Calculation method of decision variable

Abstract

To provide a calculation method of a decision variable, capable of calculating a decision variable which can achieve an effect of reverse estimation of the decision variable, without training in common, a shared model by combination learning, by members of the combination learning.SOLUTION: A reverse estimation request person can calculate, a desired decision variable by reverse estimation, according to a target result, by a learned prediction model provided from some members of combination learning, in a calculation method of the decision variable, the method includes: reverse estimation of an input parameter by providing a target result to the learned prediction model related to the method; formation of a loss function, from a difference between each learned prediction model and the target result; integration of all input parameters subjected to reverse estimation, to an integrated input parameter, integration of all loss functions to an integrated loss function, construction of an optimized problem, by the integrated input parameter and integrated loss function for determining a solution thereof, the solution determined from the optimized problem is a desired decision variable.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a method for calculating decision variables, and more particularly to a method for inversely estimating decision variables that match a target outcome using multiple trained artificial intelligence prediction models provided by members of a federated learning system. [Background technology]

[0002] With the advancement of medical engineering technology, the application of regenerative medicine in clinical treatment of diseases is becoming more and more diverse, and its research and application are attracting more and more attention from all walks of life. Currently, regenerative medicine is mainly applied to organ repair, immune cell therapy, and stem cell therapy. Cell therapy involves the procedure of culturing or processing human cells outside the body, and then transplanting the cultured cells into an individual's body for use. Therefore, how to effectively culture cells is a crucial element in regenerative medicine. Previous research has shown that the cells in each person's body have unique characteristics, and it is necessary to customize cells suitable for treatment according to each situation, which increases the complexity and difficulty of cell processes. On the other hand, when there is a large amount of cell data and bioprocess data, machine learning can generate predictive models that evaluate the effects of each cell culture condition and process parameter. Thus, machine learning can reduce the complexity and difficulty of customizing cell processes. However, most of these cell data and bioprocess data belong to very sensitive and highly confidential personal data, such as cell data of patients in various hospitals or laboratories. Conventional aggregated machine learning methods require all bioprocessing data to be aggregated in a central server for analysis and training, which may be impractical due to constraints such as regulations, privacy protection, and commercial considerations. In addition, aggregated machine learning methods may face similar problems not only in biomedical engineering, but also in any technical field where data is highly sensitive and confidential.

[0003] Federated learning is a type of machine learning technology, also known as collaborative learning, in which an algorithm is trained on multiple distributed edge devices or servers with local data samples. This method is significantly different from traditional aggregated machine learning technology, in which all local datasets are uploaded to one server for machine learning training. On the other hand, federated learning is composed of multiple members. Each member performs machine learning training on its own dataset to generate its own trained predictive model, and uploads the trained predictive model parameters to a central server. The central server assigns weights to all the trained predictive model parameters and aggregates them to form a federated model. This federated model can be redistributed to each member for use. The central server only aggregates the trained predictive model parameters, and does not directly perform machine learning analysis or processing on each member's dataset. Therefore, the highly sensitive and confidential data samples and information in each member's dataset are not leaked to the outside, and the confidentiality of the data is protected.

[0004] The federated model formed by the above federated aggregation algorithm can not only generate a prediction result for the input decision variables, but also, conversely, can inversely estimate the decision variables to be input for the target result. However, compared with developing an inverse estimation function for a unified model that is jointly trained and shared, the members participating in the federated learning method can individually develop an inverse estimation function using a prediction model trained on a local data set, but the effect of inverse estimation may not be so ideal. In addition, when a unit other than the federated member wants to cooperate with a member with good prediction effect or inverse estimation effect, it is not possible to know which federated member's prediction model best meets the needs of the unit by only testing the effect of the jointly developed federated model.

[0005] In addition, in federated learning, although very sensitive and highly confidential data samples in the datasets of each member are not leaked to the outside, the central server and each member still need to exchange model parameters. Therefore, the method of obtaining decision variables by performing inverse estimation using a jointly developed federated model not only has a high process complexity, but also requires direct or indirect exchange of model parameters between each member.

[0006] Therefore, in order to solve the problems of complexity and model confidentiality when inversely estimating decision variables using a federated model, it is necessary to develop a method that enables federated learning members to inversely estimate decision variables without having to jointly calculate a shared predictive model using a federated learning algorithm. Summary of the Invention [Problem to be solved by the invention]

[0007] In view of the above, the present invention provides a method for calculating decision variables to solve the above-mentioned problems of the prior art. [Means for solving the problem]

[0008] According to a specific embodiment of the present invention, a method for calculating a decision variable includes the steps of: providing a plurality of trained prediction models, each of which is obtained by machine learning a plurality of local data sets using a machine learning method; providing a target result, and using the trained prediction model to inversely estimate a plurality of input parameters for the target result; setting a plurality of loss functions corresponding to the trained prediction models, each of which is assigned a first weighting value to each of the loss functions, and then summing them to generate a comprehensive loss function, each of which is defined as the absolute value, square of the difference value between the function value of an objective function representing each of the trained prediction models and the target result, or a function having a monotonically increasing characteristic with respect to the difference value; generating comprehensive input parameters by assigning second weighting values ​​to each of the input parameters and then summing them, and constructing an optimization problem using the comprehensive input parameters and the comprehensive loss function; and obtaining a solution to the optimization problem to obtain the second weighting value of each of the input parameters, and calculating the comprehensive input parameters as the decision variables.

[0009] The method for calculating decision variables further includes the steps of: calculating a derivative of the overall loss function for each of the second weight values; combining the derivatives to form a direction; and inputting the overall input parameters into the trained predictive model after increasing a step size along the direction to determine whether the function value of the overall loss function decreases.

[0010] The trained predictive models are each provided by multiple members of a federated learning model.

[0011] In the method for calculating decision variables, the step of inversely estimating the input parameters for each of the target outcomes using the trained predictive model further includes a step by a first member of the members comparing the target outcome with all samples of a corresponding first local dataset among the local datasets, and selecting samples in the first local dataset that match the target outcome as reference samples for inversely estimating the input parameters.

[0012] In the method for calculating decision variables, the step of inversely estimating the input parameters for the target result using the trained prediction model further includes a step of a second member of the members comparing the target result with all samples of a corresponding second local dataset among the local datasets to obtain a plurality of anchor samples and forming a multidimensional subspace using the anchor samples; a step of determining an initial sample in the multidimensional subspace; a step of obtaining a direction of the initial sample in the multidimensional subspace, forming a relay sample for the initial sample by increasing a step size along the direction, and inputting the relay sample into a second trained prediction model provided by the second member to check whether the generated prediction result is close to the target result; and a step of continuing to input the relay sample with the determined direction and the increased step size into the second trained prediction model until the prediction result generated by the second trained prediction model matches the target result, and setting the last relay sample as the input parameter corresponding to the second trained prediction model.

[0013] In the method for calculating decision variables, the step of inversely estimating the input parameters for the target result using the trained prediction model further includes a step in which a third member among the members sets a dummy layer connected to an input terminal of a corresponding third trained prediction model to form a parameter prediction model, the dummy layer becoming the input terminal of the parameter prediction model, and a step in which the output of the parameter prediction model is set to the target result, a training data set including at least one all-one vector is input to the parameter prediction model to train the parameter prediction model, and a step in which an optimizer in a machine learning method that trained the third trained prediction model adjusts multiple arc weight values ​​between the dummy layer and the third trained prediction model, and the arc weight values ​​become the input parameters corresponding to the third trained prediction model.

[0014] In the method for calculating decision variables, the step of inversely estimating the input parameters for the target result using the trained predictive model further includes a step of a fourth member of the members comparing the target result with all samples of a corresponding fourth local dataset among the local datasets to obtain a plurality of anchor samples, forming a multidimensional subspace using the anchor samples, a step of determining an initial sample in the multidimensional subspace, and a step of obtaining a gradient by performing a minimization operation on the objective function of the fourth trained predictive model using an optimizer used by the fourth member to generate the fourth trained predictive model. and setting the opposite direction of the gradient as the direction of the initial sample; forming relay samples by increasing a step size along the direction for the initial sample, and inputting the relay samples into the fourth trained prediction model to check whether the generated prediction result is close to the target result; continuing to perform the operation of determining the direction and inputting the relay samples with the increased step size into the fourth trained prediction model until the prediction result generated by the fourth trained prediction model matches the target result, and setting the final relay sample as the input parameter corresponding to the fourth trained prediction model.

[0015] In the method for calculating decision variables, the step of inversely estimating the input parameters for the target outcome respectively using the trained predictive model further includes the steps of: providing a plurality of confirmed decision variables and calculating a first vector of the confirmed decision variables; obtaining a plurality of corresponding sample parameters from sample parameters in all samples of the fourth local dataset based on the confirmed decision variables, respectively, and calculating a second vector of the corresponding sample parameters; obtaining a reference sample by comparing the first vector with the second vector and comparing the predicted results generated by all samples of the fourth local dataset in the fourth trained predictive model with the target outcome, wherein the second vector of the reference sample is close to or identical to the first vector, and the predicted result of the reference sample is close to or identical to the target outcome; and setting the reference sample as the initial sample.

[0016] The method for calculating the decision variable further includes the steps of providing a first test sample, the first test sample including a plurality of confirmed test input parameters and a test target result, and inputting the test input parameters into a plurality of trained predictive model candidates, and determining the trained predictive model candidate whose output result matches the test target result as the trained predictive model.

[0017] The method for calculating the decision variable further includes the steps of: providing a first test sample, the first test sample including a plurality of confirmed first test input parameters and a first test target result; inputting the first test input parameters into a plurality of trained prediction model candidates, and determining the trained prediction model candidates whose output results match the first test target result as a plurality of first trained prediction model candidates; providing a second test sample, the second test sample including a plurality of confirmed second test input parameters and a second test target result; and providing the second test target results to the first trained prediction model candidates, respectively, to have the first trained prediction model candidate inversely estimate a plurality of first input parameter candidates, and determining the first trained prediction model candidate where the first input parameter candidate matches or is close to the second test input parameters or where the first input parameter candidate satisfies a validation process as the trained prediction model. Effect of the Invention

[0018] In summary, the method for calculating decision variables in the present invention can inversely estimate input parameters (i.e., decision variables) that match a specific target result by using trained predictive models provided by multiple members of the federated learning, without jointly calculating a shared model through federated learning. According to the method of the present invention, a specific member in the federated learning or a unit outside the federated learning can request inverse estimation. In addition, since it is possible to directly seek and select collaborators without going through the federated model generated by federated learning, it is possible to avoid complex machine learning and inverse estimation processes. In addition, the method for calculating decision variables in the present invention does not require members of the federated learning to exchange very sensitive and highly confidential data samples and information in the training dataset, and does not require the members of the federated learning to exchange model parameters of the trained predictive models. Therefore, not only is the complexity of the process reduced, but the confidentiality of the data is further protected. [Brief description of the drawings]

[0019] [Figure 1]FIG. 1 shows a flow chart of the steps of a method for calculating decision variables according to a specific embodiment of the present invention. [Diagram 2] FIG. 2 shows a flow chart of the further steps of step S18 in the method of calculating the decision variables of FIG. [Diagram 3] FIG. 3 shows a flow chart of the steps of a method for calculating decision variables according to another embodiment of the present invention. [Figure 4] FIG. 4 shows a flow chart of the steps of a method for calculating decision variables according to another embodiment of the present invention. [Diagram 5] FIG. 5 shows a flow chart of the steps of a method for calculating decision variables according to another embodiment of the present invention. [Figure 6] FIG. 6 shows a flow chart of the steps of a method for calculating decision variables according to another embodiment of the present invention. [Figure 7] FIG. 7 shows a flow chart of the steps of a method for calculating decision variables according to another embodiment of the present invention. [Figure 8] FIG. 8 shows a flow chart of the steps of a method for calculating decision variables according to another embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] In order to make the advantages, spirits and features of the present invention easier and clearer to understand, the present invention will be described and discussed in detail with reference to the drawings, using specific examples. It should be noted that these specific examples are merely representative specific examples of the present invention, and the specific methods, devices, conditions, materials, etc. illustrated do not limit the present invention or the corresponding specific examples. In addition, each component in the drawings is used only to represent their relative positions, and is not described based on actual ratios. In addition, the step numbers of the present invention are only used to distinguish different steps, and do not represent the order of steps. The above will be explained in advance.

[0021] Please refer to Fig. 1. Fig. 1 shows a flowchart of the steps of the calculation method of decision variables according to a specific embodiment of the present invention. As shown in Fig. 1, the calculation method of decision variables in this specific embodiment includes the following steps:

[0022] Step S10: Providing a plurality of trained predictive models, each of which is obtained by machine learning the local dataset through a machine learning method.

[0023] Step S12: Provide a target result, and inversely estimate a plurality of input parameters for the target result, respectively, using the trained prediction model.

[0024] Step S14: A plurality of loss functions corresponding to the trained prediction models are set, and different first weight values ​​are set for these loss functions, and then the loss functions are summed up to generate an overall loss function.

[0025] Step S16: A second weight value is assigned to each of the input parameters, and then the input parameters are summed to generate a comprehensive input parameter, and an optimization problem is constructed using the comprehensive input parameter and a comprehensive loss function.

[0026] Step S18: obtain a second weight value of each input parameter by solving the optimization problem, and calculate a comprehensive input parameter as a decision variable.

[0027] In this specific embodiment, the trained prediction models in step S10 are provided by the members of the federated learning, respectively. In detail, each member of the federated learning may have a respective artificial intelligence model training device and also has a respective local dataset. Thus, each member can train its own artificial intelligence prediction model individually. For example, in the case of the regenerative medicine field, each member's artificial intelligence model training device may be a computer device capable of executing a machine learning algorithm in a hospital or laboratory. The local dataset is a sample dataset for training, validation, and testing held by the hospital or laboratory. In practice, each member of the federated learning may obtain its own trained prediction model by machine learning the local dataset through Artificial Neural Networks (ANN), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or any other machine learning or neural network algorithm. The machine learning algorithm is selected according to the needs of each member. In addition, the trained prediction model in step S10 is a selected trained prediction model, and is provided only by some members of the federated learning. In practice, the trained prediction models of all members do not necessarily meet the requirements because each member of the federated learning has a different local data set, a different machine learning algorithm to be used, a different ability to adjust model parameters, and a different ability to inversely estimate decision variables using the trained prediction model. The selection of the trained prediction model will be described in a specific embodiment below.

[0028] In step S12, the members of the federated learning who provided the trained prediction models can inversely estimate the input parameters based on the target result. In detail, each member can use the same or different inverse estimation algorithm to match the target result with the output result of the trained prediction model, thereby inversely estimating the input parameters such that the output result of the trained prediction model matches the target result. The inverse estimation algorithms used by each member will be described in the specific examples below.

[0029] In step S14, a loss function can be set for each trained prediction model. The loss function is defined as the absolute value or square of the difference value between the function value of the objective function representing each trained prediction model and the target result, or any function having a monotonically increasing characteristic with respect to the difference value (for example, the fourth or sixth power of the difference value, etc.). In other words, the loss function represents the degree of approximation between the output result of each trained prediction model and the target result. The overall loss function is formed by weighting the loss functions corresponding to each trained prediction model and then summing them. In addition, in practical use, the weighting of each loss function (i.e., the first weight value) is determined based on the reliability or performance of each trained prediction model. For example, when a trained prediction model that has a high influence on the provided target result is selected and recognized by the inverse estimation requester, or when the trained prediction model is recognized to have a relatively accurate prediction rate and inverse estimation rate in the test, the first weight value of the loss function corresponding to the trained prediction model can be set higher than the loss functions corresponding to the other trained prediction models.

[0030] Similarly, in step S16, a second weighting value can be assigned to each input parameter that each trained predictive model has previously inversely estimated for the same target result. Unlike the first weighting value in step S14, the second weighting value is not a preset value, but must be determined by calculation in a subsequent step. In steps S14 and S16, a comprehensive loss function and comprehensive input parameters corresponding to each trained predictive model are generated, and an optimization problem is constructed and a solution is obtained using these to obtain optimal input parameters.

[0031] In step S18, the method of solving the optimization problem is to find the comprehensive input parameters with the smallest comprehensive loss function. Since the input parameters of each trained prediction model in the comprehensive input parameters are known, the second weight values ​​are obtained by solving the optimization problem. After obtaining all the second weight values, the comprehensive input parameters can be calculated to obtain them, which can be used as the desired decision variables.

[0032] The method for solving the optimization problem described above can be carried out by a numerical analysis method. Now, reference is made to Fig. 2. Fig. 2 shows a flowchart of further steps of step S18 in the method for calculating the decision variables in Fig. 1. As shown in Fig. 2, the above step S18 may further include the following steps:

[0033] Step S180: Calculate the derivative of the overall loss function for each second weight value respectively.

[0034] Step S182: The derivatives are combined to form a direction.

[0035] Step S184: After increasing the step size along the above direction for the overall input parameters, input the step size to all trained prediction models to determine whether the function value of the overall loss function decreases.

[0036] As described above, the overall input parameters are obtained by multiplying the multiple input parameters inversely estimated in step S12 by the second weighting values ​​and then summing them up. In other words, each second weighting value has a different effect on the function value of the final overall loss function. Therefore, in step S180 of the calculation method of the decision variables in this specific embodiment, the derivative of the overall loss function for each second weighting value is calculated. That is, the change in the function value of the overall loss function formed for a slight change in each second weighting value is calculated. Each derivative is the rate of change of the overall loss function for each different second weighting value. Therefore, as shown in step S182, the above direction can be obtained by multiplying each derivative by the unit vector formed by each second weighting value and summing them up. After the direction is determined in step S182, in step S184, an appropriate step size is set in this direction, and the overall input parameters after the step size is increased are substituted into the overall loss function to obtain an appropriate reduction value. Therefore, the new overall input parameters (advancement by one step) can obtain a function value of the overall loss function smaller than that of the previous overall input parameters. In other words, it is possible to get closer to the solution of the optimization problem. Steps S180 to S184 can be repeatedly executed until the smallest function value of the overall loss function is found, and the overall input parameters found at that time become the solution of the optimization problem. In other words, these can be used as decision variables.

[0037] For example, in the method of calculating decision variables, when trained prediction models provided by two members of the federated learning are selected, the target result T is first provided to these two members. Then, the input parameters I1 and I2 that match the target result T are inversely estimated by each trained prediction model, and two loss functions L1 and L2 corresponding to these two trained prediction models are set. The loss functions L1 and L2 are the absolute values ​​|f1-T| and |f2-T| of the difference between the functions f1 and f2 representing the trained prediction models and the target result, respectively. Next, the first weight values ​​α and 1-α are respectively assigned to the two loss functions, and then the two loss functions are summed up to generate an overall loss function L=α×|f1-T|+(1-α)×|f2-T|. In addition, the two input parameters I1 and I2 are respectively assigned second weight values ​​β and γ, and then the overall input parameter S is summed up to generate an overall loss function L=α×|f1-T|+(1-α)×|f2-T|. 1 = β × I1 + γ × I2. The optimization problem constructed by the comprehensive input parameters and the comprehensive loss function is to minimize the comprehensive loss function L by finding the optimal β, γ and comprehensive input parameters in a situation where α, T, I1, and I2 are all fixed.

[0038] In the method for finding a solution to the optimization problem described above, first, the derivatives d1 and d2 of the overall loss function L with respect to the second weight values ​​β and γ are calculated, respectively. Next, a direction g = (d1 × (1,0) + d2 × (0,1)) is set. Note that (1,0) and (0,1) are unit vectors of β and γ, respectively. After the direction is found, a step size t is set along this direction, and the overall input parameter is advanced by the step size t in that direction to obtain a new overall input parameter S 2 =S 1 -t×g can be generated. Note that the derivatives d1 and d2 represent the gradient of the overall loss function with respect to β and γ, and the direction of the step size is the opposite direction of the gradient. By adjusting the appropriate step size t, the new overall input parameter S after the step size is 2 After substituting into the overall loss function L, the reduced function value (the original overall input parameter S 1(compared with the function value obtained by inputting the new overall input parameter S 2 The prediction result obtained by the above-mentioned method will be closer to the target result. By repeating the above-mentioned steps of calculating the direction and adjusting the step size, a comprehensive input parameter that can obtain the smallest function value by substituting it into the comprehensive loss function L can be finally obtained. This comprehensive input parameter can be the decision variable to be calculated by the decision variable calculation method.

[0039] In the above specific embodiment, in the method for calculating the decision variables, a member of the federated learning or a unit outside the federated learning can request inverse estimation, and all members of the federated learning are asked to cooperate. However, in practice, it is not limited to the case where a member of the federated learning can provide a trained prediction model, and the collaborator does not have to be a member of the federated learning. It is sufficient for the collaborator to be able to independently train the trained prediction model and to use the trained prediction model to inversely estimate the input parameters.

[0040] As described above, in step S12, the members of the federated learning may first inversely estimate input parameters by the trained prediction model based on the target result provided by the inverse estimation requester. The inverse estimation method includes, but is not limited to, several aspects described below. Now, refer to FIG. 3. FIG. 3 shows a flowchart of steps of a method for calculating decision variables according to another specific embodiment of the present invention. As shown in FIG. 3, the method for calculating decision variables in this specific embodiment further includes the following steps:

[0041] Step S120: Each member of the federated learning compares all samples in the local dataset with the target result.

[0042] Step S122: The samples in the local data set that match the target result are taken as reference samples for inversely estimating the input parameters.

[0043] It should be noted that in addition to the above steps, the method of this specific embodiment further includes step S10 and steps S14 to S18 in the specific embodiment described above, but since the contents of these are the same as the corresponding steps in the specific embodiment described above, they will not be described in detail here.

[0044] In this specific embodiment, the first member of the federated learning members can use samples in the first local data set when the trained prediction model is trained as reference samples to inversely estimate the input parameters. In detail, if there are confirmed decision variables (e.g., parameters already executed in the process) in the decision variables to be inversely estimated, those confirmed decision variables can be used as vectors. Similarly, corresponding sample parameters corresponding to the confirmed decision variables in all samples in the first local data set can be used as another vector, and the two vectors can be compared to find samples that are close or matched. These samples have corresponding sample parameters similar to the confirmed decision variables. In other words, it means that these samples have high reference value. Therefore, other sample parameters that do not correspond to the confirmed decision variables in these samples can be input into the trained prediction model together with the confirmed decision variables, and it is possible to check whether the prediction result output from the trained prediction model matches the target result. If it matches the target result, the other sample parameters that do not correspond to the confirmed decision variables in this sample and the confirmed decision variables can be used as input parameters (e.g., the input parameter I1).

[0045] Please refer to Figure 4. Figure 4 shows a flowchart of the steps of a calculation method for decision variables according to another specific embodiment of the present invention. As shown in Figure 4, the calculation method for decision variables in this specific embodiment further includes the following steps:

[0046] Step S120': Each member of the federated learning compares all samples in the local dataset with the target result to obtain a number of anchor samples, and forms a multi-dimensional subspace by the anchor samples.

[0047] Step S122': Determine an initial sample in the multi-dimensional subspace.

[0048] Step S124': Obtain the direction of the initial sample in the multi-dimensional subspace, and form relay samples by increasing the step size along the direction of the initial sample, and input the relay samples into the trained prediction model to check whether the generated prediction result is close to the target result.

[0049] Step S126': The operation of inputting the relay samples whose direction has been determined and whose step size has been increased into the trained prediction model is continued until the prediction result generated by the trained prediction model matches the target result, and the last relay sample is taken as the input parameter corresponding to the trained prediction model.

[0050] It should be noted that in addition to the above steps, the method of this specific embodiment further includes step S10 and steps S14 to S18 in the specific embodiment described above, but since the contents of these are the same as the corresponding steps in the specific embodiment described above, they will not be described in detail here.

[0051] In step S120', the local data set is a data set formed of samples used to train the trained prediction model, so that both the input and output results of the samples for the trained prediction model are known. In addition, samples whose output results are close to or match the target result have high reference value, so these samples can be extracted as reference samples. Therefore, in this specific embodiment, a multidimensional subspace is formed using these samples as anchor samples, and this multidimensional subspace becomes the feasible solution space of the optimization problem. Steps S122', S124', and S126' are a method for finding the optimal feasible solution in the feasible solution space. In brief, first, an initial point (initial sample) is determined in the feasible solution space. Then, a direction of the initial point is found by a numerical analysis method, and the step size is adjusted, so that the prediction result output from the trained prediction model is advanced to the next point (relay sample) that can make the prediction result even closer to the target result. The above direction search and step size adjustment are repeated, and when the prediction result output from the trained prediction model is closest to or matches the target result, that point is used as the input parameter. It should be noted that the method of determining the initial point or initial sample in the above-mentioned multi-dimensional subspace can be specified by the members of the federated learning, or by the similar steps in the specific embodiment of FIG. 3, first find a sample that is close or coincident with the confirmed decision variable and set it as the initial sample.

[0052] Please refer to Figure 5. Figure 5 shows a flowchart of the steps of a calculation method for decision variables according to another specific embodiment of the present invention. As shown in Figure 5, the calculation method for decision variables in this specific embodiment further includes the following steps:

[0053] Step S120: The members of the federated learning set up a dummy layer that is connected to the input terminal of the trained prediction model to form a parameter prediction model. The dummy layer becomes the input terminal of the parameter prediction model.

[0054] Step S122: Set the output of the parameter prediction model as a target result, input a training data set including at least one all-one vector to the parameter prediction model to train the parameter prediction model, and adjust multiple arc weight values ​​between the dummy layer and the trained prediction model by an optimizer in the machine learning method that trained the trained prediction model, and set these arc weight values ​​as input parameters of the trained prediction model.

[0055] It should be noted that in addition to the above steps, the method of this specific embodiment further includes step S10 and steps S14 to S18 in the specific embodiment described above, but since the contents of these are the same as the corresponding steps in the specific embodiment described above, they will not be described in detail here.

[0056] The dummy layer newly added in step S120'' is used to form a parameter prediction model, which is an extension model, together with the original trained prediction model. In practice, the dummy layer includes the same number of artificial neurons as the input points in the input layer of the trained prediction model. The artificial neurons can be connected to these input points, respectively. In addition, the bias value of the activation function of each artificial neuron is set to 0. As a result, when the input value of the activation function is 1, the output is also 1. It should be noted that although the trained prediction model forms a parameter prediction model together with the dummy layer, the model parameters of the trained prediction model are fixed in the subsequent training. In addition, when 1 is input to the artificial neuron, the output is also 1, so when the output of the parameter prediction model is a fixed result (for example, set to the above target result), the optimizer adjusts the input of the trained prediction model by adjusting the arc weights between each artificial neuron and each input point of the trained prediction model. Therefore, in step S122'', the output of the parameter prediction model is set to the target result, and the arc weights obtained by the operation of inputting a training data set including an all-ones vector and training can be used as the input parameters of the trained prediction model. Similarly, in practice, one may first find a sample that approximates or matches the verified decision variables by following similar steps in the specific embodiment of FIG. 3, and use the sample parameters as the initial point for the optimizer to adjust the arc weights.

[0057] In practice, the types of optimizers used include Adaptive Moment Estimation (Adam) optimization, Stochastic Gradient Descent (SGD), Momentum, Nesterov Accelerated Gradient (NAG), Adaptive Gradient Algorithm (AdaGrad), Nesterov-accelerated Adaptive Moment Estimation (Nadam), Root Mean Square Propagation (RMSprop), Adaptive Delta (Adadelta), AdamW (Adam with Weight Decay), Adaptive Moment Estimation with Long-term Memory (AMSGrad), Adaptive Belief (AdaBelief), Layer-wise Adaptive Rate Scaling (LARS), Self-adaptive Hessian (AdaHessian), Rectified Random Algorithm (RAdam), and others. The optimizer may be Adam, Lookahead, MadGrad (Momentumized, Adaptive, and Decentralized Gradient Descent), Yogi optimizer, and AdamMax (Adaptive Moment Estimation with Maximum).

[0058] Please refer to Figure 6. Figure 6 shows a flowchart of the steps of a calculation method for decision variables according to another specific embodiment of the present invention. As shown in Figure 6, the calculation method for decision variables in this specific embodiment further includes the following steps:

[0059] Step S120': Each member of the federated learning compares all samples in the local dataset with the target result to obtain multiple anchor samples, and forms a multidimensional subspace with these anchor samples.

[0060] Step S122'): Determine an initial sample in the multidimensional subspace.

[0061] Step S124': The optimizer that generated the trained prediction model performs a minimization operation on the objective function of the trained prediction model to obtain a gradient, and sets the opposite direction of the gradient to the direction of the initial sample.

[0062] Step S126': For the initial samples, form relay samples by increasing the step size along the direction, and input the relay samples into the trained prediction model to check whether the generated prediction result is close to the target result.

[0063] Step S128': The operation of inputting the relay samples whose direction has been determined and whose step size has been increased into the trained prediction model is continued until the prediction result generated by the trained prediction model matches the target result, and the last relay sample is used as an input parameter of the trained prediction model.

[0064] It should be noted that in addition to the above steps, the method of this specific embodiment further includes step S10 and steps S14 to S18 in the specific embodiment described above, but since the contents of these are the same as the corresponding steps in the specific embodiment described above, they will not be described in detail here.

[0065] In this specific embodiment, the method of calculating the decision variables is the same as the above specific embodiment, in that an anchor sample is first found to form a feasible solution space (multidimensional subspace), and an initial sample is pre-determined within this feasible solution space. However, this specific embodiment differs from the above specific embodiment in the following respects. That is, the direction of the initial sample or the subsequent relay sample in this specific embodiment is determined by calculating it with an optimizer that trains the trained prediction model. In detail, the optimizer can be used to obtain a gradient by performing a minimization operation on the objective function representing the trained prediction model, and the opposite direction of the gradient can be the direction of the initial sample or the subsequent relay sample. Then, using the direction calculated by the optimizer, the initial sample or the relay sample is advanced in appropriate increments to approach the target result. Then, when the trained prediction model is advanced until the output of the trained prediction model matches the target result, the relay sample can be used as an input parameter of the trained prediction model. Also, similar to the specific embodiment described above, the method for determining the initial point or initial sample in the multidimensional subspace in this specific embodiment may be to first find a sample that is close to or matches the confirmed decision variable and use it as the initial sample through similar steps to the specific embodiment of FIG.

[0066] According to each of the above specific embodiments, the members of the federated learning can calculate corresponding input parameters (e.g., the above-mentioned input parameters I1 and I2) by various inverse estimation methods based on the target results provided by the inverse estimation requester. However, as described above, each member of the federated learning has a different local data set, a different machine learning algorithm to be used, a different ability to adjust model parameters, and a different inverse estimation method and ability to be used. Therefore, the inverse estimation requester must first select a trained predictive model that meets its requirements before it can achieve efficient and accurate calculation of decision variables.

[0067] Please refer to Figure 7. Figure 7 shows a flowchart of the steps of a calculation method for decision variables according to another specific embodiment of the present invention. As shown in Figure 7, the calculation method for decision variables in this specific embodiment further includes the following steps:

[0068] Step S20: Providing a test sample, the test sample including verified test input parameters and test target results.

[0069] Step S22: Input the test input parameters into a trained prediction model candidate, and the trained prediction model candidate whose output result matches the test target result is set as the trained prediction model.

[0070] It should be noted that in the method of calculating the decision variables in this specific embodiment, after the above steps are completed, other steps in the specific embodiment described above (e.g., steps S10 to S18, etc.) are subsequently executed, and finally, the decision variables requested by the inverse estimation requester are calculated. Note that these steps have been described in detail in the specific embodiment described above, so they will not be described in detail again here.

[0071] In step S20, the reverse estimation requester may provide a sample currently held in which the decision variables and results have been completely confirmed as a test sample, and use it to verify whether the trained prediction models of all members of the federated learning (defined as trained prediction model candidates in this specific embodiment) are as required by the reverse estimation requester. Both the test input parameters and the test target result in the test sample are specified. Therefore, in step S22, the trained prediction model candidate that receives the test input parameters and outputs a result that matches the test target result has a prediction accuracy that meets the requirements of the reverse estimation requester, and can therefore be used as the trained prediction model to be used for subsequent reverse estimation of decision variables.

[0072] In addition, the method for calculating decision variables in the present invention also requires that the members of the federated learning involved in the calculation be asked to perform inverse estimation, as in step S12 and its sub-steps in the specific embodiment described above. Therefore, the members of the federated learning are also asked by the inverse estimation requester for their ability to perform inverse estimation. Now, reference is made to FIG. 8. FIG. 8 shows a flowchart of the steps of the method for calculating decision variables based on another specific embodiment of the present invention. As shown in FIG. 8, the method for calculating decision variables in this specific embodiment further includes the following steps.

[0073] Step S20': Provide a first test sample, which includes a verified first test input parameter and a first test target result.

[0074] Step S22': The first test input parameters are input to the trained prediction model candidate, and the trained prediction model candidate whose output result coincides with the first test target result is set as the first trained prediction model candidate.

[0075] Step S24: Provide a second test sample, the second test sample including the verified second test input parameters and second test target results.

[0076] Step S26: provide the first trained prediction model candidate with the second test target result respectively, and allow the first trained prediction model candidate to inversely estimate the first input parameter candidate.Then, the first trained prediction model candidate whose first input parameter candidate matches or is close to the second test input parameter or whose first input parameter candidate satisfies the validity verification process is set as the trained prediction model.

[0077] It should be noted that in the method of calculating the decision variables in this specific embodiment, after the above steps are completed, other steps in the specific embodiment described above (e.g., steps S10 to S18, etc.) are subsequently executed, and finally, the decision variables requested by the inverse estimation requester are calculated. Note that these steps have been described in detail in the specific embodiment described above, so they will not be described in detail again here.

[0078] This specific embodiment differs from the above specific embodiment in the following points. That is, as shown in steps S20' to S22', in this specific embodiment, first, the test input parameters of the first test sample are provided to all members of the federated learning, and all members input the test input parameters to the trained prediction model (in this specific embodiment, the trained models of all members are defined as trained prediction model candidates), and then return the output prediction result. The reverse estimation requester selects the first trained prediction model candidate based on whether the prediction result matches the first test target result of the first test sample. At this point, the reverse estimation requester can select a prediction model whose accuracy meets the requirements from all trained prediction model candidates participating in the federated learning. Next, the second test target result of the second test sample is provided to the trained prediction model (in this specific embodiment, the first trained prediction model candidate) having high accuracy, and the first input parameter candidate is obtained by performing reverse estimation. Then, the inverse estimation requester compares the first input parameter candidate with the second test input parameters of the second test sample, and selects the first trained predictive model candidate that matches or approximates the second test input parameters as the trained predictive model to be used in the subsequent decision variable calculation step, since the first trained predictive model candidate has accurate inverse estimation capability.

[0079] In addition, in all trained prediction models, the functions that represent them may have many-to-one properties. That is, it may be possible to obtain similar prediction results for multiple different inputs to the trained prediction model. Therefore, even if the first input parameter candidate obtained by the trained prediction model candidate performing reverse estimation based on the second test target result does not match or approximate the second test input parameter of the reverse estimation requester, it is not necessarily an infeasible input parameter. Therefore, the reverse estimation ability of the trained prediction model candidate may be verified by a validity verification process. In practice, this validity verification process can be performed by the reverse estimation requester. For example, if the reverse estimation requester is also one of the members of the federated learning, the first input parameter candidate may be input into its own trained prediction model to verify whether the prediction result matches or approximates the second test target result. In addition, if the reverse estimation requester is not a member of the federated learning or wants to perform multiple validity verifications, the above validity verification process may provide the first input parameter candidate to multiple members of the federated learning and input it into their trained prediction models to verify whether their prediction results match or approximate the second test target result. Therefore, if the first trained predictive model candidate passes the above-mentioned validity verification process, it also indicates that the first trained predictive model candidate has accurate inverse estimation capability, and is selected as the trained predictive model to be used in the subsequent decision variable calculation step.

[0080] In brief, in this specific embodiment, first, by providing a confirmed test input parameter, a trained prediction model and corresponding federated learning members whose output results match expectations are selected, and then, by providing a confirmed target result, a trained prediction model and corresponding federated learning members whose inverse estimation ability matches expectations are selected.

[0081] In summary, the method for calculating decision variables in the present invention can use trained prediction models provided by multiple members of the federated learning to inversely estimate decision variables that match a specific target result without performing federated learning. The inverse estimation requester can first inversely estimate input parameters from the target result using the trained prediction models of each member of the federated learning, provide a loss function, and sum up to form a comprehensive input parameter and a comprehensive loss function. Next, an optimization problem can be constructed using the comprehensive input parameter and the comprehensive loss function to find a solution, and the comprehensive input parameter calculated for the optimization problem can be used as a decision variable. According to the method for calculating decision variables in the present invention, a specific member in the federated learning or a unit outside the federated learning can request inverse estimation. In addition, since it is possible to directly request and select collaborators without going through the federated model generated by the federated learning, it is possible to avoid complex machine learning and inverse estimation processes. In addition, the method for calculating decision variables in the present invention does not require the members of the federated learning to exchange very sensitive and highly confidential data samples and information in the training dataset, and does not require the members of the federated learning to exchange model parameters of the trained prediction models of each member. This not only reduces the complexity of the process, but also further protects the confidentiality of the data.

[0082] The detailed description of the preferred specific embodiments above is intended to more clearly describe the features and spirit of the present invention, and is not intended to limit the scope of the present invention to the preferred specific embodiments disclosed above. Rather, it is intended that various modifications and equivalent configurations be included within the scope of the claims that are to be claimed in the present invention. Therefore, the scope of the claims that are to be claimed in the present invention should be interpreted in the broadest possible manner based on the above description so as to include all possible modifications and equivalent configurations. [Explanation of symbols]

[0083] S10~S18,S180~S184,S120~S122,S120'~S126',S120''~S122'',S120''~S128''',S20~S22,S24~S26,S20'~S22' Process steps

Claims

1. A method for calculating a decision variable, comprising the steps of: providing a plurality of trained predictive models, the trained predictive models being obtained by machine learning a plurality of local datasets respectively using a machine learning method; providing a target outcome and inversely estimating a plurality of input parameters for the target outcome, respectively, with the trained predictive model; a step of setting a plurality of loss functions corresponding to the trained prediction models, respectively, assigning a first weighting value to each of the loss functions, and then summing the loss functions to generate a comprehensive loss function, each of the loss functions being defined as an absolute value, a square of a difference value between a function value of an objective function representing each of the trained prediction models and the target result, or a function having a monotonically increasing characteristic with respect to the difference value; generating a comprehensive input parameter by adding a second weight value to each of the input parameters and then summing the input parameters, and constructing an optimization problem using the comprehensive input parameter and the comprehensive loss function; obtaining the second weight value for each of the input parameters by solving the optimization problem, and calculating the overall input parameter as the decision variable.

2. The step of solving the optimization problem further comprises: calculating a derivative of the overall loss function for each of the second weight values; combining the derivatives to form a direction; 2. The method for calculating a decision variable according to claim 1, further comprising a step of inputting the overall input parameters to the trained predictive model after increasing the step size along the direction, and determining whether the function value of the overall loss function decreases.

3. The method of claim 1 , wherein the trained predictive models are provided by multiple members of a federated learning network, respectively.

4. The step of inversely estimating the input parameters for each of the target outcomes using the trained predictive model further includes: a first one of the members comparing the target outcome with all samples in a corresponding first one of the local data sets; 4. The method of claim 3, further comprising the step of: selecting samples in the first local data set that match the target outcome as reference samples for inversely estimating the input parameters.

5. The step of inversely estimating the input parameters for each of the target outcomes using the trained predictive model further includes: a second one of the members comparing the target result with all samples of a corresponding second one of the local data sets to obtain a plurality of anchor samples, and forming a multidimensional subspace with the anchor samples; determining an initial sample within the multi-dimensional subspace; obtaining a direction of the initial sample in the multi-dimensional subspace, forming relay samples for the initial sample by increasing step sizes along the direction, and inputting the relay samples into a second trained prediction model provided by the second member to determine whether a generated prediction result is close to the target result; 4. The method of claim 3, further comprising the step of: inputting the relay samples for which the direction has been determined and the step size has been increased into the second trained predictive model until the prediction result generated by the second trained predictive model matches the target result, and setting the last relay sample as the input parameter corresponding to the second trained predictive model.

6. The step of inversely estimating the input parameters for each of the target outcomes using the trained predictive model further includes: A third member among the members sets a dummy layer connected to an input terminal of a corresponding third trained prediction model to form a parameter prediction model, and the dummy layer becomes an input terminal of the parameter prediction model; 4. The method for calculating a decision variable according to claim 3, further comprising the steps of: setting an output of the parameter prediction model as the target result; inputting a training dataset including at least one all-ones vector to the parameter prediction model to train the parameter prediction model; adjusting a plurality of arc weight values ​​between the dummy layer and the third trained prediction model by an optimizer in a machine learning method that trains the third trained prediction model; and setting the arc weight values ​​as the input parameters corresponding to the third trained prediction model.

7. The step of inversely estimating the input parameters for each of the target outcomes using the trained predictive model further includes: a fourth member of the members comparing the target result with all samples of a corresponding fourth local data set of the local data sets to obtain a plurality of anchor samples, and forming a multidimensional subspace with the anchor samples; determining an initial sample within the multi-dimensional subspace; A step of obtaining a gradient by performing a minimization operation on the objective function of the fourth trained predictive model using an optimizer used by the fourth member to generate a fourth trained predictive model, and setting the opposite direction of the gradient as the direction of the initial sample; forming relay samples by increasing step sizes along the direction for the initial samples, and inputting the relay samples into the fourth trained prediction model to determine whether a generated prediction result is close to the target result; 4. The method for calculating decision variables according to claim 3, further comprising the step of: continuing to input the relay samples for which the direction has been determined and the step size has been increased into the fourth trained predictive model until the prediction result generated by the fourth trained predictive model matches the target result, and setting the last relay sample as the input parameter corresponding to the fourth trained predictive model.

8. The step of inversely estimating the input parameters for each of the target outcomes using the trained predictive model further includes: providing a plurality of validated decision variables and calculating a first vector of the validated decision variables; obtaining corresponding sample parameters from sample parameters in all samples of the fourth local data set based on the confirmed decision variables, respectively, and calculating a second vector of the corresponding sample parameters; obtaining a reference sample by comparing the first vector with the second vector and comparing the predicted results generated by all samples of the fourth local dataset in the fourth trained predictive model with the target result, the second vector of the reference sample being close to or identical to the first vector, and the predicted results of the reference sample being close to or identical to the target result; The method of claim 7 , further comprising the step of: making the reference sample the initial sample.

9. Furthermore, providing a test sample, the test sample including a plurality of validated test input parameters and test objective outcomes; 2. The method for calculating a decision variable according to claim 1, further comprising the step of inputting the test input parameters to a plurality of trained predictive model candidates, and determining the trained predictive model candidate whose output result matches the test target result as the trained predictive model.

10. Furthermore, providing a first test sample, the first test sample including a plurality of verified first test input parameters and a first test objective outcome; inputting the first test input parameters into a plurality of trained predictive model candidates, and determining the trained predictive model candidates whose output results match the first test target results as a plurality of first trained predictive model candidates; providing a second test sample, the second test sample including a plurality of verified second test input parameters and a second test objective outcome; 2. The method for calculating decision variables according to claim 1, further comprising the step of: providing the second test target results to the first trained predictive model candidate, respectively, and having the first trained predictive model candidate inversely estimate a plurality of first input parameter candidates, and determining the first trained predictive model candidate in which the first input parameter candidates match or are close to the second test input parameters or the first input parameter candidates satisfy a validation process as the trained predictive model.

Images