Method for calculating decision variable

Abstract

To provide a method for calculating decision variables that uses general machine learning platforms to efficiently find optimal input parameters that achieve expected results.SOLUTION: A method includes adding a dummy layer at an input terminal of a pre-trained neural network predictive model. The dummy layer contains the same number of artificial neurons as the number of the input terminals of the pre-trained predictive model. Each of the artificial neurons is connected to its neuron of the input terminal of the trained predictive model through a newly established link. The input value of each of the artificial neurons is set to 1, a bias value of an activation function is set to 0, and when the input value of the activation function is 1, its output value is also 1. The initial weights of the newly established links are selected and set, with these weights considered decision variables having ranges or other inter-conditional restrictions. The parameters of the pre-trained predictive model are frozen, and only the weights of the newly established links are adjusted to obtain the optimal solution using an optimizer built into a general machine-learning platform.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a method for calculating decision variables, and more particularly to a method for calculating decision variables that utilizes an optimizer built into an artificial intelligence neural network engine training platform to inversely estimate decision variables and process parameters. [Background technology]

[0002] Advances in medical engineering technology have not only had a significant impact on the healthcare field, but also increased the application of regenerative medicine in clinical practice. 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 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 tissue engineering and regenerative therapy, organ transplantation and regeneration, tissue regeneration and repair, cancer treatment, nerve regeneration, immune cell therapy, and stem cell therapy. In particular, research and applications of cell therapy in regenerative medicine are attracting increasing attention from various fields. Cell therapy involves the use of human cells by transplanting them into an individual's body after undergoing a procedure of culturing or processing cells from the human body outside the body.

[0003] Cell therapy has the advantages of personalized treatment, avoidance of rejection, replacement of organ transplants and restoration of tissue function. Currently, many governments are gradually lifting the ban on the application of cell therapy in their countries, and many scholars at home and abroad are entering the research of cell therapy one after another. As a result, cell therapy has made remarkable progress in research 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 injuries with autologous bone marrow mesenchymal stem cells. According to previous research, each human body has its own unique characteristics, so the types of cell preparations applied in each case are different according to the difference in disease and symptoms. Therefore, it is necessary to customize cells suitable for treatment according to each individual situation, which increases the complexity and difficulty of cells in the process. 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 in quality during the culture process. However, the variability of cells among different cases is extremely high, so the optimal culture parameters and environmental parameters of cell preparations applied to different cases are not completely the same. Also, each cell preparation needs to adjust each decision variable (i.e., process parameter) in the process to achieve the desired result, so each cell preparation cannot be produced with fixed decision variables.

[0004] In the conventional technology, a predictive model is obtained by training on a large data set of samples using machine learning. Then, the parameters of the predictive model are adjusted using a machine learning optimizer to minimize or maximize the objective function. This improves the accuracy of the predictive model and accelerates the convergence process of the predictive model, further reducing the time and computational cost required for machine learning and convergence. In addition, this predictive model can input various decision variables to generate predicted results of cell culture, allowing users to simulate the effects of the cell culture process they are designing in advance.

[0005] However, the above prediction models trained by machine learning cannot inversely estimate the decision variables of the cell culture process from the user's desired results. In other words, when designing a process, the user needs to try a large number of different decision variables, so that a large amount of effort and resources are spent on designing and improving the culture process. In addition, not only the prediction of parameters is required in the cell culture process, but also other machine learning models for training neural networks require inverse estimation of decision variables and process parameters.

[0006] Therefore, it is necessary to develop a method to further optimize the process by being able to inversely estimate the decision variables of the process that match the predicted results. Summary of the Invention [Problem to be solved by the invention]

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

[0008] The present invention provides a method for calculating a decision variable, which includes the steps of: providing a trained prediction model, the trained prediction model being obtained by machine learning using a machine learning method, the trained prediction model including an input layer and an output layer, the trained prediction model being used to input a number of input parameters through the input layer and generate a prediction result corresponding to the input parameters through the output layer; setting a target result corresponding to the prediction result of the trained prediction model and providing at least one confirmed input parameter among the input parameters; adding a dummy layer connected to the input layer of the trained prediction model, the dummy layer including a number 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, the prediction result of the trained prediction model being the output of the parameter prediction model; and setting a bias value of an activation function of each artificial neuron to 0. setting at least one first weight value of at least one first artificial neuron among the artificial neurons corresponding to the at least one confirmed input parameter based on the at least one confirmed input parameter; setting an output of the parameter prediction model corresponding to the target outcome, training the parameter prediction model by inputting a training data set including at least one all-ones vector into the artificial neurons of the dummy layer of the parameter prediction model, and adjusting second weight values ​​of the artificial neurons other than the at least one first artificial neuron based on the target outcome and the at least one first weight value by an optimizer in the machine learning method that generated the trained prediction model, and setting the second weight values ​​to a plurality of unconfirmed input parameters among the input parameters.

[0009] The optimizer also includes 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 Randomness (RAdam), and Optimal Randomness (UDM). 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).

[0010] The trained predictive model is obtained by machine learning a dataset using the machine learning method.

[0011] The trained predictive model includes a number of model parameters, all of which are fixed.

[0012] The trained predictive model is obtained by machine learning the dataset using any of the following: Artificial Neural Networks (ANN), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Recursive Neural Networks (RecNNs), and Complex Neural Networks.

[0013] In another aspect, the present invention provides a method for calculating decision variables used to calculate a plurality of decision variables. The method includes the steps of: providing a trained prediction model, the trained prediction model being obtained by machine learning using a machine learning method, the trained prediction model including an input layer and an output layer, the trained prediction model being used to input a plurality of input parameters through the input layer and generate a prediction result corresponding to the input parameters through the output layer; setting a target result corresponding to the prediction result of the trained prediction model; 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, a parameter prediction model being formed by the trained prediction model and the dummy layer, 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 artificial neuron to 0, so that when an input value of the activation function is 1, an output value of the activation function is 1; setting an 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 into the artificial neurons of the dummy layer of the parameter prediction model, and adjusting a reference weight value of each artificial neuron based on the target result by an optimizer in the machine learning method that generated the trained prediction model, and setting the reference weight value as a decision variable.

[0014] The optimizer also includes 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 Randomness (RAdam), and Optimal Randomness (UDM). 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).

[0015] The trained predictive model is obtained by machine learning a dataset using the machine learning method.

[0016] The trained predictive model includes a number of model parameters, all of which are fixed.

[0017] The trained predictive models are obtained by machine learning the dataset using any of the following: Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Recursive Neural Networks (RecNNs), and Complex Neural Networks. Effect of the Invention

[0018] In summary, the present invention provides a method for calculating decision variables. In the method, a dummy layer is added to the input terminal of a trained neural network prediction model. The dummy layer includes the same number of artificial neurons as the input terminal of the trained prediction model. A new link is established between each artificial neuron and the corresponding neuron at the input terminal of the trained prediction model. The input value of each artificial neuron is set to 1. When the bias value of the activation function is 0 and the input of the activation function is 1, the output is 1. Then, the initial value of the weight of the new link is selected and set, and the weight value is regarded as a decision variable. The weight value may have a range or other mutual constraints. The parameters of the trained prediction model are frozen, and only the weight value of the new link is adjusted, and an optimal solution is obtained by an optimizer built into a general machine learning platform. The goal of training is set so that the output of this parameter prediction model is the desired target result. The weight value of the new link obtained when training is completed becomes a feasible input decision variable. In addition, the method for calculating decision variables in the present invention can be applied to other machine learning models trained by neural networks in addition to the application in the field of cellular processes, to inversely estimate decision variables and process parameters. The present invention can utilize a general machine learning platform and the method embedded therein to effectively search for optimal input parameters that achieve the desired results. [Brief description of the drawings]

[0019] [Figure 1A] FIG. 1A shows a functional block diagram of a method for calculating decision variables according to a specific embodiment of the present invention. [Figure 1B] FIG. 1B shows a functional block diagram 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 steps of a method for calculating decision variables according to an illustrative embodiment of the present invention. [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. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

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

[0021] Please refer to Figure 1A, Figure 1B and Figure 2. Figure 1A and Figure 1B show a functional block diagram of a method for calculating decision variables according to a specific embodiment of the present invention. Figure 2 shows a flowchart of steps of the method for calculating decision variables according to a specific embodiment of the present invention. As shown in Figure 2, in this specific embodiment, the method for calculating decision variables includes:

[0022] Step S1: Provide a trained prediction model 10. The trained prediction model 10 is obtained by machine learning using a machine learning method. The trained prediction model 10 includes an input layer 11 and an output layer 12. The trained prediction model 10 is used to input a number of input parameters through the input layer 11 and generate a prediction result (i.e., an output result) corresponding to the input parameters through the output layer 12.

[0023] Step S2: Setting a target result corresponding to the predicted result of the trained predictive model 10, and providing at least one confirmed input parameter among the input parameters.

[0024] Step S3: A dummy layer 20 is added, which is connected to the input layer 11 of the trained prediction model 10. The dummy layer 20 includes a plurality of artificial neurons 21, each of which is connected to each input terminal 111 of the input layer 11. The trained prediction model 10 and the dummy layer 20 form a parameter prediction model A, and the prediction result of the trained prediction model 10 becomes the output of the parameter prediction model A.

[0025] Step S4: Set the bias value of the activation function of each artificial neuron 21 to 0. When the input value of the activation function is 1, the output value of the activation function is 1.

[0026] Step S5: Based on the at least one confirmed input parameter, respectively set at least one first weight value of at least one first artificial neuron (not shown) of the artificial neurons 21 corresponding to the at least one confirmed input parameter.

[0027] Step S6: Set the output of the parameter prediction model A corresponding to the target result, and input a training data set including at least one all-ones vector into the artificial neurons 21 of the dummy layer 20 of the parameter prediction model A to train the parameter prediction model A. Then, an optimizer (not shown) in the machine learning method that generated the trained prediction model 10 adjusts second weight values ​​of the artificial neurons 21 other than the at least one first artificial neuron 21 based on the target result and the at least one first weight value, and sets the second weight value as a plurality of unconfirmed input parameters among the input parameters.

[0028] In addition, in the method for calculating decision variables of this specific embodiment, the trained prediction model 10 includes a number of model parameters, all of which are fixed, and the first weight value is a fixed value and is not allowed to be changed or adjusted by the optimizer, while only the second weight value in the artificial neurons 21 other than the first artificial neuron can be adjusted by the optimizer.

[0029] 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 model that has been trained by the user himself. In addition, the dataset in this specific embodiment may be a set formed of any data that can be used for training, testing, and validation of machine learning. 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 unconfirmed input parameters. In addition, 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), recursive neural networks (RecNN), and 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 needs of the user.

[0030] In detail, in the method for calculating the decision variables in this specific embodiment, a dummy layer 20 is added to be connected to the trained prediction model 10 in step S3 to form a parameter prediction model A. The dummy layer 20 includes the same number of artificial neurons 21 as the input terminals 111 of the trained prediction model 10. A new link is also established between each of the dummy artificial neurons 21 and the corresponding neuron at the input terminal 111 of the trained prediction model 10. In addition, in step S4, the bias value of the activation function of the artificial neuron 21 in the dummy layer 20 is set to 0. When the input value of the activation function is 1, the output value of the activation function 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 S6, a training data set including a plurality of all-one vectors is input to the dummy layer 20 of the newly formed parameter prediction model A (at this time, the dummy layer 20 forms a new input layer of the parameter prediction model A) to perform training. Here, each all-one vector represents data for inputting 1 to each artificial neuron 21. The target result may be the ground truth output from the parameter prediction model A. In addition, the initial weights of the new links are selected and set, and the second weights are regarded as decision variables. There may be ranges or other mutual constraints between the weights, such as positive integer constraints or range constraints. In the training process of the parameter prediction model A, the optimizer adjusts the second weights based on the target result, the training data set, and the first weights, respectively, to inversely estimate unidentified decision variables (unidentified input parameters).

[0031] In detail, the optimizer is built into the training platform of the artificial intelligence engine and has a function of adjusting or fixing the parameters and weight values ​​of the parameter prediction model A and the trained prediction model 10. Therefore, in the training process, the optimizer fixes or freezes the parameters of the trained prediction model 10, and then adjusts the weight value 30 between the artificial neuron 21 and the input terminal 111 of the trained prediction model 10, thereby minimizing the difference value between the output value (prediction result) of the parameter prediction model A corresponding to each all-one vector in the training data set and the ground truth (target result). It should be noted that the confirmed input parameters are fixed parameters, for example, process parameters that have been executed in the process. Therefore, in the above training process, the optimizer also fixes or freezes the first weight value corresponding to the confirmed input parameter, and adjusts other weight values ​​(second weight values) other than the first weight value, thereby minimizing the difference value between the output value (prediction result) of the parameter prediction model A and the ground truth (target result). At the end of training and at the time of convergence, the second weight value obtained by adjusting by the optimizer is combined with the fixed first weight value to become the optimal input decision variable. The parameter prediction model A has a many-to-one characteristic. That is, a plurality of different input parameters can correspond to the same output result. Therefore, each training data (all-one vector) has a possibility of obtaining a set of second weight values ​​as the optimal input decision variable. In detail, when the trained prediction model used by the user has an excellent prediction effect, it is impossible to reverse estimate the input decision variable based on the trained prediction model from the prediction effect alone. However, according to the parameter prediction model in this specific embodiment, it is possible to reverse estimate the second weight value as an unconfirmed input parameter or decision variable by methods such as 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 value, and minimizing the difference between the output value and the target result.

[0032] The following describes an example of a cell process. A user may be trying to carry out a 21-day process and has already carried out the process for 7 days (i.e., already has confirmed input parameters), but may not have confirmed how to carry out the process for the 8th to 21st days and adjust the decision variables (i.e., unconfirmed input parameters) to obtain a desired target result on the 21st day that is set in advance. In response to this, according to the above method, it is possible to set the first weight value of the first artificial neuron among the artificial neurons corresponding to the confirmed input parameters based on the confirmed input parameters carried out for 7 days, and set the target result (for example, to achieve 95% cell survival rate in a 21-day cell culture process). Then, an optimizer in the machine learning method that generated the trained prediction model inputs an all-ones vector, and adjusts the second weight values ​​of the other artificial neurons other than the first artificial neuron based on the difference value between the target result and the output of the parameter prediction model and the first weight value, and sets the second weight value as the decision variable (unconfirmed input parameter) for the 8th to 21st days. That is, if the second weight value is adjusted to 2 by the optimizer, the unconfirmed input parameter becomes 2.

[0033] 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).

[0034] The optimizer of the method for calculating decision variables in this specific embodiment further includes an Adaptive Moment Estimation (Adam) optimizer (hereinafter referred to as Adam optimizer). The Adam optimizer is also 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 with adjustable parameters and weight values ​​can be used based on the needs of the user and the type of training model.

[0035] The trained prediction model in this specific embodiment can be obtained by machine learning the dataset by a machine learning method. When the method for calculating the decision variables in this specific embodiment is applied to the field of cell culture process, the dataset in this specific embodiment may further include a cell dataset. The cell dataset may include 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 thereto in practical application 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. The attribute data 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, and residential area. However, in practical applications, the source parameters of the cell sample may further include other parameters related to the cell source that may affect the cellular process. In practical applications, the trained prediction model can be selected based on the needs of a user and can be obtained by selecting different types of datasets and performing machine learning.

[0036] In addition, in this specific embodiment, 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 control 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.

[0037] The above specific embodiment is for a process that includes confirmed input parameters, i.e., a process that has already been started. However, the method for calculating decision variables in the present invention is also applicable to a process that has not yet been started and does not have confirmed input parameters. Now, please refer to FIG. 3. FIG. 3 shows a flowchart of the steps of the method for calculating decision variables according to another specific embodiment of the present invention. The method for calculating decision variables in this specific embodiment is used to calculate multiple decision variables. The method includes:

[0038] Step S1': Provide a trained prediction model 10. The trained prediction model 10 is obtained by machine learning using a machine learning method. The trained prediction model 10 includes an input layer 11 and an output layer 12. The trained prediction model 10 is used to input a number of input parameters through the input layer 11 and generate a prediction result corresponding to the input parameters through the output layer 12.

[0039] Step S2': Set a target result corresponding to the prediction result of the trained prediction model 10.

[0040] Step S3': A dummy layer 20 is added to be connected to the input layer 11 of the trained prediction model 10. The dummy layer 20 includes a plurality of artificial neurons 21 each connected to each input terminal 111 of the input layer 11. A parameter prediction model A is formed by the trained prediction model 10 and the dummy layer 20, and the prediction result of the trained prediction model 10 becomes the output of the parameter prediction model A.

[0041] Step S4': The bias value of the activation function of each of the artificial neurons 21 is set to 0. When the input value of the activation function is 1, the output value of the activation function is 1.

[0042] Step S5': Set an output of the parameter prediction model A corresponding to the target result, and input a training data set including at least one all-ones vector to the artificial neurons 21 of the dummy layer 20 of the parameter prediction model A to train the parameter prediction model A. Then, an optimizer in the machine learning method that generated the trained prediction model 10 adjusts the reference weight values ​​of each of the artificial neurons 21 based on the target result, and sets the reference weight values ​​as the decision variables.

[0043] The method for calculating decision variables in this specific embodiment is suitable for a process that does not have confirmed input parameters, i.e., the first weight value of the confirmed input parameter is an empty set. The optimizer adjusts the reference weight value of each artificial neuron to enable inverse estimation by minimizing the difference value between the output value and the target result. That is, the reference weight value can be the decision variable (i.e., the unconfirmed decision variable). It should be noted that the other steps of the method for calculating decision variables in this specific embodiment are almost the same as the corresponding steps in the above specific embodiment, and will not be described in detail here.

[0044] In summary, the present invention provides a method for calculating decision variables. In the method, a dummy layer is added to the input terminal of a trained neural network prediction model. The dummy layer includes the same number of artificial neurons as the input terminal of the trained prediction model. A new link is constructed between each dummy neuron and the corresponding neuron at the input terminal of the original model. The input value of each artificial neuron is set to 1. When the bias value of the activation function is 0 and the input of the activation function is 1, the output is 1. Then, the initial value of the weight of the new link is selected and set, and the weight value is regarded as a decision variable. The weight value may have a range or other mutual constraints. The parameters of the trained prediction model are frozen, and only the weight value of the new link is adjusted, and an optimal solution is obtained by an optimizer built into a general machine learning platform. The goal of training is set so that the output of this extended model is the desired target result. The weight value of the new link obtained when training is completed becomes a feasible input decision variable. In addition, the method for calculating decision variables in the present invention can be applied to other machine learning models trained by neural networks in addition to the application in the field of cellular processes, to inversely estimate decision variables and process parameters. The present invention can utilize a general machine learning platform and the method embedded therein to effectively search for optimal input parameters that achieve the desired results.

[0045] 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]

[0046] S1~S6, S1'~S5' steps A Parameter Prediction Model 10 Pre-trained predictive models 11 Input Layer 111 Input terminal 12 Output layer 20 Dummy layer 21 Artificial Neuron 30 Weight Value

Claims

1. Providing a trained prediction model, the trained prediction model being obtained by machine learning through a machine learning method, the trained prediction model including an input layer and an output layer, the trained prediction model being used to input a plurality of input parameters through the input layer and generate a prediction result corresponding to the input parameters through the output layer; setting a target outcome corresponding to the predicted outcome of the trained predictive model and providing at least one verified input parameter among the input parameters; adding a dummy layer connected to the input layer of the trained prediction model, the dummy layer including a plurality of artificial neurons each 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 becoming an 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 at least one confirmed input parameter, at least one first weight value of at least one first one of the artificial neurons corresponding to the at least one confirmed input parameter; setting an output of the parameter prediction model corresponding to the target outcome; 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; and adjusting second weight values ​​of the artificial neurons other than the at least one first artificial neuron based on the target outcome and the at least one first weight value by an optimizer in the machine learning method that generated the trained predictive model, and setting the second weight values ​​to a plurality of unconfirmed input parameters among the input parameters.

2. 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 2. The method of claim 1, wherein the optimization method is selected from the group consisting of Decision Optimizer (DEC), Decentralized Gradient Descent (DGD), Yogi optimizer (Yogi), and AdamMax (Adaptive Moment Estimation with Maximum).

3. The method for calculating decision variables according to claim 1 , wherein the trained predictive model is obtained by machine learning a dataset using the machine learning method.

4. The method of claim 1 , wherein the trained predictive model includes a plurality of model parameters, all of which are fixed.

5. 2. The method of claim 1, wherein the trained predictive model is obtained by machine learning the dataset using any 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).

6. used to calculate multiple decision variables, Providing a trained prediction model, the trained prediction model being obtained by machine learning through a machine learning method, the trained prediction model including an input layer and an output layer, the trained prediction model being used to input a plurality of input parameters through the input layer and generate a prediction result corresponding to the input parameters through the output layer; setting a target outcome corresponding to the predicted outcome of the trained predictive model; adding a dummy layer connected to the input layer of the trained prediction model, the dummy layer including a plurality of artificial neurons each 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 becoming an 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; setting an output of the parameter prediction model corresponding to the target outcome, 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, and adjusting a reference weight value of each of the artificial neurons based on the target outcome by an optimizer in the machine learning method that generated the trained predictive model, and setting the reference weight value as the decision variable.

7. 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 7. The method of claim 6, wherein the optimization method is selected from the group consisting of Decision Optimizer (DEC), Decentralized Gradient Descent (DGD), Yogi optimizer (Yogi), and AdamMax (Adaptive Moment Estimation with Maximum).

8. The method for calculating decision variables according to claim 6 , wherein the trained predictive model is obtained by machine learning a dataset using the machine learning method.

9. The method of claim 6 , wherein the trained predictive model includes a plurality of model parameters, all of which are fixed.

10. 7. The method of claim 6, wherein the trained predictive model is obtained by machine learning the dataset using any 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).

Images