How to calculate the 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 particularly to an algorithm based on optimization technology, an optimizer incorporated in an artificial intelligence model, and a method for inversely estimating decision variables by applying a trained prediction model.

Background Art

[0002] With the progress of medical engineering technology, currently, the application of regenerative medicine in clinical treatment of diseases is becoming increasingly diversified. Regenerative medicine is mainly a medical technology that repairs damaged tissues and organs by utilizing the regenerative ability of cells, and its application range is very wide. Also, by combining medical technologies such as tissue engineering and molecular biology, it is expected that diseases that were once considered 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, immunocyte therapy, and stem cell therapy. Also, the research and application of cell therapy in regenerative medicine have been attracting increasing attention from all sectors. Cell therapy is a method in which human cells are cultured or processed outside the body and then transplanted into individual bodies for use.

[0003] In recent years, as many governments have gradually liberalized the application of cell therapy, numerous scholars both domestically and internationally have entered the field of cell therapy research. 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. Furthermore, the quality of cell therapy products directly impacts the safety and effectiveness of treatment. Therefore, in the cell culture process, it is necessary to strictly control the growth state of cells and to immediately monitor the culture parameters and environmental parameters of cell growth to avoid contamination or deterioration of cell 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 for cell preparations applied to different cases are not exactly the same. For this reason, it is necessary to design and adjust each process parameter within the process for each cell preparation to achieve the desired results, and it is not possible to produce each cell preparation with fixed process parameters. Furthermore, due to the complexity and correlation of the process, conventional techniques typically only allow for the optimization of a single parameter, neglecting the overall interrelationships between parameters throughout the entire process.

[0004] Conventional techniques involve training a predictive model on a large dataset of samples using machine learning. This predictive model can generate predictive results for cell culture by inputting various process parameters, allowing users to simulate the effects of their designed cell culture process in advance. However, the predictive model trained by machine learning cannot inversely estimate the process parameters of the cell culture process from the user's initial results. In other words, when designing a cell culture process, users need to try a large number of different process parameters, resulting in significant effort and resources being spent on designing and improving the cell culture process. Furthermore, the problem of being unable to inversely estimate the decision variables also exists for predictive models obtained through machine learning in fields other than cell processes.

[0005] Therefore, in order to solve the problems of conventional technology, it is necessary to develop a method that allows for accurate inverse estimation of decision variables that match the target outcome using a pre-trained predictive model. [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] In view of the above, the present invention aims to provide a method for calculating decision variables in order to solve the conventional problems described above. [Means for solving the problem]

[0007] The present invention provides a method for calculating multiple decision variables awaiting confirmation, which includes a trained predictive model obtained by machine learning a dataset using a machine learning method, the trained predictive model comprising an input layer and an output layer, the dataset comprising multiple samples, and each sample comprising multiple sample parameters, the trained predictive model being used to input multiple input variables through the input layer and to generate prediction results corresponding to the input variables by the output layer, setting a target result corresponding to the prediction result of the trained predictive model, and selecting at least one sample from the samples in which the prediction result of the trained predictive model matches the target result. The method includes the steps of: setting at least one anchor sample and forming a multidimensional subspace with the at least one anchor sample, and determining an initial sample within the multidimensional subspace; obtaining a gradient by minimizing the objective function of the trained predictive model using the optimizer used to generate the trained predictive model; increasing the step size of the sample parameters of the initial sample along the opposite direction of the gradient, and then inputting this into the trained predictive 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 confirmation pending decision variable.

[0008] The method for calculating the decision variables 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 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 sample by comparing the first vector with the second vector of each of the sample and comparing the prediction result that each of the sample generates in the trained prediction model with the target result, wherein the second vector of the reference sample approximates or matches the first vector, and the prediction result of the reference sample approximates or matches the target result; and setting the reference sample as the initial sample.

[0009] The present invention further provides a trained predictive model, the trained predictive model is obtained by machine learning a dataset using a machine learning method, the dataset includes a plurality of samples, and each of the samples includes a plurality of sample parameters, the trained predictive model is used to take a plurality of decision variables as input and generate a prediction result corresponding to the decision variables, the decision variables include a step of including a plurality of confirmed decision variables and a plurality of pending decision variables, setting a target result corresponding to the prediction result of the trained predictive model, selecting a candidate sample in which the corresponding sample parameter approximates or matches the confirmed decision variable by comparing the confirmed decision variable with a plurality of corresponding sample parameters among the sample parameters corresponding to the confirmed decision variable, and inputting the other sample parameters in the candidate sample that do not belong to the corresponding sample parameter together with the confirmed decision variable into the trained predictive model, the prediction result output from the trained predictive model is The present invention provides a method for calculating a decision variable, which includes the steps of: checking whether the result matches the target result; performing the above steps for all sample parameters in the dataset, selecting at least one sample from the samples for which the prediction result of the trained prediction model matches the target result, and forming a multidimensional subspace with the at least one anchor sample; determining an initial sample in the multidimensional subspace and obtaining the direction of the initial sample in the multidimensional subspace; increasing the 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, and checking 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 that do not correspond to the confirmed decision variable as the decision variable awaiting confirmation.

[0010] The step of selecting a candidate sample in which the corresponding sample parameter approximates or matches the confirmed decision variable by comparing the confirmed decision variable with the corresponding sample parameter among the sample parameters that corresponds to the confirmed decision variable further includes the steps of: calculating a first vector of the confirmed decision variable; obtaining the corresponding sample parameter from the sample parameter of the sample based on the confirmed decision variable; calculating a second vector of the corresponding sample parameter; and comparing the first vector with the second vector of each sample, and selecting the first sample as the candidate sample if the second vector of the first sample among the samples approximates or matches the first vector.

[0011] The method for calculating the decision variable further includes the step of calculating the binding angle between the first vector and the second vector of each sample, and defining the sample corresponding to the second vector having the smallest binding angle with respect to the first vector as the first sample.

[0012] The method for calculating the decision variable further includes the step of calculating the distance between the coordinates of the tip of the first vector and the coordinates of the tip of the second vector of each sample using a distance function, and designating the sample corresponding to the second vector having the smallest distance from the first vector as the first sample.

[0013] The step of determining the initial sample in the multidimensional subspace further includes the step of setting the midpoint of the distance between at least two of the anchor samples in the multidimensional subspace as the initial sample.

[0014] The step of determining the initial sample in the multidimensional subspace further includes the step of using a linear combination of at least two of the anchor samples in the multidimensional subspace as the initial sample.

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

[0016] The step of obtaining the direction of the initial sample in the multidimensional subspace further includes setting the initial sample as the origin and calculating the second direction derivative of the origin toward each anchor sample as the direction of the initial sample by the numerical analysis method.

[0017] The trained prediction model is represented by an objective function. Furthermore, the step of obtaining the direction of the initial sample in the multidimensional subspace further includes the step of obtaining the secant derivative of the initial sample toward each anchor sample by calculating the change in the function value of the objective function based on the change in the distance of the initial sample toward each anchor sample using 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 anchor sample, and after increasing and adjusting the step size of the initial sample toward the directional derivative of each anchor sample, sequentially checking whether the prediction result output by inputting it into the trained prediction model comes closer to the target result, and after increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, the prediction result output by inputting it into the trained prediction model comes closer to the target result. If it is confirmed that the result will be closer to the target result, the step of increasing the step size of the initial sample in the directional derivative is included; if it is confirmed that the prediction result output when input to the trained predictive model does not come any closer to the target result after increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, the step of increasing the step size of the initial sample toward the directional derivative of another anchor sample and then confirming again whether the prediction result output when input to the trained predictive model comes any closer to the target result is included.

[0019] The present invention further provides a method for calculating decision variables for calculating multiple pending decision variables. The method for calculating decision variables provides a trained predictive model, which is obtained by machine learning a dataset using a neural network method, which includes an input layer and an output layer, the number of input terminals in the input layer being equal to the sum of the number of pending decision variables and multiple confirmed decision variables, the dataset includes multiple samples, each of which includes multiple sample parameters, and the method includes the step of comparing the confirmed decision variable with multiple corresponding sample parameters from the sample parameters that correspond to the confirmed decision variable, selecting a first sample in which the corresponding sample parameters approximate or match the confirmed decision variable, inputting the confirmed decision variable and other sample parameters in the first sample that do not belong to the corresponding sample parameters into the trained predictive model, and if the prediction result output from the trained predictive model matches the target result, setting the other sample parameters as initial weight values, and the method includes the step of inputting the confirmed decision variable with multiple corresponding sample parameters from the sample parameters that correspond to the confirmed decision variable, which includes the step of inputting the confirmed decision variable with other sample parameters that do not belong to the corresponding sample parameters in the first sample into the trained predictive model, and setting the other sample parameters as initial weight values. The process involves adding a dummy layer connected to the input layer, the dummy layer including a plurality of artificial neurons connected to each input terminal of the input layer, forming a parameter prediction model with the trained prediction model and the dummy layer, and the prediction result of the trained prediction model becoming the output of the parameter prediction model; setting the bias value of the activation function of each artificial neuron to 0, so that when the input value of the activation function is 1, the output value of the activation function becomes 1; setting a plurality of first weight values ​​between a plurality of first artificial neurons and the input terminals of the artificial neurons corresponding to the confirmed decision variable based on the confirmed decision variable; setting the output of the parameter prediction model corresponding to the target result; training the parameter prediction model by inputting a training dataset including at least one all-one vector to the artificial neurons of the dummy layer of the parameter prediction model, and using an optimizer in the neural network method that generated the trained prediction model, based on the target result,The process includes the step of adjusting the reference weight values ​​between each artificial neuron and the input layer, starting from the initial weight values, and setting a plurality of second reference weight values ​​corresponding to the initial weight values ​​among the reference weight values ​​as the confirmation-awaited decision variables.

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

[0021] The aforementioned first weight value is fixed.

[0022] The optimizer includes 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 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), and RAdam (Rectified The following can be selected: Adam), Lookahead, MadGrad (Momentumized, Adaptive, and Decentralized Gradient Descent), Yogi optimizer (Yogi), and AdamMax (Adaptive Moment Estimation with Maximum). [Effects of the Invention]

[0023] In summary, the method for calculating decision variables in the present invention allows for the calculation of decision variables that match the target result or ground truth by performing inverse estimation on a predictive model acquired through machine learning training. More specifically, the method for calculating decision variables in the present invention allows for the conversion of the constrained objective function of the trained predictive model into an unconstrained objective function, and then optimizing for the unconstrained objective function to find a solution. Furthermore, in the process of finding a solution, anchor samples can be set to form a multidimensional feasible solution space, and the optimal solution can be searched within this space. When finding a solution, the initial solution, direction, and step size can be considered. The initial solution can be obtained from the sample set used during the training of the trained predictive model; for example, a sample with sample parameters that are close to or match those of a confirmed decision variable can be used as the initial solution. In addition, the optimizer of the platform on which the trained predictive model was trained can obtain the gradient of either the initial solution or a point in the feasible solution space, and the inverse direction of this gradient can be used as the direction. Finally, starting from the initial solution, the direction and step size are continuously adjusted to calculate the optimal solution that matches the target result as a pending decision variable. In addition, the method for calculating the decision variable in the present invention may involve calculating the unconstrained objective function using a numerical method when determining the direction. Alternatively, the method for calculating the decision variable in the present invention may involve directly training a pre-trained predictive model with an added dummy layer using an optimizer. Then, by training the pre-trained predictive model using samples from the sample set used during training as initial weight values, arc weights are obtained between the dummy layer and the pre-trained predictive model, and these arc weights are used as the decision variable. [Brief explanation of the drawing]

[0024] [Figure 1] Figure 1 shows a flowchart of the steps for calculating the decision variable based on a specific embodiment of the present invention. [Figure 2]FIG. 2 shows a flowchart of further steps of step S3 of the method for calculating decision variables based on FIG. 1. [Figure 3] FIG. 3 shows a flowchart of steps of the method for calculating decision variables based on another specific embodiment of the present invention. [Figure 4] FIG. 4 shows a flowchart of steps of the method for calculating decision variables based on another specific embodiment of the present invention. MODE FOR CARRYING OUT THE INVENTION

[0025] Subsequently, in order to make the advantages, spirit and features of the present invention more easily and clearly understandable, specific examples will be used and described in detail with reference to the drawings. It should be noted that these specific examples are only representative specific examples of the present invention, and the specific methods, devices, conditions, materials, etc. exemplified do not limit the present invention or the corresponding specific examples. Also, each component in the drawings is used only to represent their relative positions and is not described based on actual ratios. Further, the step numbers of the present invention are only for distinguishing different steps and do not represent the order of the steps. This is explained in advance above.

[0026] Referring to FIG. 1. FIG. 1 shows a flowchart of steps of the method for calculating decision variables based on a specific embodiment of the present invention. The method for calculating decision variables in this specific embodiment is used to calculate a plurality of decision variables pending confirmation. Among the decision variables, the confirmed decision variables are the decision variables that have been confirmed by the user during operation or have actually been executed. Also, the decision variables pending confirmation are the decision variables that the user has not yet confirmed or executed. As shown in FIG. 1, in this specific embodiment, the method for calculating decision variables includes the following.

[0027] Step S1: Provide a trained predictive model obtained by machine learning the dataset using a machine learning method. The dataset contains multiple samples, and each sample contains multiple sample parameters. The input layer of the trained predictive model is used to take multiple input variables as input and output and generate prediction results corresponding to the input variables.

[0028] Step S2: Set target results corresponding to the prediction results of the trained predictive model.

[0029] Step S3: At least one sample in the dataset whose prediction result from the trained predictive model matches the target result is designated as at least one anchor sample. Furthermore, a multidimensional subspace is formed using this at least one anchor sample, and the initial sample is determined within this multidimensional subspace.

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

[0031] Step S5: The sample parameters of the initial sample are increased in step size along the opposite direction of the gradient, and then input into the trained predictive model to check whether the generated prediction results match the target results.

[0032] Step S6: If the prediction result matches the target result, the sample parameter after increasing the step size is used as the confirmation-awaited decision variable.

[0033] In this specific embodiment, the trained predictive model in step S1 may be any machine learning model obtained from an open platform that has completed training, or it may be a predictive model trained by the user themselves. Furthermore, the dataset in this specific embodiment may be a set formed from any data that can be used for training, testing, and validating machine learning. In actual applications, the method for calculating the decision variable in this specific embodiment can also be applied to any other trained machine learning model and used for inverse estimation of awaiting decision variables. Furthermore, the trained predictive model in this specific embodiment can be obtained by machine learning the dataset using an artificial neural network (ANN), convolutional neural network (CNN), recurrent neural network (RNN), or any other machine learning algorithm or neural network algorithm. The selection of a machine learning or neural network algorithm will be made according to the user's needs.

[0034] In step S2, the user can first set a target result, also known as the Ground Truth. This is the target to be achieved by the decision variable calculated using the decision variable calculation method in this invention. The decision variable calculation method performs inverse estimation based on a trained predictive model, which can be represented by an objective function. Therefore, the process of obtaining the decision variable by performing inverse estimation on the trained predictive model can be considered as finding a solution to an optimization problem for the objective function. When finding a solution to an optimization problem, it is more efficient to obtain a feasible solution space in advance and find a solution within that space. Therefore, in step S3, the above-mentioned feasible solution space can be formed using the samples from when the trained predictive model was trained. More specifically, for the samples in the dataset, the prediction results output after inputting into the trained predictive model are known. Among the samples in the dataset, those whose prediction results match the target result have considerable reference value when finding a solution for the decision variable. Therefore, these samples can be used as anchor samples to form a multidimensional subspace, and this multidimensional subspace will be the feasible solution space described above. After determining the multidimensional subspace, an initial point can be determined within this multidimensional subspace, and the solution to the optimization problem can be started. This initial point will be the initial sample in step S3.

[0035] As described above, after determining the initial samples in the feasible solution space, we can begin finding the solution to the optimization problem from the initial samples. More specifically, after advancing one step in one direction from the initial samples, the sample parameters with increased step sizes are input into the trained predictive model to check whether the output prediction result matches the target result, or whether the gap with the target result is reduced. As described in steps S5 and S6, by repeating the above step size advancement, we can finally obtain input parameters in which the output prediction result perfectly matches the target result or the gap with the target result is minimized, and these can be used as the decision variable awaiting confirmation. However, when determining the direction of the initial samples or each point in the process of finding the solution to the optimization problem, the gradient can generally be obtained by taking partial derivatives with respect to the objective function, and the opposite direction of the gradient is taken as the direction. However, the objective function of a trained predictive model is usually not symbolic, is often not differentiable, 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, analyzing the derivative (or gradient) of the objective function using numerical analysis methods consumes a lot of computational resources and can lead to problems such as overhead and inaccuracies in numerical calculations. Therefore, in step S4 of the method in this specific embodiment, the gradient can be obtained by minimizing the objective function of the trained prediction model using an optimizer that is originally intended to generate a trained model. For example, it is possible to calculate the gradient in the process of completing the minimization of the objective function using a stochastic gradient descent (SGD) optimizer. Compared to numerical analysis methods, calculating the gradient using an optimizer allows for accurate calculation of the gradient and significantly reduces computational resources.

[0036] In the specific embodiment described above, the feasible solution space can be formed from samples in the dataset. That is, samples whose prediction results match or approximate the target result are selected as anchor samples. However, in practice, not all decision variables are necessarily awaiting confirmation, and there may be situations where confirmed and awaiting decision variables exist simultaneously. For example, suppose a user is attempting to perform a 21-day cell culture process, the cell product to be produced must have a 95% cell viability rate, 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 perform the process or adjust the process parameters (i.e., awaiting confirmation decision variables) for days 8-21 to obtain the desired target result on the predetermined 21st day. If only samples whose prediction results match or approximate the target result 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 final decision variables obtained for the first 7 days will usually differ significantly from the confirmed decision variables.

[0037] Now, refer to Figure 2. Figure 2 shows a flowchart of the further steps of step S3 of the method for calculating the decision variable based on Figure 1. As shown in Figure 2, this specific embodiment differs from the specific embodiment described above in the following respects. That is, the method of this specific embodiment further includes the following.

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

[0039] Step S32: Based on the confirmed decision variable, multiple corresponding sample parameters are obtained from the sample parameters of the sample, and a second vector of the corresponding sample parameters is calculated.

[0040] Step S34: A reference sample is obtained from the samples by comparing the first vector with the second vector of each sample, and by comparing the prediction result generated by each sample in the trained prediction model with the target result. The second vector of the reference sample approximates or matches the first vector, and the prediction result approximates or matches the target result.

[0041] Step S36: Use the reference sample as the initial sample.

[0042] In the step of determining the reference sample described above, the determination is made based on whether the second vector approximates or matches the first vector. In practice, when determining whether they approximate or match, one may check whether the angle between the two vectors is close to 0 degrees or is 0 degrees. More specifically, in a particular embodiment, the step of determining the reference sample may include calculating the angle between the first vector and the second vector of each sample, and obtaining the sample corresponding to the second vector having the smallest angle between it and the first vector (close to 0 degrees or is 0 degrees) as the reference sample. Alternatively, in another particular embodiment, the step of determining the reference sample may include calculating the distance between the coordinates of the tip of the first vector and the coordinates of the tip of the second vector using a distance function, and obtaining the sample corresponding to the second vector having the smallest distance between it and the first vector as the reference sample. In practical applications, commonly used distance functions in the field of machine learning include Euclidean distance, Manhattan distance, cosine similarity, Jaccard similarity, and Mahalanobis distance. In practical use, the method for calculating the minimum distance and the range of application of the minimum distance may be set according to the user's needs.

[0043] If the vector formed by the reference sample parameters of the reference sample corresponding to a confirmed decision variable (e.g., cell process parameters for the first 7 days that have been run) approximates or matches the vector of the confirmed decision variable, it means that the reference sample parameters of the reference sample approximate or match the confirmed decision variable. Therefore, the optimal solution obtained by finding the solution to the above optimization problem using this reference sample as the initial sample is applicable to situations where some decision variables are confirmed (e.g., cell process parameters for the first 7 days) and some decision variables are awaiting confirmation (e.g., cell process parameters for days 8-21).

[0044] Now, refer to Figure 3. Figure 3 shows a flowchart of the steps for calculating the decision variable based on another specific embodiment of the present invention. As shown in Figure 3, this specific embodiment differs from the specific embodiment described above in the following respects. That is, the method for calculating the decision variable in this specific embodiment further includes the following.

[0045] Step S300: By comparing the confirmed decision variables among the decision variables with the multiple corresponding sample parameters among the sample parameters that correspond to the confirmed decision variables, a candidate sample is selected in which the corresponding sample parameter approximates or matches the confirmed decision variable.

[0046] Step S302: The other sample parameters in the candidate samples that do not belong to the corresponding sample parameters are input into the trained predictive model along with the confirmed decision variables, and it is checked whether the prediction result output from the trained predictive model matches the target result.

[0047] Step S304: Perform the above steps for all sample parameters in the dataset, and from all the samples, select at least one sample whose prediction result matches the target result from the trained predictive model as the anchor sample, and form a multidimensional subspace with the anchor sample.

[0048] Step S306: Determine the initial sample in the multidimensional subspace and obtain the orientation of the initial sample.

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

[0050] Step S60: If the prediction result matches the target result, the sample parameter after increasing the step size that does not correspond to the confirmed decision variable is set as the decision variable 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 specific embodiment described above are described in detail in the specific embodiment described above, and therefore will not be described in detail again here.

[0052] This specific embodiment differs from the above-described specific embodiment in the following respects. Specifically, in this specific embodiment, first, a sample having corresponding sample parameters is selected as a candidate sample using a confirmed decision variable. Next, the confirmed decision variable and other sample parameters in the candidate sample that do not correspond to the confirmed decision variable (i.e., the part awaiting confirmation) are combined and input into the trained prediction model. If the prediction result output from the trained prediction model matches the target result, it means that the sample parameter corresponding to the part awaiting confirmation of this candidate sample has reference value, and therefore this candidate sample can be used as an anchor sample to form a feasible solution space (multidimensional subspace). Similarly in this case, as shown in step S306, it is necessary to first start finding the solution to the optimization problem from the 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 using a numerical analysis method, and the opposite direction of the gradient is taken as the direction. Furthermore, since the confirmed decision variables are fixed, in step S50, the step size of the other sample parameters in the initial sample is increased along the aforementioned direction, and then input into the trained predictive model in combination with the confirmed decision variables to check whether the prediction result matches the target result. Alternatively, the above direction determination and step size increase operations are continued until the prediction result approaches the target result. Finally, in step S60, if the prediction result matches the target result after continuing to increase the step size, or in practical applications, if the difference between the prediction result and the target result is minimized, the sample parameters after increasing the step size that do not correspond to the confirmed decision variables are set as the decision variables awaiting confirmation.

[0053] In this specific embodiment, the step of determining candidate samples in step S300 is determined by whether the second vector approximates or matches the first vector. As described above, the first vector is a vector calculated for the confirmed decision variable, and the second vector is a vector calculated for the corresponding sample parameter corresponding to the confirmed decision variable in the sample. In practice, when determining whether these approximate or match, it may be necessary to check whether the angle between the two vectors is close to 0 degrees or is 0 degrees. More specifically, in this specific embodiment, the step of determining candidate samples may include calculating the angle between the first vector and the second vector of each sample, and obtaining the sample corresponding to the second vector having the smallest angle between it and the first vector (close to 0 degrees or is 0 degrees) as a candidate sample. Alternatively, in another specific embodiment, the step of determining candidate samples may include calculating the distance between the coordinates of the tip of the first vector and the coordinates of the tip of the second vector using a distance function, and obtaining the sample corresponding to the second vector having the smallest distance between it and the first vector as a candidate sample. In practical applications, commonly used distance functions in the field of machine learning include Euclidean distance, Manhattan distance, cosine similarity, Jaccard similarity, and Mahalanobis distance. In practical use, the method for calculating the minimum distance and the range of application of the minimum distance may be set according to the user's needs.

[0054] As described above, in step S306 of this specific embodiment, an initial sample is determined within a multidimensional subspace (feasible solution space). In practice, the method for determining the initial sample can be either the midpoint of the distance between at least two anchor samples in the multidimensional subspace, or the linear combination between at least two anchor samples in the multidimensional subspace, or a combination thereof. Each anchor sample is selected from candidate samples, and its corresponding sample parameters all match or approximate the confirmed decision variables. Therefore, the corresponding sample parameters in the initial sample determined in the above step also match or approximate the confirmed decision variables.

[0055] Furthermore, as mentioned above, when determining the direction for an initial sample or a subsequent sample after increasing the step size at least once within a multidimensional subspace, the gradient of the objective function is calculated using numerical analysis, and the direction opposite to the gradient is taken as the direction. In practice, the method for determining the direction described above can be any one or a combination of the following steps: calculating the direction derivative of the initial sample along each anchor sample as the direction of the initial sample (or subsequent sample) using numerical analysis; setting the initial sample (subsequent sample) as the origin and calculating the direction derivative of the origin toward each anchor sample as the direction of the initial sample (subsequent sample) using numerical analysis; or obtaining the secant direction derivative of the initial sample (subsequent sample) toward each anchor sample by 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 using numerical analysis.

[0056] If the initial samples proceed in the opposite direction to the feasible solution, the function value may worsen or deviate from expectations. Therefore, the method for determining the direction can be used to sequentially search for directions in which the function value may be favorable. According to another specific embodiment, the method for determining the direction may include the steps of: sequentially increasing and adjusting the step size of the initial samples toward the directional derivative of each anchor sample; sequentially checking whether the prediction result output by inputting into a trained predictive model after increasing and adjusting the step size of the initial samples toward the directional derivative of each anchor sample approaches the target result; if it is confirmed that the prediction result output by inputting into a trained predictive model after increasing the step size of the initial samples toward the directional derivative of any of the anchor samples approaches the target result, the step of increasing the step size of the initial samples with that directional derivative as the direction; and if it is confirmed that the prediction result output by inputting into a trained predictive model after increasing the step size of the initial samples toward the directional derivative of any of the anchor samples does not approach the target result, the step of increasing the step size of the initial samples toward the directional derivative of another anchor sample and then checking again whether the prediction result output by inputting into the trained predictive model approaches the target result. Next, after advancing one step increment from the initial sample to reach the next subsequent or intermediate sample, the above steps are performed again to explore a new direction, and then the process is advanced with the next step increment. In summary, in the method for determining the direction in this specific embodiment, first, the step increment is repeatedly adjusted toward the direction of the first anchor sample from the initial sample, and it is checked whether the prediction result comes closer to the target result after increasing the step increment. Then, different step increments are repeatedly tried, and after finding an appropriate step increment, the process is advanced in that direction with that step increment. However, if an appropriate step increment that brings the prediction result closer to the target result cannot be found even after adjusting the step increment a certain number of times, the above adjustment is repeated for the initial sample toward the direction of the next anchor sample.In practice, if an appropriate step size that yields better function values ​​(smaller difference between predicted and target results) cannot be found for all anchor sample directions, the calculation of the decision variable is stopped. Note that the above steps are cyclical search steps. If the initial sample progresses to a subsequent or intermediate sample in the next feasible solution space, the subsequent or intermediate sample may be set as the current solution, and the above steps may be repeated.

[0057] As described above, the method for calculating decision variables in this specific embodiment involves obtaining the input decision variables by finding the solution to an optimization problem for the objective function of a trained predictive model. However, there may be interrelationships between these decision variables. For example, within a process, each process parameter (decision variable) is not arbitrarily set, but may have various constraints depending on the actual situation. For example, there may be inequalities or equalities between multiple process parameters. Therefore, when constructing and executing an optimization problem, these constraints must be considered in order to fit the final decision variables to the actual situation. In this specific embodiment, by adding a barrier function and a penalty function to the objective function of a trained predictive model, it is possible to transform an inherently constrained optimization problem into an unconstrained optimization problem, and therefore it is not necessary to consider the constraints in the subsequent process of finding the solution. This is because the constraints 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 so that they satisfy the inequality constraint. The barrier function includes a barrier region and a barrier function value. When the process parameters satisfy the inequality constraint, they are outside the barrier region, and the corresponding barrier function value is 0. That is, it does not affect the prediction results of the trained predictive model. On the other hand, when the process parameters do not satisfy the inequality constraint, they are inside the barrier region, and the corresponding barrier function value becomes large, which has a significant impact on the prediction results of the trained predictive model.

[0059] Furthermore, a penalty function is set to constrain the decision variables so that they satisfy the equality constraint. The penalty function includes a penalty region and a penalty function value. When the decision variables satisfy the equality constraint, they fall within the penalty region, and the corresponding penalty function value is 0. On the other hand, when the decision variables do not satisfy the equality constraint, they fall outside the penalty region, and the corresponding penalty function value becomes large, significantly impacting the prediction results of the trained predictive model. Therefore, by inputting sample parameters into a trained predictive model with a barrier function and / or penalty function, it is possible to not only check whether the generated prediction results match the target results, but also to determine whether the input parameters satisfy the actual constraints. If the input decision variables do not satisfy the conditions of the barrier function and / or penalty function, the prediction results will deviate significantly, meaning that the decision variables cannot be used because they do not satisfy the conditions.

[0060] The actual constraints mentioned above include, for example, that in the 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 practical applications, penalty functions and barrier functions are commonly used in machine learning to restrict the parameters of machine learning models to a reasonable range to fit real-world conditions in natural environments. For example, a cell culture process may have constraints such as ensuring that the cell culture temperature and relative humidity are not negative, and that the concentration of components in the culture medium does not harm the cells. By setting barrier and penalty functions to restrict the range of parameters, it becomes possible to comprehensively consider the relationships between parameters and find the combination of process parameters that best approaches the desired target result. In addition, in practical applications, barrier functions may be applied to constraints on pending decision variables, and penalty functions may be applied to constraints on confirmed decision variables. These can be used interchangeably depending on the user's needs.

[0062] The methods for calculating the decision variables in each of the above specific embodiments are applicable to the field of cell culture processes. Furthermore, when applied to the field of cell culture processes, the dataset used to train the trained predictive model may also include a cell dataset. The cell dataset includes multiple cell samples, and the sample parameters for each cell sample include source parameters and culture parameters. In practice, the cell dataset may be any data obtained from an open platform, or it may be data collected by the user themselves. In addition, the types of cell samples may include immune cells (e.g., dendritic cells (DC cells), cytokine-induced killer cells (CIK), tumor-infiltrating lymphocytes (TILs), natural killer cells (NK cells), and CAR-T cells), stem cells (e.g., peripheral blood stem cells, adipose-derived stem cells, bone marrow mesenchymal stem cells), chondrocytes, fibroblasts, etc., but are not limited to these in actual applications and should be determined according to the type of cell culture the user wants to perform. Moreover, in this specific embodiment, the source parameters for each cell sample may further include attribute data of the source of each cell sample. Each cell sample in a cell dataset may include data on the cell source and attribute data associated with that source. Attribute data may include physiological data of the source, or other source-related data such as the source's sex, age, medical history, living environment, and residential area. However, in practical applications, the source parameters of a cell sample may also include other parameters that may influence cellular processes and are related to the cell source.

[0063] Furthermore, the culture parameters for each cell sample include human, equipment, material, method, and environmental parameters for each cell sample. The cell culture process involves many steps, each of which is related to many culture parameters. These culture parameters include, for example, human-related parameters such as the sex 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 precision of the temperature and humidity control of the cell culture apparatus; material-related parameters such as the material of the cell culture dish and the components, ratios, and formulations 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 environmental-related parameters such as the ambient temperature, humidity, and carbon dioxide concentration of the cell culture environment.

[0064] While the above example illustrates its application to the field of cell culture processes, the method for calculating the decision variables in the present invention is not limited to applications in the field of cell processes. It may also be applied to any other field where results can be predicted by predictive models obtained through machine learning, and used for inverse estimation of the decision variables to be input in order to obtain the target result.

[0065] In the specific embodiment described above, the step of calculating the gradient by numerical analysis in the method for calculating the decision variable may further include the following steps, if described in detail: The amount of change in the sample parameter for each sample is set, this amount of change is added to the vector dimension of the sample parameter, and the sample parameter after adding the amount of change is input into the objective function of the trained prediction model to obtain the change in the function. After performing the above steps for all sample parameters, the gradient of the sample parameter for the first sample is calculated by concatenating the change in all functions with the sample parameter after adding the amount of change in all vector dimensions. In practice, the calculation of the gradient is not limited to this, and other methods may be selected and used to calculate the gradient depending on the user's needs, the type of machine learning model, or different usage situations.

[0066] In addition to the specific embodiments presented above, the method for calculating the decision variable in the present invention may also have other embodiments. Refer to Figure 4. Figure 4 shows a flowchart of the steps for calculating the decision variable based on another specific embodiment of the present invention. As shown in Figure 4, the method for calculating the decision variable in this specific embodiment includes the following:

[0067] Step S40: Provide a trained predictive model obtained by machine learning the dataset using a neural network method. The trained predictive model includes an input layer and an output layer. The number of input terminals in the input layer is equal to the sum of the number of pending and confirmed decision variables. The dataset contains multiple samples, and each sample contains multiple sample parameters.

[0068] Step S41: The confirmed decision variable is compared with the corresponding sample parameter among all sample parameters that corresponds to the confirmed decision variable, and a first sample is selected in which the corresponding sample parameter approximates or matches the confirmed decision variable. The confirmed decision variable and the other sample parameters in the first sample that do not belong to the corresponding sample parameter are then input into the trained predictive model. If the prediction result output from the trained predictive model matches the target result, the other sample parameters are used as initial weight values.

[0069] Step S42: A dummy layer is added, connected to the input layer of the trained prediction model. The dummy layer contains multiple artificial neurons, each connected to a separate input terminal of the input layer. The trained prediction model and the dummy layer form a parameter prediction model, and the prediction results of the trained prediction model become the output of the parameter prediction model.

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

[0071] Step S44: Based on the confirmed decision variables, set multiple first weight values ​​between the first artificial neuron and the input terminal, which are among the artificial neurons corresponding to the confirmed decision variables.

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

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

[0074] The trained predictive model in this specific embodiment is obtained by machine learning the dataset using artificial neural networks (ANNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), recursive neural networks (RecNNs), complex neural networks, or any other machine learning algorithm or neural network algorithm. The selection of the machine learning or neural network algorithm is made according to the user's needs.

[0075] In detail, the method for calculating the decision variables in this specific embodiment involves forming a parameter prediction model by adding a dummy layer connected to the input layer of the trained prediction model in step S42. The dummy layer contains the same number of artificial neurons as the input terminals of the input layer in the trained prediction model. Furthermore, a new link is constructed between each dummy artificial neuron and its corresponding input terminal. Subsequently, in step S43, the bias value of the activation function of the artificial neurons in the dummy layer is set to 0. This ensures that 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 becomes 1 (i.e., the output to the artificial neuron is 1). In this specific embodiment, the activation function makes it possible to store its own numerical value in the second weight value. Subsequently, in step S45, the newly formed dummy layer of the parameter prediction model (at this point, the dummy layer forms a new input layer of the parameter prediction model) is trained by inputting a training dataset containing multiple all-one vectors. Here, each all-one vector represents data that inputs 1 to each artificial neuron. The aforementioned target result can be the Ground Truth output from the parameter prediction model. During training, the optimizer can adjust the arc weights between the artificial neuron and the input terminal of the trained prediction model. It should be noted that in the training process of step S45, both the parameters and the first weight values ​​in the trained prediction model are fixed or frozen. Therefore, the optimizer cannot adjust these, and only adjusts the second weight value to minimize the difference between the output value of the parameter prediction model (prediction result) and the Ground Truth (target result).

[0076] At the end of training and upon convergence, the second weight values, adjusted by the optimizer, are combined with the fixed first weight values ​​to form the optimal decision variables (including confirmed and pending decision variables). Furthermore, 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) has the potential to acquire one set of second weight values ​​as the optimal input decision variable. In detail, if the pre-trained prediction model used by the user has excellent predictive effects, it is impossible to inversely estimate the input decision variables based solely on the predictive effects using the pre-trained prediction model. However, according to the parameter prediction model in this specific embodiment, it is possible to inversely estimate the second weight values ​​as unconfirmed input parameters or decision variables by methods such as adding a dummy layer, fixing the bias values ​​of the artificial neurons in the dummy layer, fixing the numerical values ​​input during training, freezing the parameters of the pre-trained prediction model, freezing the first weight values, and minimizing the difference between the output value and the target result.

[0077] In practical applications, optimizers for calculating decision variables 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 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), and RAdam (Rectified You can choose from one of the following: Adam, Lookahead, MadGrad (Momentumized, Adaptive, and Decentralized Gradient Descent), Yogi Optimizer (Yogi), or AdamMax (Adaptive Moment Estimation with Maximum).

[0078] The Adaptive Moment Estimation (Adam) optimizer described above (hereinafter referred to as the Adam optimizer) is used to adjust the second weight values ​​of artificial neurons. The Adam optimizer is commonly used to train various deep learning models, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers. The Adam optimizer can be applied to different tasks, such as image classification, natural language processing, and language translation. By adjusting the learning rate based on changes in the gradient of each parameter, the Adam optimizer allows for different learning rates in different directions. This contributes to better control over the convergence speed of machine learning training. In practical applications, the choice of optimizer is not limited to this; other optimizers with adjustable parameters and weight values ​​can be used based on user needs 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 to adjust the second weight value, which may become a pending decision variable, will decrease. Therefore, in steps S40 to S41 of this specific embodiment, the samples used to train the trained predictive model are 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 determine whether they are similar or match, and for each sample, it is confirmed whether the output of the trained predictive model matches the target result. If there is a sample in the dataset that matches both, it means that the prediction result of that sample not only matches the target set by the user, but the sample parameters corresponding to the confirmed decision variable (e.g., process parameters already executed in the process) are also similar or match, and that it fits the actual situation. Therefore, this sample can be used as an initial point when calculating the pending decision variable. For example, if the first 7 days of a 21-day cell culture process have already been performed, then the confirmed decision variable for the 7 days is already available. Therefore, the initial sample must not only meet the user's goal (e.g., culturing 95% viable cells) based on the prediction results output from the trained predictive model, but also have sample parameters for the first seven days of the sample that are the same as or approximately the same as the process parameters performed. Otherwise, the subsequent process parameters that the optimizer adjusts and obtains based on the initial sample will deviate from the actual situation.

[0080] It should be noted that the step actually performed in step S41 in this specific embodiment is the same as the method for comparing whether the first vector and the second vector are approximate or coincident, as described above. For example, a sample with the smallest included angle between the first vector and the second vector is obtained, or a sample with the smallest distance between the coordinates of the tip of the first vector and the coordinates of the tip of the second vector is calculated and obtained using a distance function. Detailed steps are described in detail in the specific embodiment above, so they will not be described in detail again here.

[0081] In summary, the present invention provides a method for calculating decision variables. This method involves selecting a cell reference sample in a trained predictive model through comparison, and then inversely estimating the unconfirmed decision variable by increasing the step size along the opposite direction of the gradient for the sample parameters of the cell reference sample. Furthermore, by setting barrier and penalty functions to limit the parameter range, the combination of process parameters that best approximates the desired target result is found, taking into account the relationships between parameters. Additionally, the unconfirmed decision variable is inversely estimated by setting and fixing sample parameters in the dataset, input values ​​to artificial neurons in the artificial neuron layer, and bias values, and then adjusting the second weight values ​​based on the target result and the first weight value using an optimizer. This effectively optimizes the process method and achieves the target result of the process. Moreover, the method for calculating decision variables in the present invention is not limited to applications in the field of cell processes, but may also be applied to any other trained machine learning model to inversely estimate process parameters and unconfirmed decision variables.

[0082] The detailed description of the above preferred specific embodiments is intended to more clearly describe the features and spirit of the invention and does not limit the scope of the invention by the preferred specific embodiments disclosed above. Rather, it is intended that various modifications and equivalent configurations are covered within the claims of the invention. Therefore, the claims of the invention should be interpreted most broadly based on the above description to cover all possible modifications and equivalent configurations. [Explanation of Symbols]

[0083] Steps S1-S6, S30-S36, S300-S306, S50-S60, S40-S50

Claims

1. A method for calculating decision variables for calculating multiple confirmation-awaited decision variables, A pre-trained predictive model is provided, the pre-trained predictive model is obtained by machine learning a dataset using a machine learning method, the pre-trained predictive model includes an input layer and an output layer, the dataset includes multiple samples, and each sample includes multiple sample parameters, the pre-trained predictive model is used to input multiple input variables through the input layer and to generate prediction results corresponding to the input variables through the output layer, The steps include receiving a target result corresponding to the prediction result of the aforementioned trained prediction model, The steps include selecting at least one of the aforementioned samples, in which the prediction result of the trained prediction model matches the target result, as at least one anchor sample, forming a multidimensional subspace with the at least one anchor sample, and determining the initial sample within the multidimensional subspace; The steps include obtaining the gradient by minimizing the objective function of the trained predictive model using the optimizer used to generate the trained predictive model, The steps include: increasing the step size of the sample parameters of the initial sample along the opposite direction of the gradient, and then inputting this into the trained prediction model to check whether the generated prediction result matches the target result; A method for calculating a decision variable, in which a computer performs the steps of: if the prediction result matches the target result, the sample parameter after increasing the step size is used as the decision variable awaiting confirmation.

2. The steps include providing multiple confirmed decision variables and calculating a first vector of the confirmed decision variables, Based on the confirmed decision variable, the steps include obtaining a plurality of corresponding sample parameters from the sample parameters of the sample and calculating a second vector of the corresponding sample parameters, A reference sample is obtained from the samples by comparing the first vector with the second vector of each sample, and by comparing the prediction result generated by each sample in the trained prediction model with the target result, wherein the second vector of the reference sample approximates or matches the first vector, and the prediction result of the reference sample approximates or matches the target result. A method for calculating a computer-based decision variable according to claim 1, comprising the step of using the reference sample as the initial sample.

3. A method for calculating the decision variable, A pre-trained predictive model is provided, the pre-trained predictive model is obtained by machine learning a dataset using a machine learning method, the dataset includes multiple samples, and each sample includes multiple sample parameters, the pre-trained predictive model is used to take multiple decision variables as input and generate a predictive result corresponding to the decision variables, the decision variables include a step of including multiple confirmed decision variables and multiple pending decision variables, The steps include receiving a target result corresponding to the prediction result of the aforementioned trained prediction model, The steps include: selecting candidate samples in which the corresponding sample parameters approximate or match the confirmed decision variable by comparing the confirmed decision variable with a plurality of corresponding sample parameters among the sample parameters that correspond to the confirmed decision variable; The steps include inputting the other sample parameters in the candidate samples that do not belong to the corresponding sample parameters, along with the confirmed decision variables, into the trained prediction model, and confirming whether the prediction result output from the trained prediction model matches the target result, The steps include: selecting candidate samples for all sample parameters in the dataset and confirming whether the prediction result matches the target result; selecting at least one of the candidate samples whose prediction result of the trained prediction model matches the target result as at least one anchor sample; and forming a multidimensional subspace with at least one anchor sample. The steps include determining an initial sample within the multidimensional subspace and obtaining the orientation of the initial sample within the multidimensional subspace, After increasing the step size along the aforementioned direction, the confirmed decision variable and other sample parameters in the initial sample that do not correspond to the confirmed decision variable are input into the trained prediction model to verify whether the generated prediction result matches the target result. A method for calculating a decision variable, in which a computer performs the steps of: if the prediction result matches the target result, the sample parameter after increasing the step size which does not correspond to the confirmed decision variable becomes the decision variable awaiting confirmation.

4. The step of selecting a candidate sample in which the corresponding sample parameter approximates or matches the confirmed decision variable by comparing the confirmed decision variable with the corresponding sample parameter among the sample parameters that corresponds to the confirmed decision variable further includes: The steps include: calculating the first vector of the confirmed decision variable, Based on the confirmed decision variable, the steps include obtaining the corresponding corresponding sample parameters from the sample parameters of the sample, The steps include calculating a second vector of the corresponding sample parameters, A method for calculating a decision variable performed by a computer according to claim 3, comprising the step of 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 the first sample among the samples approximates or matches the first vector.

5. Furthermore, A method for calculating a decision variable performed by a computer according to claim 4, comprising the steps of: calculating the binding angle between the first vector and the second vector of each sample, and setting the sample corresponding to the second vector having the smallest binding angle with respect to the first vector as the first sample.

6. Furthermore, A method for calculating a decision variable performed by a computer according to claim 4, comprising the steps of: calculating the distance between the coordinates of the tip of the first vector and the coordinates of the tip of the second vector of each sample using a distance function; and selecting the sample corresponding to the second vector having the smallest distance from the first vector as the first sample.

7. The step of determining the initial sample in the multidimensional subspace is further, A method for calculating a computer-based decision variable according to claim 3, comprising the step of setting the initial sample to the midpoint of the distance between at least two of the anchor samples in the multidimensional subspace.

8. The step of determining the initial sample in the multidimensional subspace is further, A method for calculating a computer-based decision variable according to claim 3, comprising the step of using a linear combination of at least two of the anchor samples in the multidimensional subspace as the initial sample.

9. The step of obtaining the orientation of the initial sample in the multidimensional subspace is further, A method for calculating a decision variable performed by a computer according to claim 3, comprising the step of calculating the first directional derivative of the initial sample along each of the anchor samples as the direction of the initial sample by numerical analysis.

10. The step of obtaining the orientation of the initial sample in the multidimensional subspace is further, A method for calculating a decision variable performed by a computer according to claim 3, comprising the steps of setting the initial sample as the origin and calculating the second direction derivative of the origin toward each anchor sample as the direction of the initial sample by a numerical analysis method.

11. The trained predictive model is represented by an objective function, and the step of obtaining the direction of the initial sample in the multidimensional subspace is further, A method for calculating a decision variable performed by a computer according to claim 3, comprising the step of obtaining the secant direction derivative of the initial sample toward each anchor sample by calculating the change in the function value of the objective function based on the change in the distance of the initial sample toward each anchor sample using a numerical analysis method.

12. The step of obtaining the orientation of the initial sample in the multidimensional subspace is further, The steps include sequentially increasing and adjusting the step size of the initial sample toward the directional derivative of each anchor sample, The steps include: increasing and adjusting the step size of the initial samples toward the directional derivative of each anchor sample, and then sequentially checking whether the prediction result output by inputting it into the trained prediction model comes closer to the target result; After increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, if it is confirmed that the prediction result output by inputting it into the trained prediction model approaches the target result further, the step of increasing the step size of the initial sample toward the directional derivative, with the directional derivative as the direction, A method for calculating a computer-operated decision variable according to claim 3, comprising the step of increasing the step size of the initial sample toward the directional derivative of any of the anchor samples, and if it is confirmed that the prediction result output by inputting it into the trained predictive model does not come any closer to the target result, then increasing the step size of the initial sample toward the directional derivative of another anchor sample, and then checking again whether the prediction result output by inputting it into the trained predictive model comes any closer to the target result.

13. A method for calculating decision variables for calculating multiple confirmation-awaited decision variables, A pre-trained predictive model is provided, which is obtained by machine learning a dataset using a neural network method, the pre-trained predictive model includes an input layer and an output layer, the number of input terminals in the input layer is equal to the sum of the number of pending decision variables and multiple confirmed decision variables, the dataset includes multiple samples, each of which includes multiple sample parameters, and the steps are as follows: The steps include: comparing the confirmed decision variable with a plurality of corresponding sample parameters from the sample parameters that correspond to the confirmed decision variable, selecting a first sample in which the corresponding sample parameters approximate or match the confirmed decision variable; inputting the confirmed decision variable and other sample parameters in the first sample that do not belong to the corresponding sample parameters into the trained prediction model; and if the prediction result output from the trained prediction model matches the target result, using the other sample parameters as initial weight values. The steps include: adding a dummy layer connected to the input layer of the trained prediction model, the dummy layer including a plurality of artificial neurons connected to each input terminal of the input layer, a parameter prediction model being formed by the trained prediction model and the dummy layer, and the prediction result of the trained prediction model becoming the output of the parameter prediction model; The steps include setting the bias value of the activation function of each artificial neuron to 0, so that when the input value of the activation function is 1, the output value of the activation function becomes 1, Based on the confirmed decision variable, the steps include setting a plurality of first weight values ​​between a plurality of first artificial neurons and the input terminal, respectively, among the artificial neurons corresponding to the confirmed decision variable, A method for calculating a decision variable, in which a computer performs the following steps: setting the output of the parameter prediction model corresponding to the target result, training the parameter prediction model by inputting a training dataset containing at least one all-one vector into the artificial neurons of the dummy layer of the parameter prediction model, and using an optimizer in a neural network method that has generated the trained prediction model, adjusts a plurality of second reference weight values ​​between each of the artificial neurons and the input layer, starting from the initial weight values, based on the target result, and sets the second reference weight values ​​as the confirmation-awaited decision variable.

14. The method for calculating a decision variable performed by a computer according to claim 13, wherein the trained predictive model includes a plurality of model parameters, all of which are fixed.

15. The method for calculating a decision variable executed by a computer according to claim 14, wherein the first weight value is fixed.

16. The method for calculating a decision variable executed by a computer according to claim 14, wherein the neural network method includes one of the following: artificial neural networks (ANN), convolutional neural networks (CNNs), recurrent neural networks (RNNs), recursive neural networks (RecNNs), and complex neural networks.

17. The optimizer further supports Adaptive Moment Estimation (Adam) optimization, Stochastic Gradient Descent (SGD), Momentum, Nesterov Accelerated Gradient (NAG), Adaptive Gradient Algorithm (AdaGrad), Nadam (Nesterov-Accelerated Adaptive Moment Estimation), and RMSprop (Root Mean Square). 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 A method for calculating a decision variable to be executed by a computer according to claim 14, which is selected from any of the following: Decentralized Gradient Descent, Yogi optimizer (Yogi), and AdamMax (Adaptive Moment Estimation with Maximum).

Images