# US 11,443,235 B2 · Identifying optimal weights to improve prediction accuracy in machine learning techniques > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 11,443,235 B2 - Document ID: US11443235B2 - Status: Granted - Kind: B2 - Application number: 16/684,396 - Filed: 2019-11-14 - Published: 2021-05-20 - Granted: 2022-09-13 - Earliest claimed: 2019-11-14 - Assignees: International Business Machines Corporation - Applicants: International Business Machines Corporation - Inventors: Jing Xu; Si Er Han; Steven George Barbee; Xue Ying Zhang; Ji Hui Yang - Web page: https://patentgrep.com/patent/US11443235B2 ## CPC - G06N 3/08: Learning methods - G06N 20/00: Machine learning - G06N 3/045: Combinations of networks - G06N 3/09: Supervised learning - G06N 3/092: Reinforcement learning - G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn - G06N 5/02: Knowledge representation; Symbolic representation - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract A computer-implemented method, system and computer program product for improving prediction accuracy in machine learning techniques. A teacher model is constructed, where the teacher model generates a weight for each data case. The current student model is then trained using training data and the weights generated by the teacher model. After training the current student model, the current student model generates state features, which are used by the teacher model to generate new weights. A candidate student model is then trained using training data and these new weights. A reward is generated by comparing the current student model with the candidate student model using training and testing data, which is used to update the teacher model if a stopping rule has not been satisfied. Upon a stopping rule being satisfied, the weights generated by the teacher model are deemed to be the “optimal” weights which are returned to the user. ## Claims ### Claim 1 A computer-implemented method for improving prediction accuracy in machine learning techniques, the method comprising: constructing a teacher model, wherein said teacher model generates a weight for each data case; training a current student model using training data and weights generated by said teacher model; generating state features by said current student model; generating new weights by said teacher model using said state features; training a candidate student model using said training data and said new weights; generating a reward by comparing said current student model with said candidate student model using said training data and testing data to determine which is better at predicting an observed target; updating said teacher model with said reward in response to a stopping rule not being satisfied; and returning said new weights and said current student model to a user in response to said stopping rule being satisfied, wherein said returned student model provides a prediction of said observed target. ### Claim 2 (depends on 1) The method as recited in claim 1 further comprising: determining whether said candidate student model generates a better prediction of said observed target than said current student model based on how close the prediction is to said observed target. ### Claim 3 (depends on 2) The method as recited in claim 2 further comprising: updating said current student model with said candidate student model and updating current weights with said new weights in response to said candidate student model generating a better prediction of said observed target than said current student model. ### Claim 4 (depends on 3) The method as recited in claim 3 further comprising: generating new state features by said updated student model using said new weights; and generating a second set of new weights by said teacher model using said new state features. ### Claim 5 (depends on 4) The method as recited in claim 4 further comprising: training said candidate student model using said training data and said second set of new weights; and generating a reward by comparing said updated student model with said candidate student model using said training data and said testing data to determine which is better at predicting said observed target. ### Claim 6 (depends on 2) The method as recited in claim 2 further comprising: generating a second set of new weights by said updated teacher model using said state features in response to said candidate student model not generating a better prediction of said observed target than said current student model. ### Claim 7 (depends on 6) The method as recited in claim 6 further comprising: training said candidate student model using said training data and said second set of new weights; and generating a reward by comparing said current student model with said candidate student model using said training data and said testing data to determine which is better at predicting said observed target. ### Claim 8 (depends on 1) The method as recited in claim 1, wherein said stopping rule comprises one or more of the following: reaching a specified number of trials, reaching a specified training time, converging of a prediction accuracy, and a user-initiated termination. ### Claim 9 (depends on 1) The method as recited in claim 1, wherein said teacher model comprises a neural network, wherein said student model comprises one of the following: a decision tree and a neural network. ### Claim 10 A computer program product for improving prediction accuracy in machine learning techniques, the computer program product comprising a computer readable storage medium having program code embodied therewith, the program code comprising the programming instructions for: constructing a teacher model, wherein said teacher model generates a weight for each data case; training a current student model using training data and weights generated by said teacher model; generating state features by said current student model; generating new weights by said teacher model using said state features; training a candidate student model using said training data and said new weights; generating a reward by comparing said current student model with said candidate student model using said training data and testing data to determine which is better at predicting an observed target; updating said teacher model with said reward in response to a stopping rule not being satisfied; and returning said new weights and said current student model to a user in response to said stopping rule being satisfied, wherein said returned student model provides a prediction of said observed target. ### Claim 11 (depends on 10) The computer program product as recited in claim 10, wherein the program code further comprises the programming instructions for: determining whether said candidate student model generates a better prediction of said observed target than said current student model based on how close the prediction is to said observed target. ### Claim 12 (depends on 11) The computer program product as recited in claim 11, wherein the program code further comprises the programming instructions for: updating said current student model with said candidate student model and updating current weights with said new weights in response to said candidate student model generating a better prediction of said observed target than said current student model. ### Claim 13 (depends on 12) The computer program product as recited in claim 12, wherein the program code further comprises the programming instructions for: generating new state features by said updated student model using said new weights; and generating a second set of new weights by said teacher model using said new state features. ### Claim 14 (depends on 13) The computer program product as recited in claim 13, wherein the program code further comprises the programming instructions for: training said candidate student model using said training data and said second set of new weights; and generating a reward by comparing said updated student model with said candidate student model using said training data and said testing data to determine which is better at predicting said observed target. ### Claim 15 (depends on 11) The computer program product as recited in claim 11, wherein the program code further comprises the programming instructions for: generating a second set of new weights by said updated teacher model using said state features in response to said candidate student model not generating a better prediction of said observed target than said current student model. ### Claim 16 (depends on 15) The computer program product as recited in claim 15, wherein the program code further comprises the programming instructions for: training said candidate student model using said training data and said second set of new weights; and generating a reward by comparing said current student model with said candidate student model using said training data and said testing data to determine which is better at predicting said observed target. ### Claim 17 (depends on 10) The computer program product as recited in claim 10, wherein said stopping rule comprises one or more of the following: reaching a specified number of trials, reaching a specified training time, converging of a prediction accuracy, and a user-initiated termination. ### Claim 18 A system, comprising: a memory for storing a computer program for improving prediction accuracy in machine learning techniques; and a processor connected to said memory, wherein said processor is configured to execute the program instructions of the computer program comprising: constructing a teacher model, wherein said teacher model generates a weight for each data case; training a current student model using training data and weights generated by said teacher model; generating state features by said current student model; generating new weights by said teacher model using said state features; training a candidate student model using said training data and said new weights; generating a reward by comparing said current student model with said candidate student model using said training data and testing data to determine which is better at predicting an observed target; updating said teacher model with said reward in response to a stopping rule not being satisfied; and returning said new weights and said current student model to a user in response to said stopping rule being satisfied, wherein said returned student model provides a prediction of said observed target. ### Claim 19 (depends on 18) The system as recited in claim 18, wherein the program instructions of the computer program further comprise: determining whether said candidate student model generates a better prediction of said observed target than said current student model based on how close the prediction is to said observed target. ### Claim 20 (depends on 19) The system as recited in claim 19, wherein the program instructions of the computer program further comprise: updating said current student model with said candidate student model and updating current weights with said new weights in response to said candidate student model generating a better prediction of said observed target than said current student model. ### Claim 21 (depends on 20) The system as recited in claim 20, wherein the program instructions of the computer program further comprise: generating new state features by said updated student model using said new weights; and generating a second set of new weights by said teacher model using said new state features. ### Claim 22 (depends on 21) The system as recited in claim 21, wherein the program instructions of the computer program further comprise: training said candidate student model using said training data and said second set of new weights; and generating a reward by comparing said updated student model with said candidate student model using said training data and said testing data to determine which is better at predicting said observed target. ### Claim 23 (depends on 19) The system as recited in claim 19, wherein the program instructions of the computer program further comprise: generating a second set of new weights by said updated teacher model using said state features in response to said candidate student model not generating a better prediction of said observed target than said current student model. ### Claim 24 (depends on 23) The system as recited in claim 23, wherein the program instructions of the computer program further comprise: training said candidate student model using said training data and said second set of new weights; and generating a reward by comparing said current student model with said candidate student model using said training data and said testing data to determine which is better at predicting said observed target. ### Claim 25 (depends on 18) The system as recited in claim 18, wherein said stopping rule comprises one or more of the following: reaching a specified number of trials, reaching a specified training time, converging of a prediction accuracy, and a user-initiated termination.