination unit 501 is used to determine the initial amount of training data and the initial reward for each terminal device participating in ensemble learning;

[0076] Decision unit 502 is used to perform N rounds of iterative decision-making on the initial training data volume and initial reward of each terminal device according to the decision-making order using an incentive algorithm, to obtain the decision result corresponding to each round of iteration; wherein, the incentive algorithm is determined based on the computing and communication costs of each terminal device, and is used to decide the training data volume and reward corresponding to each terminal device, so as to maximize the accuracy of the ensemble learning model and minimize the reward of each terminal device after N rounds of iteration; the decision result corresponding to the iteration includes the training data volume and reward corresponding to each terminal device; N is an integer greater than or equal to 1;

[0077] The determining unit 501 is further configured to use the training data volume and reward corresponding to each terminal device whose decision result satisfies the preset convergence condition as the target training data volume and target reward corresponding to each terminal device; wherein, the target training data volume corresponding to each terminal device is used to train the local model corresponding to each terminal device; and the target reward corresponding to each terminal device is the reward obtained by each terminal device training the local model based on the corresponding target training data volume; to obtain the trained local model corresponding to each terminal device, and to integrate the trained local models corresponding to each terminal device to obtain the ensemble learning model, and to predict the data to be predicted based on the ensemble learning model.

[0078] In this embodiment, the initial training data volume and initial reward for each terminal device participating in ensemble learning can be determined. Using an incentive algorithm, the initial training data volume and initial reward for each terminal device are iteratively decided in N rounds according to the decision order, yielding the decision result for each iteration. The incentive algorithm is determined based on the computational and communication costs of each terminal device and is used to decide the training data volume and reward for each terminal device, aiming to maximize the accuracy of the ensemble learning model and minimize the rewards paid to each terminal device after N iterations. The training data volume and reward for each terminal device whose decision result satisfies a preset convergence condition are taken as the target training data volume and target reward for each terminal device. By utilizing the incentive mechanism (incentive algorithm) on the ensemble learning server, considering both the accuracy of the ensemble learning model and the reasonableness of the rewards for each terminal device, the accuracy of the ensemble learning model can be maximized while minimizing the rewards paid to each terminal device. This allows for adaptive decision-making and incentivizes terminal devices with different computational and communication costs to participate in ensemble learning.

[0079] The integrated server in the embodiments of this application is described in detail below. Please refer to [link / reference]. Figure 6Another embodiment of the integrated server in this application includes:

[0080] The determination unit 601 is used to determine the initial amount of training data and the initial reward for each terminal device participating in ensemble learning;

[0081] Decision unit 602 is used to perform N rounds of iterative decision-making on the initial training data volume and initial reward of each terminal device according to the decision-making order using an incentive algorithm, to obtain the decision result corresponding to each round of iteration; wherein, the incentive algorithm is determined based on the computing and communication costs of each terminal device, and is used to decide the training data volume and reward corresponding to each terminal device, so as to maximize the accuracy of the ensemble learning model and minimize the reward of each terminal device after N rounds of iteration; the decision result corresponding to the iteration includes the training data volume and reward corresponding to each terminal device; N is an integer greater than or equal to 1;

[0082] The determining unit 601 is further configured to use the training data volume and reward corresponding to each terminal device whose decision result satisfies the preset convergence condition as the target training data volume and target reward corresponding to each terminal device; wherein, the target training data volume corresponding to each terminal device is used to train the local model corresponding to each terminal device; and the target reward corresponding to each terminal device is the reward obtained by each terminal device training the local model based on the corresponding target training data volume; to obtain the trained local model corresponding to each terminal device, and to integrate the trained local models corresponding to each terminal device to obtain the ensemble learning model, and to predict the data to be predicted based on the ensemble learning model.

[0083] The integrated server also includes:

[0084] The obtaining unit 603 is used to obtain the total cost corresponding to each terminal device, the total cost including computing cost and communication cost;

[0085] The determining unit 601 is further configured to sort the total cost of each terminal device in ascending order of cost, and determine the ascending order of cost as the decision order.

[0086] The decision unit 602 is specifically configured to, for the first iteration, determine the target incentive algorithm corresponding to the first iteration based on the initial training data volume and initial reward of each terminal device; and use the target incentive algorithm corresponding to the first iteration to perform the first iteration on the initial training data volume and initial reward of each terminal device in the decision order, to obtain the decision result corresponding to the first iteration; the decision result corresponding to the first iteration includes the training data volume and reward of each terminal device after the first iteration. For the nth iteration, the decision unit 602 determines the target incentive algorithm corresponding to the nth iteration based on the training data volume and reward of each terminal device after the (n-1)th iteration; and uses the target incentive algorithm corresponding to the nth iteration to perform the nth iteration on the training data volume and reward of each terminal device after the (n-1)th iteration, to obtain the decision result corresponding to the nth iteration; the decision result corresponding to the nth iteration includes the training data volume and reward of each terminal device after the nth iteration; where 2≤n≤N.

[0087] The decision unit 602 is specifically configured to, for the terminal device making the first decision in the first round of iteration, determine the target incentive algorithm corresponding to the terminal device making the first decision in the first round of iteration based on the initial training data volume and initial reward of each terminal device; and use the target incentive algorithm corresponding to the terminal device making the first decision in the first round of iteration to perform the decision in the first round of iteration on the terminal device making the first decision, to obtain the decision result of the terminal device making the first decision in the first round of iteration; and for the terminal device making the m-th decision in the first round of iteration, based on the first round of iteration corresponding to the terminal device making the first decision to the (m-1)-th decision, [further details about the first round of iteration are needed]. Given the corresponding training data volume and reward, and the initial training data volume and initial reward for each of the terminal devices making decisions from the m-th to the M-th decisions, determine the target incentive algorithm corresponding to the terminal device making the m-th decision in the first round of iteration; and use the target incentive algorithm corresponding to the terminal device making the m-th decision in the first round of iteration to make decisions for the terminal device making the m-th decision in the first round of iteration, to obtain the decision result of the terminal device making the m-th decision in the first round of iteration; where 2≤m≤M, and the target incentive algorithm corresponding to the first round of iteration includes the target incentive algorithms corresponding to each of the terminal devices making decisions from the 1st to the Mth decisions in the first round of iteration.

[0088] The decision unit 602 is specifically configured to, for the terminal device making the first decision in the nth iteration, determine the target incentive algorithm corresponding to the terminal device making the first decision in the nth iteration based on the amount of training data and reward after the (n-1)th iteration for each terminal device; and use the target incentive algorithm corresponding to the terminal device making the first decision in the nth iteration to perform the decision for the terminal device making the first decision in the nth iteration, thereby obtaining the decision result of the terminal device making the first decision in the nth iteration; and for the terminal device making the mth decision in the nth iteration, based on the nth iteration for each of the terminal devices making the first decision to the (m-1)th decisions, determine the target incentive algorithm corresponding to the terminal device making the first decision in the nth iteration. Given the corresponding training data volume and reward, and the training data volume and reward for each of the terminal devices making decisions from the m-th to the M-th decisions in the (n-1)th iteration, determine the target incentive algorithm corresponding to the terminal device making the m-th decision in the n-th iteration; and use the target incentive algorithm corresponding to the terminal device making the m-th decision in the n-th iteration to make decisions for the terminal device making the m-th decision in the n-th iteration, to obtain the decision result of the terminal device making the m-th decision in the n-th iteration; where 2≤m≤M, and the target incentive algorithm corresponding to the n-th iteration includes the target incentive algorithms corresponding to each of the terminal devices making decisions from the 1st to the Mth decisions in the n-th iteration.

[0089] The obtaining unit 603 is also used to obtain the target accuracy function of each local model of the ensemble learning and the sum of the rewards of each terminal device of the ensemble learning.

[0090] The determining unit 601 is further configured to use the maximum value of the difference between the target accuracy function and the sum of rewards function as the incentive algorithm.

[0091] The obtaining unit 603 is used to determine the total number of terminal devices participating in ensemble learning, the total amount of training data, and the number of local models with prediction errors corresponding to each training data; and to obtain the diversity function among the local models based on the total number of terminal devices participating in ensemble learning, the total amount of training data, and the number of models with prediction errors corresponding to each training data, and to determine the average accuracy of each local model; and to determine the accuracy function of each local model based on the total number of terminal devices participating in ensemble learning and the average accuracy of each local model, and to use the sum of the diversity function among the local models and the accuracy function of each local model as the target accuracy function of each local model in the ensemble learning.

[0092] The determining unit 601 is further configured to determine the weight values ​​of the target accuracy function;

[0093] The determining unit 601 is specifically used to multiply the weight value by the target accuracy function to obtain a product, and to use the maximum value of the difference between the product and the sum of rewards function as the incentive algorithm.

[0094] In this embodiment, each unit in the integration server performs as described above. Figure 2 The operation of the integrated server in the illustrated embodiment will not be described in detail here.

[0095] Please refer to the following: Figure 7 Another embodiment of the integrated server 700 in this application includes:

[0096] Central processing unit 701, memory 705, input / output interface 704, wired or wireless network interface 703, and power supply 702;

[0097] Memory 705 is either a short-term storage memory or a persistent storage memory;

[0098] The central processing unit 701 is configured to communicate with the memory 705 and execute instructions stored in the memory 705 to perform the aforementioned operations. Figure 2 The method in the illustrated embodiment.

[0099] This application also provides a computer-readable storage medium, which includes instructions that, when executed on a computer, cause the computer to perform the aforementioned actions. Figure 2 The method in the illustrated embodiment.

[0100] This application also provides a computer program product containing instructions, which, when run on a computer, causes the computer to perform the aforementioned... Figure 2 The method in the illustrated embodiment.

[0101] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0103] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0104] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0105] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0106] If the integrated unit is implemented as 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 this application, in essence, 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. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. An integrated learning training data volume and reward determination method, characterized in that, The method comprises the following steps: determining the initial training data amount and the initial reward of each terminal device participating in the integrated learning; using an incentive algorithm to make N rounds of iteration decisions on the initial training data amount and the initial reward of each terminal device in a decision order, to obtain a corresponding decision result of each round of iteration; wherein the incentive algorithm is determined based on the calculation and communication costs of each terminal device, and is used to decide the training data amount and the reward of each terminal device, so that the accuracy of the integrated learning model reaches a maximum value and the reward paid to each terminal device reaches a minimum value after N rounds of iteration; the iteration corresponding decision result comprises the training data amount and the reward corresponding to each terminal device; N is an integer greater than or equal to 1; the training data amount and the reward corresponding to each terminal device that meets the preset convergence condition are taken as the target training data amount and the target reward corresponding to each terminal device; wherein the target training data amount corresponding to each terminal device is used to train the local model corresponding to each terminal device respectively; and the target reward corresponding to each terminal device is the reward obtained by training the local model based on the corresponding target training data amount; to obtain the trained local model corresponding to each terminal device, and integrate the trained local model corresponding to each terminal device to obtain the integrated learning model, and predict the to-be-predicted data according to the integrated learning model.

2. The method of claim 1, wherein, Before the step of using the incentive algorithm to make N rounds of iteration decisions on the initial training data amount and the initial reward of each terminal device in a decision order, to obtain a corresponding decision result of each round of iteration, the method further comprises: obtaining the total cost corresponding to each terminal device, wherein the total cost comprises the calculation cost and the communication cost; sorting the total cost of each terminal device in a cost from low to high order, and determining the cost from low to high order as the decision order.

3. The method of claim 1, wherein, The step of using the incentive algorithm to make N rounds of iteration decisions on the initial training data amount and the initial reward of each terminal device in a decision order, to obtain a corresponding decision result of each round of iteration, comprises: for the first round of iteration, determining the target incentive algorithm corresponding to the first round of iteration based on the initial training data amount and the initial reward of each terminal device; and using the target incentive algorithm corresponding to the first round of iteration to make the first round of iteration on the initial training data amount and the initial reward of each terminal device in the decision order, to obtain the corresponding decision result of the first round of iteration; the corresponding decision result of the first round of iteration comprises the training data amount and the reward of each terminal device after the first round of iteration; for the nth round of iteration, determining the target incentive algorithm corresponding to the nth round of iteration based on the training data amount and the reward of each terminal device after the (n-1)th round of iteration; and using the target incentive algorithm corresponding to the nth round of iteration to make the nth round of iteration on the training data amount and the reward of each terminal device after the (n-1)th round of iteration in the decision order, to obtain the corresponding decision result of the nth round of iteration; the corresponding decision result of the nth round of iteration comprises the training data amount and the reward of each terminal device after the nth round of iteration; wherein 2≤n≤N.

4. The method of claim 3, wherein, The target incentive algorithm corresponding to the first round of iterations is determined based on the initial training data amount and the initial reward of each terminal device; and the target incentive algorithm corresponding to the first round of iterations is used to perform the first round of iterations on the initial training data amount and the initial reward of each terminal device according to the decision sequence, to obtain a decision result corresponding to the first round of iterations, including: For the terminal device of the first decision in the first round of iterations, the target incentive algorithm corresponding to the terminal device of the first decision in the first round of iterations is determined based on the initial training data amount and the initial reward of each terminal device; and the target incentive algorithm corresponding to the terminal device of the first decision in the first round of iterations is used to perform the decision of the first round of iterations on the terminal device of the first decision, to obtain a decision result of the terminal device of the first decision corresponding to the first round of iterations. For the terminal device of the mth decision in the first round of iterations, the target incentive algorithm corresponding to the terminal device of the mth decision in the first round of iterations is determined based on the training data amount and the reward corresponding to the first round of iterations of the terminal device of the first decision to the terminal device of the m-1th decision, and the initial training data amount and the initial reward corresponding to the terminal device of the mth decision to the terminal device of the Mth decision; and the target incentive algorithm corresponding to the terminal device of the mth decision in the first round of iterations is used to perform the decision of the first round of iterations on the terminal device of the mth decision, to obtain a decision result of the terminal device of the mth decision corresponding to the first round of iterations; wherein 2≤m≤M, and the target incentive algorithm corresponding to the first round of iterations includes the target incentive algorithm corresponding to the terminal device of the first decision to the terminal device of the Mth decision in the first round of iterations.

5. The method of claim 3, wherein, The target incentive algorithm corresponding to the nth round of iterations is determined based on the training data amount and the reward of each terminal device after the (n-1)th round of iterations; and the target incentive algorithm corresponding to the nth round of iterations is used to perform the nth round of iterations on the training data amount and the reward of each terminal device corresponding to the (n-1)th round of iterations according to the decision sequence, to obtain a decision result corresponding to the nth round of iterations, including: For the terminal device of the first decision in the nth round of iterations, the target incentive algorithm corresponding to the terminal device of the first decision in the nth round of iterations is determined based on the training data amount and the reward of each terminal device after the (n-1)th round of iterations; and the target incentive algorithm corresponding to the terminal device of the first decision in the nth round of iterations is used to perform the decision of the nth round of iterations on the terminal device of the first decision, to obtain a decision result of the terminal device of the first decision corresponding to the nth round of iterations. For the mth decision-making terminal device in the nth iteration, a target incentive algorithm corresponding to the mth decision-making terminal device in the nth iteration is determined based on the training data amount and the reward corresponding to the nth iteration of the respective first to (m-1)th decision-making terminal devices, and the training data amount and the reward corresponding to the (n-1)th iteration of the respective (mth to Mth) decision-making terminal devices; and the mth decision-making terminal device in the nth iteration is subjected to decision-making in the nth iteration by using the target incentive algorithm corresponding to the mth decision-making terminal device in the nth iteration, to obtain a decision-making result of the mth decision-making terminal device in the nth iteration; wherein 2≤m≤M, and the target incentive algorithm corresponding to the nth iteration includes the respective target incentive algorithms corresponding to the first to Mth decision-making terminal devices in the nth iteration.

6. The method of claim 1, wherein, Before the method of using the incentive algorithm to perform N iterations of decision-making on the initial training data amount and the initial reward of each terminal device in a decision-making order to obtain a decision-making result corresponding to each iteration, the method further comprises: obtaining a target accuracy function of each local model of the ensemble learning; obtaining a sum function of the rewards of each terminal device of the ensemble learning; taking a maximum value of a difference between the target accuracy function and the sum function of the rewards as the incentive algorithm.

7. The method of claim 6, wherein, The method of obtaining the target accuracy function of each local model of the ensemble learning comprises: determining a total number of terminal devices participating in the ensemble learning, a total training data amount, and a number of local models that are predicted to be wrong for each training data; and obtaining a diversity function between the local models based on the total number of terminal devices participating in the ensemble learning, the total training data amount, and the number of models predicted to be wrong for each training data; determining an average accuracy of each local model; and determining an accuracy function of each local model based on the total number of terminal devices participating in the ensemble learning and the average accuracy of each local model; taking a sum of the diversity function between the local models and the accuracy function of each local model as the target accuracy function of each local model of the ensemble learning.

8. The method of claim 6, wherein, The method of taking a maximum value of a difference between the target accuracy function and the sum function of the rewards as the incentive algorithm further comprises: determining a weight value of the target accuracy function; The method of taking a maximum value of a difference between the target accuracy function and the sum function of the rewards as the incentive algorithm comprises: multiplying the weight value and the target accuracy function to obtain a product; taking a maximum value of a difference between the product and the sum function of the rewards as the incentive algorithm.

9. An integrated server, characterized by The method comprises: a central processing unit and a memory; the memory is a volatile memory or a persistent memory; the central processing unit is configured to communicate with the memory and perform instruction operations in the memory to execute the method of any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium comprises instructions that, when executed on a computer, cause the computer to perform the method of any one of claims 1 to 8.

Images