Calculation method of decision variable

Abstract

To provide a calculation method of a decision variable for calculating a checkup waiting decision variable.SOLUTION: A calculation method of a decision variable first provides a trained prediction model generated by training a data set by a machine learning method. Subsequently, the method converts an objective function of the trained prediction model into an unconstrained objective function from a constrained objective function, and obtains a solution of an optimization problem for the unconstrained objective function. The step of obtaining the solution calculates a gradient so as to easily obtain the solution by an optimizer when the trained prediction model is trained. An initialization sample when the solution of the optimization problem is obtained can be determined by using a sample in the dataset when the trained prediction model is trained. The calculation method of the decision variable adds a dummy layer in front of the trained prediction model, and starts to acquire an arc weighted value between the dummy layer and an input layer from the initialization sample as a checkup waiting determination variable, by the optimizer.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a method for calculating decision variables, and in particular to an algorithm based on optimization techniques, an optimizer embedded in an artificial intelligence model, and a method for inversely estimating decision variables by applying a trained predictive model. [Background technology]

[0002] With the progress of medical engineering technology, the application of regenerative medicine in clinical treatment of diseases is becoming more and more diverse. Regenerative medicine is a medical technology that mainly uses the regenerative ability of cells to repair damaged tissues and organs, and has a very wide range of applications. In addition, by combining medical technologies such as tissue engineering and molecular biology, it is expected that diseases that were once difficult to treat, such as diabetes, neurological diseases, cardiovascular diseases, and cancer, can be improved and treated. Currently, regenerative medicine is mainly applied to organ repair, immune cell therapy, and stem cell therapy. In addition, research and applications of cell therapy in regenerative medicine are attracting more and more attention from various fields. Cell therapy involves the transplantation of human body cells into an individual's body after a procedure of culturing or processing the cells outside the body.

[0003] In recent years, many governments have gradually lifted the ban on the application of cell therapy, and many scholars at home and abroad have been entering the research of cell therapy one after another. As a result, cell therapy has made remarkable progress in the treatment of many diseases, such as the treatment of skin defects with autologous fibroblasts, the treatment of knee joint cartilage defects with autologous chondrocytes, and the treatment of spinal cord injury with autologous bone marrow mesenchymal stem cells. In addition, the quality of cell therapy products has a direct impact on the safety and efficacy of the treatment. Therefore, in the cell culture process, it is necessary to strictly control the cell growth state and monitor the culture parameters and environmental parameters of cell growth in real time to avoid cell contamination or deterioration of quality during the culture process. In addition, previous research has shown that the variability of cells is extremely high between different cases, so the optimal culture parameters and environmental parameters of cell preparations applied to different cases are not completely the same. Therefore, it is necessary to design and adjust each process parameter in the process for each cell preparation to achieve the desired result, and therefore each cell preparation cannot be produced with fixed process parameters. In addition, due to the complexity and interrelationship of the process, the prior art usually only allows optimization of a single parameter, while neglecting the overall mutual influences between each parameter in the entire process.

[0004] In the prior art, a prediction model is obtained by training on a large number of sample data sets using machine learning. This prediction model can generate predicted results of cell culture by inputting various process parameters, so that a user can simulate the effect of the cell culture process to be designed in advance. However, the above prediction model trained by machine learning cannot reverse-estimate the process parameters of the cell culture process from the user's desired results. In other words, when designing a cell culture process, a user needs to try a large number of different process parameters, so a large amount of effort and resources are spent on designing and improving the cell culture process. In addition, there is a problem that the decision variables cannot be reverse-estimated for prediction models obtained by machine learning in other fields besides cell processes.

[0005] Therefore, to address the problems of the prior art, it is necessary to develop a method that can use a trained predictive model to accurately inversely estimate decision variables that match a target outcome. Summary of the Invention [Problem to be solved by the invention]

[0006] In view of the above, the present invention has an object to provide a method for calculating decision variables in order to solve the above-mentioned problems in the prior art. [Means for solving the problem]

[0007] The method for calculating decision variables for calculating a plurality of decision variables to be confirmed in the present invention includes the steps of: providing a trained prediction model, the trained prediction model being obtained by machine learning a dataset through a machine learning method, the trained prediction model including an input layer and an output layer, the dataset including a plurality of samples, and each of the samples including a plurality of sample parameters; the trained prediction model being used to input a plurality of input variables through the input layer and generate a prediction result corresponding to the input variables through the output layer; setting a target result corresponding to the prediction result of the trained prediction model; and selecting at least one sample among the samples, in which the prediction result of the trained prediction model is consistent with the target result. as at least one anchor sample, forming a multidimensional subspace using the at least one anchor sample, and determining an initial sample in the multidimensional subspace; obtaining a gradient by minimizing an objective function of the trained prediction model using an optimizer used in generating the trained prediction model; increasing a step size of the sample parameters of the initial sample in a direction opposite to the gradient, and then inputting the sample parameters to the trained prediction model to check whether the generated prediction result matches the target result; and if the prediction result matches the target result, setting the sample parameters after increasing the step size as the decision variables awaiting confirmation.

[0008] The method for calculating decision variables further includes the steps of providing a plurality of confirmed decision variables and calculating a first vector of the confirmed decision variables; respectively obtaining a plurality of corresponding sample parameters from the sample parameters of the sample based on the confirmed decision variables and calculating a second vector of the corresponding sample parameters; obtaining a reference sample from the samples by comparing the first vector with the second vector of each of the samples and comparing the predicted result and the target result generated by each of the samples in the trained prediction model, where the second vector of the reference sample is close to or consistent with the first vector, and the predicted result of the reference sample is close to or consistent with the target result; and setting the reference sample as the initial sample.

[0009] The present invention further provides a trained prediction model, the trained prediction model being obtained by machine learning a dataset by a machine learning method, the dataset including a plurality of samples, each of the samples including a plurality of sample parameters, the trained prediction model being used to input a plurality of decision variables to generate a prediction result corresponding to the decision variables, the decision variables including a plurality of confirmed decision variables and a plurality of decision variables to be confirmed; setting a target result corresponding to the prediction result of the trained prediction model; comparing the confirmed decision variables with a plurality of corresponding sample parameters corresponding to the confirmed decision variables among the sample parameters, respectively, to select a candidate sample whose corresponding sample parameters are close to or consistent with the confirmed decision variables; inputting other sample parameters not belonging to the corresponding sample parameters in the candidate sample together with the confirmed decision variables into the trained prediction model, and the prediction result output from the trained prediction model is compared with the previous step. performing the above steps for all sample parameters in the dataset, and selecting at least one sample among the samples, for which the prediction result of the trained prediction model matches the target result, as at least one anchor sample, and forming a multidimensional subspace using the at least one anchor sample; determining an initial sample in the multidimensional subspace and obtaining a direction of the initial sample in the multidimensional subspace; after increasing a step size along the direction, inputting the confirmed decision variable and other sample parameters in the initial sample that do not correspond to the confirmed decision variable into the trained prediction model to check whether the generated prediction result matches the target result; and, if the prediction result matches the target result, selecting the sample parameters after increasing the step size that do not correspond to the confirmed decision variable as the decision variables awaiting confirmation.

[0010] The step of selecting the candidate sample whose corresponding sample parameters approximate or match the confirmed decision variables by comparing the confirmed decision variables with the corresponding sample parameters among the sample parameters that correspond to the confirmed decision variables further includes the steps of calculating a first vector of the confirmed decision variables, obtaining the corresponding corresponding sample parameters from the sample parameters of the samples based on the confirmed decision variables, calculating a second vector of the corresponding sample parameters, and comparing the first vector with the second vector of each of the samples, and selecting the first sample as the candidate sample if the second vector of a first sample among the samples approximates or matches the first vector.

[0011] The method for calculating decision variables further includes a step of calculating an included angle between the first vector and the second vector of each of the samples, and determining as the first sample the sample corresponding to the second vector that has the smallest included angle with the first vector.

[0012] The method for calculating the decision variables further includes a step of calculating a distance between the coordinates of the end point of the first vector and the coordinates of the end point of the second vector of each of the samples using a distance function, and determining the sample corresponding to the second vector having the smallest distance between itself and the first vector as the first sample.

[0013] Determining the initial sample in the multi-dimensional subspace further includes determining the initial sample as a midpoint of a distance between at least two of the anchor samples in the multi-dimensional subspace.

[0014] Determining the initial sample in the multi-dimensional subspace further includes determining the initial sample as a linear combination of at least two of the anchor samples in the multi-dimensional subspace.

[0015] The step of obtaining the orientation of the initial sample in the multidimensional subspace further includes a step of calculating, by a numerical analysis method, a first directional derivative of the initial sample along each of the anchor samples as the orientation of the initial sample.

[0016] The step of obtaining the orientation of the initial sample within the multidimensional subspace further includes a step of setting the initial sample to an origin and calculating, by the numerical analysis method, a second directional derivative of the origin toward each of the anchor samples as the orientation of the initial sample.

[0017] The trained prediction model is represented by an objective function, and obtaining the orientation of the initial sample in the multi-dimensional subspace further includes obtaining a secant directional derivative of the initial sample toward each of the anchor samples by calculating a change in a function value of the objective function based on a change in a distance of the initial sample toward each of the anchor samples through the numerical analysis method.

[0018] The step of obtaining the direction of the initial sample in the multidimensional subspace further includes the steps of: sequentially increasing and adjusting the step size of the initial sample toward the directional derivative of each of the anchor samples; sequentially checking whether the prediction result input to the trained prediction model and output after increasing and adjusting the step size of the initial sample toward the directional derivative of each of the anchor samples is closer to the target result; and sequentially checking whether the prediction result input to the trained prediction model and output after increasing the step size of the initial sample toward the directional derivative of any of the anchor samples is closer to the target result. the step of increasing the step size of the initial sample in the directional derivative if it is confirmed that the prediction result becomes closer to the target result; and the step of increasing the step size of the initial sample toward the directional derivative of one of the anchor samples if it is confirmed that the prediction result input to the trained prediction model and output does not become closer to the target result, the step of increasing the step size of the initial sample toward the directional derivative of another of the anchor samples, and then checking again whether the prediction result input to the trained prediction model and output becomes closer to the target result.

[0019] The present invention further provides a method for calculating a plurality of decision variables to be confirmed, the method for calculating the decision variables includes the steps of: providing a trained prediction model, the trained prediction model being obtained by machine learning a dataset using a neural network method, the trained prediction model including an input layer and an output layer, the number of input terminals of the input layer being equal to the total number of the decision variables to be confirmed and a plurality of confirmed decision variables, the dataset including a plurality of samples, each of the samples including a plurality of sample parameters; comparing the confirmed decision variables with a plurality of corresponding sample parameters among the sample parameters corresponding to the confirmed decision variables, respectively, to select a first sample whose corresponding sample parameters are approximate or match the confirmed decision variables, inputting the confirmed decision variables together with other sample parameters not belonging to the corresponding sample parameters in the first sample into the trained prediction model, and if the prediction result output from the trained prediction model matches the target result, setting the other sample parameters as initial weight values; and inputting the trained prediction model into the trained prediction model. a dummy layer is added to be connected to the input layer, the dummy layer includes a plurality of artificial neurons respectively connected to each input terminal of the input layer, the trained prediction model and the dummy layer form a parameter prediction model, and the prediction result of the trained prediction model is the output of the parameter prediction model; a bias value of an activation function of each of the artificial neurons is set to 0, and when an input value of the activation function is 1, an output value of the activation function is set to 1; a plurality of first weight values ​​between a plurality of first artificial neurons among the artificial neurons corresponding to the confirmed decision variables and the input terminals are set based on the confirmed decision variables; an output of the parameter prediction model corresponding to the target result is set; and a training data set including at least one all-one vector is input to the artificial neurons of the dummy layer of the parameter prediction model to train the parameter prediction model, and the trained prediction model is generated by an optimizer in the neural network method based on the target result,adjusting a reference weight value between each of the artificial neurons and the input layer based on the initial weight value, and setting a plurality of second reference weight values ​​corresponding to the initial weight value among the reference weight values ​​as the confirmation waiting decision variables.

[0020] The trained predictive model includes a number of model parameters, all of which are fixed.

[0021] The first weight value is fixed.

[0022] The optimizers 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), Adam with Weight Decay (AdamW), 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 optimization algorithm is selected from among Adam, Lookahead, MadGrad (Momentumized, Adaptive, and Decentralized Gradient Descent), Yogi optimizer (Yogi), and AdamMax (Adaptive Moment Estimation with Maximum). Effect of the Invention

[0023] In summary, the method for calculating decision variables in the present invention can calculate decision variables that match the target result or ground truth by performing inverse estimation on a prediction model trained and acquired by machine learning. In detail, the method for calculating decision variables in the present invention can convert the constrained objective function of the trained prediction model into an unconstrained objective function and optimize the unconstrained objective function to obtain a solution. In addition, in the process of obtaining a solution, it is possible to set anchor samples to form a multidimensional feasible solution space and search for an optimal solution within this space. When obtaining a solution, an initial solution, a direction, and a step size can be taken into consideration. The initial solution can be obtained from the sample set used when training the trained prediction model, and for example, a sample having a sample parameter that is close to or matches the confirmed decision variable is taken as the initial solution. In addition, the optimizer of the platform on which the trained prediction model is trained can obtain the gradient of the initial solution or any point in the feasible solution space, and the reverse direction of the gradient can be taken as the direction. Finally, starting from the initial solution, the direction and step size are continuously adjusted to calculate an optimal solution that matches the target result as a decision variable waiting for confirmation. In addition, the method for calculating the decision variables in the present invention may use a numerical method to calculate the unconstrained objective function when determining the direction. In addition, the method for calculating the decision variables in the present invention may use an optimizer to directly train the trained prediction model to which a dummy layer has been added. Then, the trained prediction model is trained using samples in the sample set used in training the trained prediction model as initial weight values ​​to obtain arc weights between the dummy layer and the trained prediction model, and the arc weights are used as decision variables. [Brief description of the drawings]

[0024] [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 S3 of the method for calculating the decision variables according to 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. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0025] 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.

[0026] Please refer to FIG. 1. FIG. 1 shows a flowchart of steps of a method for calculating decision variables according to a specific embodiment of the present invention. The method for calculating decision variables in this specific embodiment is used to calculate a number of decision variables to be confirmed. Among the decision variables, the confirmed decision variables are decision variables that have been confirmed by the user during operation or have actually been executed. Meanwhile, the decision variables to be confirmed are decision variables that have not been confirmed or executed by the user. As shown in FIG. 1, in this specific embodiment, the method for calculating decision variables includes the following:

[0027] Step S1: Provide a trained prediction model obtained by machine learning a dataset through a machine learning method. The dataset includes a number of samples. Each sample includes a number of sample parameters. The input layer of the trained prediction model is used to input a number of input variables and output and generate a prediction result corresponding to the input variables.

[0028] Step S2: Set a target result corresponding to the predicted result of the trained predictive model.

[0029] Step S3: At least one sample in the dataset whose prediction result output from the trained prediction model coincides with the target result is set as at least one anchor sample, and a multidimensional subspace is formed by the at least one anchor sample, and an initial sample is determined in the multidimensional subspace.

[0030] Step S4: The optimizer used to generate the trained predictive model obtains a gradient by minimizing the objective function of the trained predictive model.

[0031] Step S5: For the sample parameters of the initial sample, the step size is increased along the opposite direction of the gradient, and then input into the trained prediction model to check whether the generated prediction result is consistent with the target result.

[0032] Step S6: If the predicted result matches the target result, the sample parameters after increasing the step size are set as the confirmation waiting decision variables.

[0033] The trained prediction model in step S1 in this specific embodiment may be any model that has completed machine learning training and is available on an open platform, or may be a prediction model that has been trained by the user himself. The dataset in this specific embodiment may be a set formed of any data that can be used for machine learning training, testing, and validation. In practical applications, the calculation method of the decision variables in this specific embodiment may be applied to any other trained machine learning model for inverse estimation of the decision variables waiting for confirmation. The trained prediction model in this specific embodiment is obtained by machine learning the dataset using Artificial Neural Networks (ANN), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or any other machine learning or neural network algorithm. The machine learning or neural network algorithm is selected according to the needs of the user.

[0034] In step S2, first, the user can set a target result, also called ground truth. This is a goal to be achieved by the decision variables calculated by the decision variable calculation method of the present invention. In the decision variable calculation method, inverse estimation is performed based on a trained prediction model, which can be expressed by an objective function. Therefore, the process of obtaining decision variables by performing inverse estimation on the trained prediction model can be considered as finding a solution to an optimization problem for the objective function. In finding a solution to an optimization problem, it is more efficient if a feasible solution space is obtained in advance and a solution can be found within that space. Therefore, in step S3, the feasible solution space can be formed using samples when the trained prediction model is trained. In detail, for the samples in the data set, the predicted results that are output after inputting them into the trained prediction model are known. Among the samples in the data set, samples whose predicted results match the target results have considerable reference value in finding a solution for the decision variables. Therefore, a multidimensional subspace can be formed using these samples as anchor samples, and this multidimensional subspace is the feasible solution space. After the multidimensional subspace is determined, an initial point is determined in this multidimensional subspace, and the solution to the optimization problem is started to be found. This initial point becomes the initial sample in step S3.

[0035] As described above, after determining the initial sample in the feasible solution space, it is sufficient to start searching for a solution to the optimization problem from the initial sample. In detail, after proceeding by one step size in one direction from the initial sample, the sample parameters with the increased step size are input to the trained prediction model to check whether the output prediction result matches the target result or whether the gap with the target result becomes smaller. As described in steps S5 and S6, by repeating the above-mentioned step size progression, the input parameters are finally obtained such that the output prediction result completely matches the target result or the gap with the target result is minimized, and these can be used as decision variables to be confirmed. However, when determining the direction of each point in the process of searching for the solution to the initial sample or the optimization problem, the gradient can generally be obtained by a method of performing partial differentiation on the objective function, and the opposite direction of the gradient is set as the direction. However, the objective function of the trained prediction model is usually not symbolic, and in many cases cannot be differentiated, and may even be discontinuous. Therefore, it is difficult to directly calculate the derivative of this objective function, and the gradient can be analyzed by changing to a numerical analysis method. However, when the derivative (or gradient) of the objective function is analyzed by the numerical analysis method, a lot of computational resources are consumed, and problems such as overhead and inaccuracy in the numerical calculation occur. Therefore, in step S4 of the method of this specific embodiment, the gradient can be obtained by minimizing the objective function of the trained prediction model using an optimizer that is originally used to generate the trained model. For example, it is possible to calculate the gradient in the process of completing the minimization of the objective function using an optimizer of the Stochastic Gradient Descent (SGD). Compared to the numerical analysis method, calculating the gradient using an optimizer makes it possible to accurately calculate the gradient and significantly save computational resources.

[0036] In the above specific embodiment, the feasible solution space can be formed by samples of the data set. That is, samples whose predicted results match or are close to the target results are selected as anchor samples. However, in practice, not all decision variables are necessarily decision variables waiting for confirmation, and there may be a situation where confirmed decision variables and decision variables waiting for confirmation exist at the same time. For example, a user may be performing a 21-day cell culture process, and the cell preparation to be produced must have a cell viability of 95%, and the user has already performed the first 7 days of the process (i.e., already has confirmed decision variables for 7 days), but has not confirmed how to adjust the process parameters (i.e., decision variables waiting for confirmation) for the subsequent 8th to 21st days to obtain the desired target results on the pre-set 21st day. If only samples whose predicted results match or are close to the target results are selected as anchor samples, and initial samples are randomly selected within the feasible solution space formed by the anchor samples to find a solution to the optimization problem, the decision variables for the first 7 days finally obtained will usually be significantly different from the confirmed decision variables.

[0037] Now, refer to Fig. 2. Fig. 2 shows a flow chart of further steps of step S3 of the method for calculating decision variables based on Fig. 1. As shown in Fig. 2, this specific embodiment is different from the above-mentioned specific embodiment in the following points. That is, the method of this specific embodiment further includes:

[0038] Step S30: Provide a plurality of confirmed decision variables and calculate a first vector of the confirmed decision variables.

[0039] Step S32: According to the confirmed decision variables, obtain a plurality of corresponding sample parameters from the sample parameters of the samples respectively, and calculate a second vector of the corresponding sample parameters.

[0040] Step S34: Obtain a reference sample from each sample by comparing the first vector with the second vector of each sample, and comparing the predicted result generated by each sample in the trained prediction model with the target result, where the second vector of the reference sample is close to or coincides with the first vector, and the predicted result is close to or coincides with the target result.

[0041] Step S36: The reference sample is set as the initial sample.

[0042] In the step of determining the reference sample, the determination is made based on whether the second vector is close to or coincident with the first vector. In practice, when determining whether they are close to or coincident with each other, it may be confirmed whether the included angle between the two vectors is close to or equal to 0 degrees. In detail, in a specific embodiment, the step of determining the reference sample may include a step of calculating an included angle between the first vector and the second vector of each sample, and obtaining a sample corresponding to the second vector having the smallest included angle (close to or equal to 0 degrees) between the first vector as the reference sample. Alternatively, in another specific embodiment, the step of determining the reference sample may include a step of calculating a distance between the coordinates of the end point of the first vector and the coordinates of the end point of the second vector by a distance function, and obtaining a sample corresponding to the second vector having the smallest distance between the first vector as the reference sample. In practical applications, the distance functions commonly used in the field of machine learning include Euclidean Distance, Manhattan Distance, Cosine Similarity, Jaccard Similarity, and Mahalanobis Distance. In practice, the calculation method of the minimum distance and the application scope of the minimum distance can be set according to the needs of the user.

[0043] The fact that the vector formed by the reference sample parameters of the reference sample corresponding to the confirmed decision variables (e.g., the cellular process parameters for the first 7 days of execution) is close to or coincides with the vector of the confirmed decision variables means that the reference sample parameters of the reference sample are close to or coincident with the confirmed decision variables. Therefore, the optimal solution obtained by solving the above optimization problem using this reference sample as the initial sample is applicable to the situation where some decision variables have been confirmed (e.g., the cellular process parameters for the first 7 days) and some decision variables are waiting to be confirmed (e.g., the cellular process parameters for the 8th to 21st days).

[0044] Now, please refer to Fig. 3. Fig. 3 shows a flow chart of steps of a method for calculating decision variables according to another embodiment of the present invention. As shown in Fig. 3, this embodiment is different from the above-mentioned embodiment in the following points. That is, the method for calculating decision variables in this embodiment further includes:

[0045] Step S300: By comparing a confirmed decision variable among the decision variables with a plurality of corresponding sample parameters corresponding to the confirmed decision variable among the sample parameters, a candidate sample whose corresponding sample parameters are close to or consistent with the confirmed decision variable is selected.

[0046] Step S302: Other sample parameters not belonging to the corresponding sample parameters in the candidate sample are input into the trained prediction model together with the confirmed decision variables, and it is confirmed whether the prediction result output from the trained prediction model is consistent with the target result.

[0047] Step S304: The above steps are performed for all sample parameters in the dataset, and among all the samples, at least one sample whose prediction result of the trained prediction model matches the target result is selected as an anchor sample, and a multidimensional subspace is formed using the anchor samples.

[0048] Step S306: Determine an initial sample in the multi-dimensional subspace, and obtain the direction of the initial sample.

[0049] Step S50: After increasing the step size along the direction, the confirmed decision variables and other sample parameters that do not correspond to the confirmed decision variables in the initial sample are input into the trained prediction model to check whether the generated prediction result is consistent with the target result.

[0050] Step S60: If the predicted result coincides with the target result, the sample parameters after the step size is increased that do not correspond to the confirmed decision variables are set as decision variables awaiting confirmation.

[0051] It should be noted that the steps in the method of this specific embodiment that are the same as or correspond to those in the above specific embodiment are described in detail in the above specific embodiment, and therefore will not be described in detail again here.

[0052] This specific embodiment differs from the above specific embodiment in the following points. That is, in this specific embodiment, first, a sample having a corresponding sample parameter is selected as a candidate sample using the confirmed decision variable. Next, the confirmed decision variable and other sample parameters (i.e., the part waiting for confirmation) in the candidate sample that do not correspond to the confirmed decision variable are combined and input into the trained prediction model. Then, if the prediction result output from the trained prediction model matches the target result, it means that the sample parameter corresponding to the part waiting for confirmation of this candidate sample has reference value, so this candidate sample can be used as an anchor sample for forming a feasible solution space (multidimensional subspace). In this case as well, as shown in step S306, first, it is necessary to start solving the optimization problem from an initial sample in the multidimensional subspace, and it is also necessary to determine the direction and step size. In this specific embodiment, when determining the direction in step S306, the gradient of the objective function is calculated by a numerical analysis method, and the opposite direction of the gradient is set as the direction. Also, since the confirmed decision variable is fixed, in step S50, the other sample parameters in the initial sample are increased in step size along the direction, and then combined with the confirmed decision variable and input into the trained prediction model to confirm whether the predicted result matches the target result. Alternatively, the above-mentioned direction determination and step size increase operations are continued to be performed to continue to bring the predicted result closer to the target result until it matches the target result. Finally, in step S60, the step size is continued to be increased, and if the predicted result matches the target result, or in practical use, if the difference value between the predicted result and the target result becomes the smallest, the sample parameters after the step size increase that do not correspond to the confirmed decision variable are set as decision variables waiting for confirmation.

[0053] In this specific embodiment, the step of determining the candidate sample in step S300 is based on whether the second vector is close to or coincides with the first vector. As described above, the first vector is a vector calculated for the confirmed decision variables, and the second vector is a vector calculated for the corresponding sample parameters corresponding to the confirmed decision variables in the sample. In practice, when determining whether they are close to or coincide with each other, it may be confirmed whether the included angle between the two vectors is close to or equal to 0 degrees. In detail, in a specific embodiment, the step of determining the candidate sample may include a step of calculating an included angle between the first vector and the second vector of each sample, and acquiring a sample corresponding to the second vector having the smallest included angle (close to or equal to 0 degrees) with the first vector as the candidate sample. Alternatively, in another specific embodiment, the step of determining the candidate sample may include a step of calculating a distance between the coordinates of the end point of the first vector and the coordinates of the end point of the second vector by a distance function, and acquiring a sample corresponding to the second vector having the smallest distance with the first vector as the candidate sample. In practical applications, the distance functions commonly used in the field of machine learning include Euclidean Distance, Manhattan Distance, Cosine Similarity, Jaccard Similarity, and Mahalanobis Distance. In practice, the calculation method of the minimum distance and the application scope of the minimum distance can be set according to the needs of the user.

[0054] As described above, in step S306 in this specific embodiment, an initial sample is determined in a multi-dimensional subspace (feasible solution space). In practice, the method for determining the initial sample can be either or a combination of the following steps: the midpoint of the distance between at least two anchor samples in the multi-dimensional subspace is determined as the initial sample; or the linear combination between at least two anchor samples in the multi-dimensional subspace is determined as the initial sample. Each anchor sample is selected from the candidate samples, and its corresponding sample parameters are consistent or approximate to the confirmed decision variables. Therefore, the corresponding sample parameters in the initial samples determined in the above steps are also consistent or approximate to the confirmed decision variables.

[0055] In addition, as described above, when determining the direction for the initial sample or for the subsequent samples after increasing the step size at least once in the multi-dimensional subspace, a gradient is calculated for the objective function by a numerical analysis method, and the direction is the opposite direction of the gradient. In practice, the method for determining the direction can be one of the following steps or a combination thereof: calculating the directional derivative of the initial sample along each anchor sample as the direction of the initial sample (or subsequent sample) by a numerical analysis method; setting the initial sample (subsequent sample) as the origin and calculating the directional derivative of the origin toward each anchor sample as the direction of the initial sample (subsequent sample) by a numerical analysis method; or calculating the change in the function value of the objective function based on the change in the distance of the initial sample (subsequent sample) toward each anchor sample by a numerical analysis method to obtain the secant directional derivative of the initial sample (subsequent sample) toward each anchor sample.

[0056] If the initial samples go in a direction opposite to the feasible solution, the function value may become worse or not match expectations, so the method of determining the direction may also iterate through directions where the function value may be better. According to another specific embodiment, the method for determining a direction may include the steps of: sequentially increasing and adjusting the step size of the initial sample toward the directional derivative of each anchor sample; sequentially checking whether the predicted result input to the trained prediction model and outputted is closer to the target result after increasing and adjusting the step size of the initial sample toward the directional derivative of each anchor sample; if it is confirmed that the predicted result input to the trained prediction model and outputted is closer to the target result after increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, increasing the step size of the initial sample with the directional derivative as a direction; if it is confirmed that the predicted result input to the trained prediction model and outputted is not closer to the target result after increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, increasing the step size of the initial sample toward the directional derivative of another of the anchor samples, and then checking again whether the predicted result input to the trained prediction model and outputted is closer to the target result. Then, after proceeding one step from the initial sample to reach the next subsequent sample or relay sample, the above steps are further executed to search for a new direction and proceed with the next step. In summary, in the method of determining the direction in this specific embodiment, the step size is first repeatedly adjusted from the initial sample toward the direction of the first anchor sample, and it is confirmed whether the predicted result is closer to the target result after the step size is increased. Then, different step sizes are repeatedly tried, and after finding a suitable step size, the step size is advanced along the direction. However, if a suitable step size that makes the predicted result closer to the target result cannot be found even after adjusting the step size a certain number of times, the above adjustment is repeated again for the initial sample toward the direction of the next anchor sample.In practice, if no suitable step size is found in which the function value is better (the difference between the predicted result and the target result is smaller) in all anchor sample directions, the method of calculating the decision variables is stopped. Note that the above steps are cyclical search steps. When the initial sample has progressed to the subsequent sample or intermediate sample in the next feasible solution space, the subsequent sample or intermediate sample may be set as the current solution, and the above steps may be performed again.

[0057] As described above, the method of calculating the decision variables in this specific embodiment obtains the decision variables to be input by solving the optimization problem for the objective function of the trained prediction model. However, there may be a mutual relationship between these decision variables. For example, in a process, each process parameter (decision variable) is not arbitrarily set, but may have various constraint conditions according to the actual situation. For example, there may be inequality or equality constraint conditions between multiple process parameters. Therefore, when constructing and executing an optimization problem, the decision variables finally obtained cannot be adapted to the actual situation unless these constraint conditions are taken into consideration. In this specific embodiment, by adding a barrier function and a penalty function to the objective function of the trained prediction model, the optimization problem that is originally constrained can be converted into an optimization problem without constraints, so there is no need to consider the constraint conditions in the process of finding the solution thereafter. This is because the constraint conditions are directly added to the objective function in the form of a barrier function and a penalty function.

[0058] In this specific embodiment, the barrier function is set to constrain the decision variables to satisfy the inequality constraint. The barrier function includes a barrier region and a barrier function value. When the process parameters satisfy the inequality constraint, the process parameters are outside the barrier region, and the corresponding barrier function value is 0. That is, the prediction result of the trained prediction model is not affected. On the other hand, when the process parameters do not satisfy the inequality constraint, the process parameters are inside the barrier region, and the corresponding barrier function value is large, which greatly affects the prediction result of the trained prediction model.

[0059] In addition, the penalty function is set to restrict the relationship between the decision variables to satisfy the constraint condition of the equality. The penalty function includes a penalty region and a penalty function value. When the relationship between the decision variables satisfies the constraint condition of the equality, the relationship is within the penalty region, and the corresponding penalty function value obtained is 0. On the other hand, when the relationship between the decision variables does not satisfy the constraint condition of the equality, the relationship is outside the penalty region, and the corresponding penalty function value obtained is large, which greatly affects the prediction result of the trained prediction model. Therefore, when sample parameters are input into a trained prediction model having a barrier function and / or a penalty function, not only can it be confirmed whether the generated prediction result matches the target result, but also whether the input parameters satisfy the actual constraint condition. When the input decision variables do not satisfy the conditions of the barrier function and / or the penalty function, the prediction result is significantly dissociated, which means that the decision variables cannot be used because they do not satisfy the conditions.

[0060] The practical constraints mentioned above may be, for example, that in a cell process, the combined antibiotic concentration of Streptomycin and Amphotericin B added to the cell culture medium on day 10 must be 200 μg / mL or less, and the sum of the serum concentration values ​​in the cell culture medium on days 8 and 10 must be 20% or less.

[0061] In practice, penalty functions and barrier functions are commonly used in machine learning to restrict the parameters of a machine learning model to a reasonable range so as to fit the actual situation in natural environmental conditions. For example, a cell culture process may have conditions that restrict the cell culture temperature and relative humidity to not become negative values, and the concentrations of ingredients in the medium to not damage the cells. By setting the barrier function and the penalty function to restrict the range of parameters, it is possible to find a combination of process parameters that is closest to the desired target result by comprehensively considering the relationship between the parameters. In addition, in practical applications, the barrier function may be applied to the constraint of the decision variables waiting for confirmation, and the penalty function may be applied to the constraint of the confirmed decision variables. They can be used interchangeably according to the needs of the user.

[0062] The calculation method of the decision variables in each of the above specific embodiments can be applied to the field of cell culture process. In addition, when applied to the field of cell culture process, the dataset used to train the trained prediction model may further include a cell dataset. The cell dataset includes a plurality of cell samples. And the sample parameters of each cell sample include source parameters and culture parameters. In practice, the cell dataset may be any data obtained from an open platform, or may be data collected by the user himself. In addition, the type of cell sample may be immune cells (e.g., dendritic cells (DC cells), cytokine-induced killer cells (CIK), tumor-infiltrating lymphocytes (TIL), natural killer cells (NK cells) and CAR-T cells), stem cells (e.g., peripheral blood stem cells, adipose stem cells, bone marrow mesenchymal stem cells), chondrocytes, fibroblasts, etc., but is not limited to these in actual applications and may be determined according to the type of cell culture that the user wants to perform. In addition, in this specific embodiment, the source parameters of each cell sample may further include attribute data of the source of each cell sample. Each cell sample in the cell dataset may include a cell source corresponding to the cell and attribute data of the source, which may be physiological data of the source or other data related to the source, such as the source's gender, age, medical history, living environment, residential area, etc. However, in practical applications, the source parameters of the cell sample may further include other parameters related to the cell source that may affect cellular processes.

[0063] The culture parameters of each cell sample further include human, equipment, material, method, and environment parameters for each cell sample. The cell culture process includes many steps, and each step is related to many culture parameters. The culture parameters include human-related parameters such as the gender and age of the cell source and the experience and stability of the cell culture operator, equipment-related parameters such as the type and grade of the cell operation platform, and the stability and accuracy of the temperature and humidity control of the cell culture device, material-related parameters such as the material of the cell culture dish, the components and ratios and mixtures of the cell culture medium, method-related parameters such as the technique of the cell culture operator and the method of the cell culture process, and environment-related parameters such as the environmental temperature, humidity, and carbon dioxide concentration of the cell culture environment.

[0064] Although the above describes an example of application to the field of cell culture processes, the method for calculating decision variables of the present invention is not limited to application to the field of cell processes, and may be applied to any other field in which results can be predicted using a predictive model obtained by machine learning, and may be used for inverse estimation of decision variables that should be input to obtain a target result.

[0065] In the method for calculating the decision variables in the above-mentioned specific embodiment, the step of calculating the gradient by the numerical analysis method may further include the following steps. That is, the change amount of the sample parameter in the sample is set, the change amount is added to the vector dimension of the sample parameter, and the sample parameter after the change amount is added is input to the objective function of the trained prediction model to obtain the change value of the function. Then, after performing the above steps for all sample parameters, the change values ​​of all functions and the sample parameters to which the change amounts in all vector dimensions have been added are linked to calculate the gradient of the sample parameter of the first sample. Note that in practice, the calculation of the gradient is not limited to this, and other methods may be selected to calculate the gradient according to the needs of the user, the type of the machine learning model, or different usage situations.

[0066] In addition to the specific embodiment provided above, the calculation method of the decision variables in the present invention may have other aspects. Now, refer to FIG. 4. FIG. 4 shows a flowchart of the steps of the calculation method of the decision variables according to another specific embodiment of the present invention. As shown in FIG. 4, the calculation method of the decision variables in this specific embodiment includes:

[0067] Step S40: A trained prediction model is provided by machine learning the dataset through a neural network method. The trained prediction model includes an input layer and an output layer. The number of input terminals of the input layer is equal to the total number of decision variables to be confirmed and the confirmed decision variables. The dataset includes a plurality of samples. Each sample includes a plurality of sample parameters.

[0068] Step S41: Compare the confirmed decision variable with the corresponding sample parameters corresponding to the confirmed decision variable among all sample parameters, and select a first sample whose corresponding sample parameters are close to or consistent with the confirmed decision variable. Then, input the confirmed decision variable and other sample parameters not belonging to the corresponding sample parameters in the first sample into a trained prediction model, and if the prediction result output from the trained prediction model is consistent with the target result, set the other sample parameters as initial weights.

[0069] Step S42: Add a dummy layer to be connected to the input layer of the trained prediction model. The dummy layer includes a plurality of artificial neurons each connected to each input terminal of the input layer. The trained prediction model and the dummy layer form a parameter prediction model, and the prediction result of the trained prediction model becomes the output of the parameter prediction model.

[0070] Step S43: Set the bias value of the activation function of each artificial neuron to 0. When the input value of the activation function is 1, the output value of the activation function is 1.

[0071] Step S44: According to the confirmed decision variables, a plurality of first weight values ​​between a first artificial neuron among the artificial neurons corresponding to the confirmed decision variables and an input terminal are respectively set.

[0072] Step S45: Set the output of the parameter prediction model corresponding to the target result, and input a training data set including at least one all-one vector into each artificial neuron of the dummy layer of the parameter prediction model to train the parameter prediction model.Then, the optimizer in the neural network method that generates the trained prediction model adjusts a second reference weight value between the input layer and other artificial neurons other than the first artificial neuron based on the target result, starting from the initial weight value, and sets the second reference weight value as a confirmation waiting decision variable.

[0073] In this specific embodiment, the confirmed decision variables are fixed decision variables, so the first weight values ​​are fixed and are not allowed to be varied or adjusted by the optimizer, and only the second weight values ​​between the artificial neurons other than the first artificial neuron and the input layer are adjustable by the optimizer.

[0074] In this specific embodiment, the trained predictive model is obtained by machine learning the dataset by Artificial Neural Networks (ANN), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Recursive Neural Networks (RecNN) and Complex Neural Networks, or any other machine learning or neural network algorithms. The choice of machine learning or neural network algorithm is based on the needs of the user.

[0075] In detail, in the method for calculating the decision variables in this specific embodiment, a dummy layer is added to be connected to the input layer of the trained prediction model in step S42, thereby forming a parameter prediction model. The dummy layer includes the same number of artificial neurons as the input terminals of the input layer in the trained prediction model. A new link is also constructed between each dummy artificial neuron and its corresponding input terminal. Next, in step S43, the bias value of the activation function of the artificial neuron in the dummy layer is set to 0. As a result, when the input value of the activation function is 1 (i.e., when the input to the artificial neuron is 1), the output value of the activation function is 1 (i.e., the output to the artificial neuron is 1). In this specific embodiment, the activation function makes it possible to make the second weight value hold its own numerical value. Next, in step S45, a training data set including a plurality of all-one vectors is input to the dummy layer of the newly formed parameter prediction model (at this time, the dummy layer forms a new input layer of the parameter prediction model) to perform training. Here, each all-one vector represents data for inputting 1 to each artificial neuron. The target result can be the ground truth output from the parameter prediction model. During the training process, the optimizer can adjust the arc weights between the artificial neurons and the input terminals of the trained prediction model. It should be noted that during the training process in step S45, the parameters and the first weight values ​​in the trained prediction model are both fixed or frozen. Therefore, the optimizer cannot adjust them, and only adjusts the second weight value to minimize the difference value between the output value (prediction result) of the parameter prediction model and the ground truth (target result).

[0076] At the end of training and convergence, the second weight values ​​adjusted by the optimizer are combined with the fixed first weight values ​​to become optimal decision variables (including confirmed and awaiting confirmation decision variables). In addition, the parameter prediction model has a many-to-one characteristic. That is, multiple different input parameters can correspond to the same output result. Therefore, each training data (all-one vector) may obtain a set of second weight values ​​as optimal input decision variables. In detail, if the trained prediction model used by the user has excellent predictive effect, it is impossible to reverse estimate the input decision variables based on the trained prediction model from the predictive effect alone. However, according to the parameter prediction model in this specific embodiment, it is possible to reverse estimate the second weight values ​​as unconfirmed input parameters or decision variables by adding a dummy layer, fixing the bias value of the artificial neuron in the dummy layer, fixing the numerical value input during training, freezing the parameters of the trained prediction model, freezing the first weight values, minimizing the difference between the output value and the target result, etc.

[0077] In practical applications, the optimizers used to calculate decision variables 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. You can choose from among Adam, Lookahead, MadGrad (Momentumized, Adaptive, and Decentralized Gradient Descent), Yogi optimizer, and AdamMax (Adaptive Moment Estimation with Maximum).

[0078] The above-mentioned Adaptive Moment Estimation (Adam) optimizer (hereinafter referred to as Adam optimizer) is used to adjust the second weight value of the artificial neuron. The Adam optimizer is generally used to train various deep learning models, including convolutional neural networks (CNN), recurrent neural networks (RNN), transformers, etc. The Adam optimizer is applied to different tasks, such as image classification, natural language processing, language translation, etc. The Adam optimizer adjusts the learning rate based on the change in the gradient of each parameter, allowing different learning rates in different directions. This contributes to better control the convergence speed of machine learning training. In practical applications, the selection of the optimizer is not limited to this, and other optimizers that can adjust parameters and weight values ​​can be used based on the needs of the user and the type of training model.

[0079] In this specific embodiment, the optimizer can start adjusting the second weight value from an initial point. In practice, if the initial point is well selected, the computational resources and time required for adjusting the second weight value that may be a decision variable waiting for confirmation are reduced. Therefore, in steps S40 to S41 of this specific embodiment, the sample used to train the trained prediction model is used as a reference. That is, the confirmed decision variable and the corresponding sample parameters corresponding to the confirmed decision variable in each sample are compared to see if they are approximate or match, and the output of the trained prediction model for the sample is confirmed to be matched to the target result. If a sample that matches both exists in the data set, not only does the predicted result of the sample match the target set by the user, but the sample parameters of the part corresponding to the confirmed decision variable (e.g., the process parameters that have been implemented in the process) are also approximate or matched, indicating that the sample is suitable for the actual situation. Therefore, this sample can be used as an initial point when calculating the decision variables waiting for confirmation. For example, if the first 7 days of cell culture have already been performed out of a 21-day cell culture process, the confirmed decision variables for the 7 days are already available. Therefore, the sample as the initial point not only needs to satisfy the user's goal (e.g., to culture 95% viable cells) with the predicted results output from the trained prediction model, but also needs to have the sample parameters in the first 7 days of the sample be the same or close to the process parameters already run, otherwise the subsequent process parameters obtained by the optimizer through adjustment based on the initial point will be out of touch with the actual situation.

[0080] It should be noted that the steps actually performed in step S41 in this specific embodiment are the same as the above-mentioned method of comparing whether the first vector and the second vector are similar or identical. For example, the sample having the smallest included angle between the first vector and the second vector is obtained, or the sample having the smallest distance between the coordinates of the end point of the first vector and the coordinates of the end point of the second vector is calculated and obtained by a distance function. Note that the detailed steps are described in detail in the above specific embodiment, so they will not be described here again.

[0081] In summary, the present invention provides a method for calculating decision variables, in which the method selects a cell reference sample in a trained prediction model by comparing, and increases the step size along the opposite direction of the gradient for the sample parameters of the cell reference sample to inversely estimate the disconfirmed decision variables. In addition, the barrier function and the penalty function are set to restrict the range of the parameters, and the relationship between the parameters is comprehensively considered to find the combination of process parameters that is closest to the desired target result. In addition, the sample parameters in the data set, the input values ​​and bias values ​​to the artificial neurons in the artificial neuron layer are set and fixed, and the second weight value is adjusted by the optimizer according to the target result and the first weight value, respectively, to inversely estimate the disconfirmed decision variables. In this way, the method of the process is effectively optimized and the target result of the process is achieved. In addition, the method of calculating decision variables in the present invention is not limited to application in the field of cell processes, and may also be applied to any other trained machine learning model to inversely estimate the process parameters and disconfirmed decision variables.

[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] S1~S6, S30~S36, S300~S306, S50~S60, S40~S50 Step

Claims

1. 1. A method for computing a plurality of awaiting confirmation decision variables, comprising: providing a trained prediction model, the trained prediction model being obtained by machine learning a dataset through a machine learning method, the trained prediction model including an input layer and an output layer, the dataset including a plurality of samples, and each of the samples including a plurality of sample parameters, the trained prediction model being used to input a plurality of input variables through the input layer and generate a prediction result corresponding to the input variables through the output layer; setting a target outcome corresponding to the predicted outcome of the trained predictive model; Among the samples, at least one sample in which the prediction result of the trained prediction model matches the target result is selected as at least one anchor sample, and a multidimensional subspace is formed by the at least one anchor sample; and an initial sample is determined in the multidimensional subspace; A step of obtaining a gradient by minimizing an objective function of the trained predictive model using an optimizer used to generate the trained predictive model; increasing the step size of the sample parameters of the initial samples along the opposite direction of the gradient, and then inputting the sample parameters to the trained prediction model to check whether the generated prediction results match the target results; if the predicted result matches the target result, setting the sample parameters after increasing the step size as the decision variables awaiting confirmation.

2. providing a plurality of validated decision variables and calculating a first vector of the validated decision variables; obtaining corresponding sample parameters from the sample parameters of the samples based on the confirmed decision variables, respectively, and calculating a second vector of the corresponding sample parameters; obtaining a reference sample from the samples by comparing the first vector with the second vector of each of the samples and comparing the predicted result generated by each of the samples in the trained prediction model with the target result, wherein the second vector of the reference sample is close to or coincides with the first vector and the predicted result of the reference sample is close to or coincides with the target result; The method of claim 1 , further comprising the step of: selecting said reference sample as said initial sample.

3. A method for calculating a decision variable, comprising the steps of: providing a trained prediction model, the trained prediction model being obtained by machine learning a dataset by a machine learning method, the dataset including a plurality of samples, each of the samples including a plurality of sample parameters, the trained prediction model being used to input a plurality of decision variables to generate a prediction result corresponding to the decision variables, the decision variables including a plurality of confirmed decision variables and a plurality of decision variables to be confirmed; setting a target outcome corresponding to the predicted outcome of the trained predictive model; selecting a candidate sample whose corresponding sample parameters are similar to or match the confirmed decision variable by comparing the confirmed decision variable with a plurality of corresponding sample parameters among the sample parameters, the corresponding sample parameters being similar to or match the confirmed decision variable; inputting other sample parameters not belonging to the corresponding sample parameters in the candidate sample together with the confirmed decision variables into the trained prediction model, and confirming whether the prediction result output from the trained prediction model is consistent with the target result; For all sample parameters in the dataset, the steps of selecting the candidate samples and checking whether the prediction result is consistent with the target result are performed, and among the candidate samples, at least one sample whose prediction result of the trained prediction model is consistent with the target result is selected as at least one anchor sample, and a multidimensional subspace is formed by the at least one anchor sample; determining an initial sample within the multidimensional subspace and obtaining an orientation of the initial sample within the multidimensional subspace; after increasing the step size along the direction, inputting the confirmed decision variables and other sample parameters that do not correspond to the confirmed decision variables in the initial sample into the trained prediction model to verify whether the generated prediction result is consistent with the target result; if the predicted outcome matches the target outcome, then setting the sample parameters after increasing the step size that do not correspond to the confirmed decision variables as the decision variables awaiting confirmation.

4. The step of comparing the confirmed decision variables with the corresponding sample parameters among the sample parameters, which correspond to the confirmed decision variables, to select the candidate samples whose corresponding sample parameters are close to or match the confirmed decision variables further includes: calculating a first vector of the identified decision variables; obtaining corresponding sample parameters from the sample parameters of the samples based on the confirmed decision variables, respectively; calculating a second vector of corresponding sample parameters; 4. The method of claim 3, further comprising the step of: comparing the first vector with the second vector of each of the samples; and selecting the first one of the samples as the candidate sample if the second vector of the first one of the samples is close to or matches the first vector.

5. Furthermore, 5. The method of claim 4, further comprising the step of: calculating an included angle between the first vector and the second vector of each of the samples; and determining as the first sample the sample corresponding to the second vector that has the smallest included angle with the first vector.

6. Furthermore, 5. The method for calculating a decision variable according to claim 4, further comprising the step of calculating a distance between a coordinate of an end point of the first vector and a coordinate of an end point of the second vector of each of the samples by a distance function, and determining as the first sample the sample corresponding to the second vector having the smallest distance between the sample and the first vector.

7. The step of determining the initial sample within the multi-dimensional subspace further comprises:

4. The method of claim 3, further comprising the step of: determining as the initial sample a midpoint of a distance between at least two of the anchor samples in the multi-dimensional subspace.

8. The step of determining the initial sample within the multi-dimensional subspace further comprises: The method of claim 3 , further comprising the step of: determining the initial sample as a linear combination of at least two of the anchor samples in the multi-dimensional subspace.

9. The step of obtaining the orientation of the initial sample within the multi-dimensional subspace further comprises:

4. The method of claim 3, further comprising the step of: calculating, by numerical analysis, a first directional derivative of said initial sample along each of said anchor samples as said direction of said initial sample.

10. The step of obtaining the orientation of the initial sample within the multi-dimensional subspace further comprises:

4. The method of claim 3, further comprising the step of setting the initial sample at an origin and calculating, by a numerical analysis method, a second directional derivative of the origin towards each of the anchor samples as the direction of the initial sample.

11. The trained prediction model is expressed by an objective function, and obtaining the orientation of the initial sample in the multi-dimensional subspace further comprises:

4. The method of claim 3, further comprising: calculating a change in a function value of the objective function based on a change in distance of the initial sample towards each of the anchor samples using a numerical analysis method to obtain a secant derivative of the initial sample towards each of the anchor samples.

12. The step of obtaining the orientation of the initial sample within the multi-dimensional subspace further comprises: Incrementing and adjusting the step size of the initial sample towards the directional derivative of each of the anchor samples in turn; Increasing and adjusting the step size of the initial sample towards the directional derivative of each of the anchor samples, and then checking in turn whether the prediction result outputted from the trained prediction model is closer to the target result; increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, if it is determined that the prediction result input to the trained prediction model and output is closer to the target result, increasing the step size of the initial sample in the directional derivative with the directional derivative as the direction; 4. The method of claim 3, further comprising: if it is determined that the prediction result input to the trained prediction model and output does not get any closer to the target result after increasing the step size of the initial sample toward the directional derivative of one of the anchor samples, then increasing the step size of the initial sample toward the directional derivative of another of the anchor samples and then checking again whether the prediction result input to the trained prediction model and output gets any closer to the target result.

13. 1. A method for computing a plurality of awaiting confirmation decision variables, comprising: providing a trained prediction model, the trained prediction model being obtained by machine learning a dataset using a neural network method, the trained prediction model including an input layer and an output layer, the number of input terminals of the input layer being equal to the sum of the number of the decision variables to be confirmed and the number of confirmed decision variables, the dataset including a number of samples, each of the samples including a number of sample parameters; a step of comparing the confirmed decision variable with a plurality of corresponding sample parameters among the sample parameters, which correspond to the confirmed decision variable, to select a first sample whose corresponding sample parameters are approximate or identical to the confirmed decision variable, inputting the confirmed decision variable together with other sample parameters not belonging to the corresponding sample parameters in the first sample into the trained prediction model, and setting the other sample parameters as initial weight values ​​when the prediction result output from the trained prediction model matches the target result; adding a dummy layer connected to the input layer of the trained prediction model, the dummy layer including a plurality of artificial neurons respectively connected to each input terminal of the input layer, the trained prediction model and the dummy layer forming a parameter prediction model, and the prediction result of the trained prediction model being the output of the parameter prediction model; setting a bias value of an activation function of each of the artificial neurons to 0, such that when an input value of the activation function is 1, an output value of the activation function is 1; determining, based on the confirmed decision variables, a plurality of first weight values ​​between a plurality of first artificial neurons among the artificial neurons corresponding to the confirmed decision variables and the input terminals; and training the parameter prediction model by inputting a training data set including at least one all-ones vector to the artificial neurons of the dummy layer of the parameter prediction model to generate the trained prediction model. The neural network method includes a step of adjusting a plurality of second reference weight values ​​between each of the artificial neurons and the input layer based on the target result, starting from the initial weight values, and setting the second reference weight values ​​as the decision variables to be confirmed.

14. The method of claim 13 , wherein the trained predictive model includes a plurality of model parameters, all of which are fixed.

15. 15. The method of claim 14, wherein the first weight value is fixed.

16. 15. The method of claim 14, wherein the neural network method comprises one of Artificial Neural Networks (ANN), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Recursive Neural Networks (RecNNs) and Complex Neural Networks (Complex Neural Networks).

17. The optimizer may further include Adaptive Moment Estimation (Adam) optimization, Stochastic Gradient Descent (SGD), Momentum, Nesterov Accelerated Gradient (NAG), Adaptive Gradient Algorithm (AdaGrad), Nadam (Nesterov-accelerated Adaptive Moment Estimation), RMSprop (Root Mean Square), and other optimization methods. Propagation), Adadelta (Adaptive Delta), AdamW (Adam with Weight Decay), AMSGrad (Adaptive Moment Estimation with Long-term Memory), AdaBelief (Adaptive Belief), LARS (Layer-wise Adaptive Rate) Scaling), Self-adaptive Hessian (AdaHessian), RAdam (Rectified Adam), Lookahead (Lookahead), MadGrad (Momentumized, Adaptive, and 15. The method of claim 14, wherein the optimization method is selected from the group consisting of Decentralized Gradient Descent, Yogi optimizer (Yogi) and AdamMax (Adaptive Moment Estimation with Maximum).

Images