SYSTEM AND METHOD FOR AUTOMATED DATA SCIENCE PROCESSES - Patent application

An integrated framework addresses the limitations of existing data science processes by automating tasks and integrating offline and online processes, enhancing model performance and adaptability through customizable and extensible AutoML techniques.

JP7591128B2Active Publication Date: 2024-11-27HITACHI VANTARA LLC
View PDF 12 Cites 0 Cited by

Patent Information

Application Number
JP2023510484
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-08-20
Publication Date
2024-11-27
Estimated Expiration
2040-08-20

AI Technical Summary

Technical Problem

Existing data science processes lack comprehensive support for descriptive, exploratory, predictive, prescriptive, and autonomous learning components, are not customizable, and do not integrate offline and online processes, requiring significant manual labor and being limited to static machine learning libraries.

Method used

An integrated, customizable, and extensible framework that supports descriptive, exploratory, predictive, prescriptive, and autonomous tasks, seamlessly combining offline and online processes, and automates data science tasks using AutoML techniques.

Benefits of technology

Reduces manual labor, improves model performance and development time, and enables real-time adaptation to dynamic systems by integrating offline and online processes, supporting various machine learning libraries and automating common tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007591128000001
    Figure 0007591128000001
  • Figure 0007591128000002
    Figure 0007591128000002
  • Figure 0007591128000003
    Figure 0007591128000003
Patent Text Reader

Abstract

Example embodiments described herein are directed to systems and methods for automated, autonomous, self-learning machine learning model generation and deployment, which may include generating predictive and prescriptive models through an offline learning process in a first system, controlling the operation of a second system through deploying the predictive and prescriptive models to the second system, and autonomously updating the predictive and prescriptive models from feedback from the second system through an online learning process once the prescriptive and predictive models have been deployed to the second system.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates generally to data science techniques, and more particularly, to data science processes and automated machine learning (AutoML). [Background technology]

[0002] In the prior art, the data science process defines a methodology for efficiently and effectively delivering analytical solutions and intelligent applications. Automated Machine Learning (AutoML) is a system of frameworks that can automatically build models for given data, thereby automating a maximum number of steps in the ML pipeline, minimizing human effort, and improving model performance.

[0003] Prior art implementations include AutoML frameworks that include a predictive component. Prior art AutoML frameworks include static implementations that are specialized to a particular ML implementation and are generally not customizable. Additionally, such prior art implementations do not utilize an automated data science process. Summary of the Invention [Problem to be solved by the invention]

[0004] Prior art implementations have no components to support descriptive and exploratory analysis, prescriptive, automated, or autonomous learning components. Exemplary implementations described herein are directed to supporting these components, in addition to predictive components, through a unified, highly customizable and extensible framework configured to provide an automated data science process.

[0005] Traditional data science processes have several challenges. First, traditional data science processes are not comprehensive enough to support value-driven tasks. There is a need for a comprehensive process that allows data science professionals to understand how to specifically drive value. Second, traditional data science processes focus primarily on offline processes only. In the prior art, no online data science process has been proposed, and no data science process exists that combines both offline and online processes. There is a need for a data science process that supports both offline and online processes. Online processes may be important in real-world systems, as they aid in automation and autonomous learning to obtain best-fit models based on real-time data in dynamic systems. Furthermore, traditional data science processes require a significant amount of manual work by humans. There is a need for a system that performs tasks in the data science process automatically.

[0006] Prior art implementations of AutoML frameworks suffer from several deficiencies. First, prior art AutoML frameworks only address the "prediction" aspect of data science tasks, while not automating common tasks in other components of the data science process. There is a need for a system that automates common tasks in the data science process.

[0007] In prior art implementations, each AutoML library supports only one underlying machine learning library. There is a need to support a unified and extensible system that supports a variety of machine learning libraries. [Means for solving the problem]

[0008] To address the above needs, example implementations described herein include a comprehensive data science system that is descriptive, exploratory, predictive, prescriptive, automated, and autonomous, and configured to support value-driven tasks, including offline and online processes, which are seamlessly integrated together as one system.

[0009] In the exemplary implementation described herein, there is an offline process that defines the methodology and workflow for all data science tasks on historical data, which corresponds to the first four components that support the descriptive, exploratory, predictive, and prescriptive aspects of the system.

[0010] In the exemplary implementation described herein, there is an online process that defines the methodology and workflow for all data science tasks on real-time data, which addresses the last four components that support the predictive, prescriptive, automated, and autonomous aspects of the system.

[0011] In the exemplary implementations described herein, there is an automated system for data science processes where AutoML is applied to each step of the process to reduce manual effort and optimize outcomes. A unified, customizable and extensible framework is introduced and provided to support the system.

[0012] Aspects of the disclosure may include a method for automated, autonomous, self-learning machine learning model generation and deployment, the method including generating predictive and prescriptive models through an offline learning process in a first system, controlling operation of a second system through deploying the predictive and prescriptive models to the second system, and autonomously updating the prescriptive and predictive models from feedback from the second system through an online learning process once the prescriptive and predictive models have been deployed to the second system.

[0013] Aspects of the disclosure may include a system for automated, autonomous, self-learning machine learning model generation and deployment, the system including means for generating predictive and prescriptive models through an offline learning process in a first system, means for controlling operation of the second system through deploying the predictive and prescriptive models to the second system, and means for autonomously updating the prescriptive and predictive models from feedback from the second system through an online learning process once the prescriptive and predictive models have been deployed to the second system.

[0014] Aspects of the disclosure may include a non-transitory computer-readable medium storing instructions for automated and autonomous generation and deployment of self-learning machine learning models, the instructions including generating predictive and prescriptive models through an offline learning process in a first system, controlling operation of a second system through deploying the predictive and prescriptive models to the second system, and autonomously updating the prescriptive and predictive models from feedback from the second system through an online learning process once the prescriptive and predictive models have been deployed to the second system.

[0015] Aspects of the disclosure may include a system for automated, autonomous, self-learning machine learning model generation and deployment, the system comprising a first system with a processor configured to generate predictive and prescriptive models through an offline learning process, control operation of a second system through deployment of the predictive and prescriptive models to the second system, and autonomously update the prescriptive and predictive models from feedback from the second system through an online learning process once the prescriptive and predictive models have been deployed to the second system. [Brief description of the drawings]

[0016] [Figure 1a] 1 illustrates an example of an automated system for both offline and online processes, according to an exemplary implementation.

[0017] [Figure 1b] 1 illustrates an example flow of an offline process, according to an example implementation.

[0018] [Diagram 2] 1 illustrates an example workflow of an online system when integrated with an offline system, according to an example implementation.

[0019] [Diagram 3] 1 illustrates an example of automated adaptive cohort learning, according to an exemplary implementation.

[0020] [Figure 4] 1 illustrates an example of a solution configuration file according to an exemplary implementation.

[0021] [Figure 5a] FIG. 1 illustrates an example flow diagram for automated, autonomous, self-learning machine learning model generation and deployment, according to an example implementation. [Figure 5b]FIG. 1 illustrates an example flow diagram for automated, autonomous, self-learning machine learning model generation and deployment, according to an example implementation.

[0022] [Figure 6] 1 illustrates a number of systems and a management device according to an exemplary implementation.

[0023] [Figure 7] 1 illustrates an exemplary computing environment including an exemplary computing device suitable for use in some exemplary embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] The following detailed description provides details of the drawings and exemplary implementations of the present application. For clarity, reference numbers and descriptions of elements that overlap between drawings are omitted. Terms used throughout the description are presented by way of example and are not intended to be limiting. For example, the use of the term "automatic" may include fully automatic or semi-automatic implementations that include user or administrator control over certain aspects of the implementation, depending on the desired implementation of the person skilled in the art implementing the implementation of the present application. Selection may be performed by a user via a user interface or other input means, or may be implemented via a desired algorithm. The exemplary implementations described herein may be utilized alone or in combination, and the functionality of the exemplary implementations may be implemented via any means according to the desired implementation.

[0025] Example implementations described herein are directed to a system that supports a comprehensive data science process that helps drive business value, including descriptive, exploratory, predictive, prescriptive, automated, and autonomous components.

[0026] The Describe component covers the understanding and evaluation of the business problem, data, and technology. The Describe step provides prioritization, feasibility, return on investment, and viable business use cases through the use of data statistics, dictionaries, dashboards, and reports. The Describe step utilizes case value determination that identifies key performance indicators / metrics that measure the success of the system. The Describe step further includes data understanding and sanity checks while utilizing the value chain to further evaluate the feasibility and risks of technology, data, and analytics and avoid siloed analysis.

[0027] The exploration component includes exploratory data analysis and data engineering through hypothesis building and development, data exploration (e.g., correlation analysis, outlier analysis, trend, centrality, seasonality analysis, etc.), building data visualizations, example analysis, data engineering (e.g., integration, normalization), and identifying variables that can be used as better predictors.

[0028] The predictive component is directed to building predictive models based on knowledge learned from the descriptive and exploratory components. The predictive component may include analysis using statistical learning and machine learning (codifying causal relationships), "goodness of fit" metrics, codifying patterns / relationships embedded in the data, data enrichment, feature engineering, asset and physical behavior modeling, and supervised / unsupervised / enhanced machine learning.

[0029] The prescriptive component is directed to generating prescriptive, actionable recommendations in response to predictions from the predictive component and may include root cause analysis, explainable artificial intelligence (AI), generating actionable recommendations, mining and incorporating asset and physical behavior models, measuring decision effectiveness, and performing constraint-based optimization.

[0030] The automation components focus on integrating with operating environments or systems and automating the application of prescription actions to operating environments or control systems. The automation components may include model deployment and execution, model lifecycle management, automating the application of prescription actions to operating environments or control systems, smart products or devices, intelligent applications, and decision effectiveness. Automation hardware components including programmable logic controllers (PLCs), robots, variable frequency drives (VFDs), smart products or devices, sensors, etc. may be programmed to support automation tasks such as automating the application of prescription actions. The automation components may drive several benefits including, but not limited to, increased productivity, throughput, and revenue, improved product quality, enhanced health and safety, reduced utility / energy costs, reduced downtime of industrial systems, and reduced time to diagnose system failures and reduced maintenance / repair costs.

[0031] The autonomous components focus on continuous learning based on new data and feedback and may include reinforcement learning, transfer learning, automatic learning (e.g., knowledge store, knowledge graph, user behavior, etc.), cohort learning across multiple subsystems (or domains), and automation of continuous learning based on new data / evidence and / or error / feedback. The automated and autonomous components are integrated together to support the automation of application of prescriptive actions, obtain new data, error, and feedback, retrain predictive and prescriptive models, and make intelligent decisions regarding prescriptive actions regarding controlling the system to work more efficiently and effectively by leveraging the ability of autonomous learning. This is in contrast to prior art automated systems where rule-based models or algorithms are designed and maintained by humans to support the automation of the system.

[0032] Because many tasks in the system are generic across different data science tasks, the example implementations described herein automate the generic components of each component using AutoML techniques to reduce development time and improve performance of the solution.

[0033] Figure 1(a) illustrates an example of an automated system for both offline and online processes according to an exemplary implementation. In the exemplary implementation of Figure 1(a), an exemplary subset of generic content that may be automated is shown.

[0034] 1(a), the offline system includes a description component 100, an exploration component 110, a prediction component 120, and a prescription component 130. The offline system operates on offline data or historical data. Outputs from the offline system include data analysis results from the description component 100 and exploration component 110, a prediction model from the prediction component 120, and a prescription model from the prescription component 130.

[0035] The descriptive component 100 is configured to provide a descriptive analysis 101 (e.g., data sanity check, statistics, etc.) as output when a problem and a dataset are provided as input. The exploration component 110 is configured to provide an exploratory analysis 111 (e.g., correlation, outliers, etc.) of a given dataset if the given dataset is of sufficient quality to perform the analysis. The predictive component 120 is configured to build a predictive model based on the exploratory analysis if the exploratory analysis is of sufficient quality to build a predictive model. If not, further descriptive or exploratory analysis may be performed until sufficient quality is achieved. The predictive component 120 may perform many activities for predictive modeling, including solution initialization 121, solution selection 122, hyperparameter optimization 123, and post-processing 124. The prescriptive component 130 is configured to build a prescriptive model 131 (e.g., root cause analysis, recommendations, explainable AI, etc.) based on the predictive model. If the predictive model is not of sufficient quality to build a prescriptive model, the predictive model may be reconstructed by the predictive component 120.

[0036] In one example of an online system, the online system is configured to take online or real-time data as input. Outputs from the online system include predictions, actionable recommendations, control signals to control or machine systems, and continuously trained models. In an exemplary implementation of the online system, the prediction component 120 is configured to deploy models from the offline system to the online system to perform predictions based on the input of real-time data. The prescription component 130 is configured to generate prescription actions from prescription models based on predictions provided by the prediction component 120.

[0037] The automation component 140 is configured to automatically apply prescribed actions to the control system and change the operation of the underlying machines / assets. The automation component 140 may include device / equipment / process / model integration and deployment 141, as well as DevOps improvements and automation of application of recommended actions 142.

[0038] The autonomous component 150 is configured to solicit error and / or feedback based on comparing predictions with associated prescriptive actions against observed incidents and action items. Using real-time data and the current version of the model, the information is utilized to retrain and improve the predictive and prescriptive models. The retrained model is evaluated against reserved test data. If the retrained model has better performance, the retrained model is deployed to the online system, otherwise the current version of the model remains. The autonomous component 150 may include self-learning (e.g., reinforcement learning) 151, continuous learning 152, transfer learning 153, and cohort learning 154.

[0039] Figure 1(b) shows an example flow of an offline process according to an example implementation. First, the description component 100 is called as shown in Figure 1(a). At 1000, a determination is made whether the data is sufficient to support a descriptive analysis. If it is sufficient (yes), the flow proceeds to the exploration component 110 as shown in Figure 1(a) to determine an exploratory analysis, and if it is not sufficient (no), the flow proceeds to 1010 to request more data from the system.

[0040] At 1001, a determination is made whether the descriptive and exploratory analysis are of sufficient quality to build a model. If so (yes), the flow proceeds to the predictive component 120 to generate a predictive model from the exploratory and predictive analysis. If not (no), the flow returns to 1010 to obtain more data to perform the analysis.

[0041] At 1002, a determination is made whether the predictive model generated by the prediction component 120 has sufficient accuracy for the historical data. If so (yes), the flow proceeds to the prescription component 130 to generate a prescription model. If not (no), the flow returns to the exploration component 110 to generate further exploratory analysis.

[0042] At 1003, a determination is made whether the prescription model generated by the prescription component 130 has sufficient accuracy. If so (yes), the flow ends and the online system is invoked when the predictive and prescription models are deployed online to the underlying system. If not (no), the flow returns to the prediction component 120 to regenerate the predictive model.

[0043] Figure 2 illustrates an example workflow of an online system when integrated with an offline system, according to an example implementation. As shown in Figure 2, the online system operates to apply predictive and prescriptive models to online data (or real-time data) to generate predictions and actionable recommendations, automatically apply the recommendations to the control system, and continuously self-learn based on new data, errors, and feedback. An example workflow is as follows:

[0044] Initially, the offline system 210 is configured to generate and provide predictive and prescriptive models for the online system, after which the predictive and prescriptive models are deployed to the online system at 211. At 212, the models are executed in the online system to perform predictions 221 on real-time data 220. Also, prescriptive actions 231 are generated from the prescriptive models based on the predictions 221 and automatically applied to the control system 230. The control system / DevOps 230 then responds to the prescriptive actions by changing the configuration of assets and how machines work, thereby affecting the real-time data collected from the sensors.

[0045] Errors are generated automatically by comparing predictions 221 with observed values ​​in 220 in real time. Feedback regarding predictive and prescriptive actions can also be generated automatically by analyzing operator logs. Errors, feedback, and new data are combined to retrain both predictive and prescriptive models through autonomous learning 200.

[0046] Depending on the desired implementation, several techniques for autonomous learning 200 may be used, including continuous learning, transfer learning, reinforcement learning, fine-tuning techniques, boosting, etc. Depending on the desired implementation, the continuous learning may be done in real-time or based on a predefined schedule. The retrained model is evaluated against reserved test data, and if the model has better performance, it is deployed to the online system at 211, otherwise the current version of the model remains in the online system.

[0047] In another exemplary implementation, the retrained model can be deployed in the online system in parallel with the existing model. Both the existing model and the retrained model are run on real-time data to generate predictions and prescription actions for a predefined period. However, during the predefined period, only the predictions and prescription actions generated from the existing model are enabled. After the predefined period ends, the existing model and the retrained model are evaluated and compared based on the predictions and prescription actions for the predefined period, and the model with better performance is left in the online system and enabled.

[0048] Throughout the exemplary implementation, the online system attempts to find a "best fit" solution in real time based on the real time conditions of the underlying industrial system. The metrics and objectives analyzed by the model itself can be configured according to the desired implementation based on the underlying industrial system.

[0049] As shown in Figure 2, once a model is constructed, it is deployed to a physical runtime environment. Access to a physical runtime environment (e.g., a customer environment) can be difficult and costly. To address such issues, an example implementation includes a package that contains a model that can be easily deployed to a physical runtime environment.

[0050] To achieve such an exemplary implementation, the model files are rendered in a standard format, and the package of model files is made portable to any physical runtime environment so that the model can be deployed and executed. To solve the first problem, the exemplary implementation uses a standard pickle file format for the model files and MLFlow to manage the lifecycle of the model. To solve the second problem, the exemplary implementation packages the MLFlow package (including the model files) into a Docker image, which is a virtual runtime environment that contains the MLFlow package and can run on top of the physical runtime environment facilities.

[0051] Machine learning lifecycle management is an efficient way of managing the lifecycle of a model, including model building, model deployment, model versioning, etc. In the exemplary implementation described herein, there is further emphasis on standardizing the format of the provided model so that it can be used in different execution environments. In the exemplary implementation described herein, one implementation is to use a pickle file format for the model file and package it into the MLFlow package, which is an existing model lifecycle management tool. Model versioning is based on model performance, and when a new model is built, the new model is evaluated against reserved test data based on metrics. If the evaluation results are good (and better than the existing model), the new model is kept, otherwise it is discarded. Another exemplary implementation of selecting a retrained model is described in

[0047] .

[0052] Autonomous learning is an artificial intelligence (AI) technique that encourages machines (or other learners) to learn and analyze by themselves, without human assistance or supervision, in a changing and uncertain world. With autonomous learning, the learner is autonomously responsible for its own learning process. Autonomous learning is also called self-directed learning.

[0053] In an example implementation, after a model is deployed in a dynamic and time-changing physical runtime environment, the model is automatically tuned and improved without human assistance or supervision based on new real-time data, errors, and feedback collected in the physical runtime environment, as shown in Figure 2. Several techniques may be used for this purpose, including reinforcement learning, continuous learning, transfer learning, cohort learning, etc.

[0054] Reinforcement learning is an area of ​​machine learning concerned with how a software agent should take actions in order to maximize some notion of cumulative reward through trial and error in an uncertain and potentially complex environment. Reinforcement learning starts with random trials and ends with complex tactics to achieve a goal.

[0055] In the exemplary implementation described herein, reinforcement learning starts with a deployed offline model (which is better than random trials), makes predictions, and generates prescription actions. The application of prescription actions changes the behavior of the physics runtime environment, including the collected data. New data, model errors, and feedback, as well as the problem targets, are collected and used to calculate the next prescription action. Several state-of-the-art techniques, including Deep Q Network (DQN), Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), Twin Delayed Deep Deterministic Policy Gradient (TD3), etc., may be used according to the desired implementation.

[0056] An example is provided to illustrate the various elements in reinforcement learning modeling. In one example, a model is constructed to optimize production from a plant. To build an offline model, key factors for the plant's production are determined and used as features, and the production is used as a target to build a supervised regression model. Once the offline model is built, it can be deployed to an online system, where reinforcement learning is used in addition to the offline model to optimize production, as follows:

[0057] "State": A state defines the observations that the learning agent receives from the environment. In this example, it is the current plant state that includes sensor readings, operating parameters, and current production.

[0058] "Policy": A policy defines how the learning agent behaves at a given time in the current state of the environment. In this example, the policy defines a mapping from the perceived current state of the plant to actions to be taken to optimize production.

[0059] Reward: The reward defines the goal of the reinforcement learning problem. In this example, the reward is the "yield of the plant." The reward indicates the outcome of an action in an immediate sense, where each action immediately leads to a reward.

[0060] "Value function": The value of a state is the total reward the learning agent can expect to get from that state. In this example, it is the total output of the plant from that state over a period of time, say a month. The value function defines the reward in the long run.

[0061] It should be noted that the machine learning techniques used in the offline and online models may differ: in this example, the offline model is based on supervised learning techniques, while the online model is based on reinforcement learning techniques.

[0062] Continuous learning is a machine learning technique that facilitates incremental and continuous improvement of an existing model for the same problem with newly collected data in the system. In an exemplary implementation, after a model is built from an offline system, the model is deployed to an online system. New real-time data collected in the online system is used to retrain the model and improve it to achieve better performance. The new model adapts to the new data to make itself the best-fit model. Retraining of the model can be done on a predefined schedule, which can be real-time, daily, weekly, etc., depending on the training needs and costs. In continuous learning, the machine learning techniques used in the offline and online models are the same.

[0063] Transfer learning is a machine learning technique that facilitates building a model for a problem by leveraging knowledge learned from related problems. Reasons for using transfer learning include: there is not enough past data for a new problem; the training cost is very high on data from the new problem; and the new problem is very similar to an existing problem, so the new problem can benefit from a model built for the existing problem. In the exemplary implementation described herein, after an offline model is built for a problem, the offline model is deployed to an online system for another problem. The model is retrained and improved for the new problem using real-time data from the online system. In this way, transfer learning can be an efficient and effective way to build a model for a new problem. In transfer learning, the machine learning techniques used in the offline model and the online model are the same.

[0064] FIG. 3 illustrates an example of automated adaptive cohort learning according to an exemplary implementation. An industrial system may include several similar components working simultaneously. If the entire system is too large and / or the online system does not have enough computing power to build a model, one design is to use automated adaptive cohort learning 300 to distribute the online model building work to each subsystem at 320, combine the models from the subsystems at 330, and distribute the combined model back to each subsystem at 331, as shown in FIG. 3. The subsystems may be identified based on the hierarchical design of the entire system, i.e., the asset model. The workflow of automated adaptive cohort learning 300 may include the following:

[0065] At 310, a global offline predictive model and a global offline prescriptive model are trained based on all historical data of the system. Assume that building an offline model is a one-time task, and the requirement of computing hardware resources for building the offline model can be met. At 320, the global predictive model and the global prescriptive model are deployed to each subsystem. Each subsystem is installed in the online system as shown in FIG. 2, and follows the workflow as shown in FIG. 2. In each subsystem of the local model 320, each pair of local models (predictive model and prescriptive model) is updated by autonomous learning, as shown in FIG. 2.

[0066] At some predefined schedule (e.g., daily), the local models in the subsystems are ensembled together to create a global model at 330. Depending on how the models are evaluated, there are two design schemes that can be adopted to ensemble the local models into a global model. One exemplary design scheme can include global evaluation. Each pair of local models (predictive and prescriptive models) is run for a predefined period (e.g., 3 hours) against new global data across the system. The global evaluation design scheme then measures the performance score of each pair of local models based on predefined evaluation metrics, which can be business metrics or the like according to the desired implementation. The output of the model evaluation is a vector, with each element being the performance score of a pair of local models (predictive and prescriptive models). Finally, the global evaluation design scheme finds the highest performance score and selects the corresponding model.

[0067] Another design scheme may include local evaluation. Each pair of local models (predictive and prescriptive models) is run for a predefined period (e.g., 3 hours) on new local data in each subsystem. The local evaluation design scheme then measures the performance score of each pair of local models based on predefined evaluation metrics, which may be business metrics or the like according to a desired implementation. The output of the model evaluation is a matrix, with each element being the performance score of the pair of local models (predictive and prescriptive models) on the local data in each subsystem. Finally, the local evaluation design scheme may check how many times each pair of local models was ranked highest based on the performance scores across all local data, and select the model with the most wins.

[0068] In another example design scheme for local evaluation, an exemplary implementation may calculate an average performance score across the local dataset for each pair of local models (predictive and prescriptive models) and select the model with the highest average performance score as the global model.

[0069] In another example design scheme for local evaluation, an exemplary implementation may convert the performance score matrix into a ranking matrix, calculate an average ranking across the local dataset for each pair of local models (predictive and prescriptive models), and select the model with the highest rank as the global model.

[0070] In another example design scheme for local evaluation, instead of selecting one pair of global models (predictive and prescriptive models), an example implementation may take the local model pair with the best performance score for each local dataset and select a set of model (predictive and prescriptive) pairs for all subsystems.

[0071] Finally, the selected model 311 is deployed to each subsystem of the local model 320. Through such an exemplary implementation, the online system can be automated and adaptive to real-time data.

[0072] The exemplary implementations described herein provide a comprehensive data science process that includes offline and online processes that work seamlessly together as one process. To support the development of an automated system for the data science process described herein, the exemplary implementations include a unified, highly customizable, and highly extensible framework that provides an easy-to-use interface integrated with various underlying machine learning libraries to meet the needs of data science professionals at various skill levels. The framework organizes each solution into two components as follows:

[0073] Solution Configuration File: The solution configuration file defines the actions or items at every step of the data science process. For each step shown in Figure 1(a), the solution configuration file has a section to specify the configuration and instructions on how to execute each step. Each section may also have subsections that define the details of the work that needs to be done.

[0074] Main Program: The main program is the entry point to each solution. It reads the solution configuration file, follows the definitions of each component, and executes the components accordingly.

[0075] A solution configuration file is characterized by the following main attributes:

[0076] Unified: Machine learning methods from multiple machine learning libraries can be specified in a single solution configuration file.

[0077] Extensible: The solution is easily extended with state-of-the-art techniques by adding sections to the solution configuration file.

[0078] Customizable: Optimized generic solution templates are provided for each solution type, such as classification, regression, recommendation engine, clustering, and anomaly detection. Additionally, users can easily customize the solution configuration files to suit their needs.

[0079] FIG. 4 illustrates an example of a solution configuration file according to an exemplary implementation. Data science processes, including offline and online processes as described herein and as shown in FIG. 1(a), are executed according to the solution configuration file 400 illustrated herein. Each solution has a solution configuration file, which is used to specify all the details of the solution. In the exemplary implementation described herein, the solution configuration file 400 has a section for each component illustrated in FIG. 1(a) that specifies the configuration and instructions on how to execute each component in the online and offline processes. The configuration may include a section for each component. If further specifications are required for each component, subsections are used to specify such details.

[0080] In the description component 100, the description section 401 may include a description of the data source and other metadata information related to the data source. Such metadata information may be utilized by the description component function 101, such as for sanity checks, identifying the amount of missing values, or information about basic statistics such as the mean, median, standard deviation, etc.

[0081] In the exploration component 110, the exploration section 402 may include data analysis 111 such as correlation analysis, outlier analysis, time series data analysis (trend, centrality, seasonality), etc.

[0082] In the prediction component 120, the functionality of the prediction component 120 is directed to defining the type of problem to be solved (e.g., classification problem, etc.). Processes initialized include solution initialization 121, solution selection 122, hyperparameter optimization 123, and model post-processing 124. The solution configuration file may include prediction 403, which may include solution initialization parameters indicating the location of a new dataset and a solution initialization model, solution selection parameters indicating the location of a new dataset and a solution selection model, hyperparameter optimization parameters indicating a list of subsections for optimizing candidate solutions, and model post-processing parameters. Examples of solution initialization, solution selection, hyperparameter optimization, and model post-processing that may utilize solution configuration file parameters are described in International Application PCT / US2020 / 039647, entitled “AUTOMATED MACHINE LEARNING: A UNIFIED, CUSTOMIZABLE, AND EXTENSIBLE SYSTEM,” the contents of which are incorporated by reference in their entirety for all purposes.

[0083] In the example solution configuration file of FIG. 4, prediction section 404 primarily specifies the configuration of hyperparameter optimization. Hyperparameter optimization may include information on what data is used, how features are processed, what models are used, model algorithms for each of the models, how the models are evaluated, model parameters, and other parameters, according to the desired implementation. Subsections of hyperparameter optimization may be data 403-1, features 403-2, models 403-3, feature specification 403-4, model specification 403-5, and model evaluation 403-6. Data 403-1 may indicate the data source or data file from which data is extracted and may include parameters such as labels, types of features involved, data splitting, etc., according to the desired implementation. Data 403-1 may also specify inputs and outputs. Features 403-2 indicate parameters related to feature extraction, such as enabling / disabling extraction of related features, feature selection based on variance thresholds, etc., according to the desired implementation. Model 403-3 indicates a model to be applied to data 403-1, such as, but not limited to, a logistic regression, a random forest classifier, a gradient boosting classifier, a support vector machine classifier, a neural network classifier, etc., depending on the desired implementation. Feature specification 403-4 may indicate parameters for the features 403-2 being incorporated. Model specification 403-5 indicates parameters for the model 403-3 being incorporated. Evaluation 403-6 indicates metrics used to evaluate model 403-3.

[0084] In prescription component 130 (and corresponding details 131), prescription section 404 may include techniques and associated parameters for calculating feature importance for the model, techniques and associated parameters for performing root cause analysis of the predictions, and techniques and associated parameters for optimizing the feature space to obtain the best target value, as shown at 131.

[0085] In the automation component 140, the automation section 405 can include methods and associated parameters for deploying offline models to online systems, as shown at 141, and automating the application of prescription actions to the online systems, as shown at 142. Other automation tasks include integration with online systems (devices, machines, etc.) through APIs and / or data interfaces.

[0086] In the autonomous component 150, the autonomous section 406 may include techniques and associated parameters for performing autonomous learning on real-time data, including one or more of the following components: reinforcement learning 151, continuous learning 152, transfer learning 153, and cohort learning 154. The autonomous component must also run within an online system and configure the lifecycle management of the model.

[0087] For some components (e.g., model 403-3), the solution configuration file specifications may include which model algorithm in the underlying machine learning library needs to be invoked, which parameters need to be passed to the model algorithm, etc. The solution configuration file 400 is used to define each problem to be taken into account in each solution. For example, in a machine learning scenario, there may be classification and regression problems, whereby the problem is specified by the entire solution including how the data is pre-processed, how feature engineering is performed, what model is selected, parameter tuning, post-processing, evaluation, etc.

[0088] The component shown in FIG. 1(a) parses a solution configuration file and obtains machine learning libraries to execute according to the file. An exemplary implementation utilizes a solution configuration file to define and specify a solution, and uses reflection techniques provided in many programming languages ​​(e.g., Python, Java, etc.) to read and parse the solution configuration file. Through such an exemplary implementation, the outcome is that various analysis libraries and machine learning libraries can be used together in one solution.

[0089] In an exemplary implementation, the main program is the entry point to each solution and is configured to read the solution configuration file and execute each module accordingly, following the definitions of each module.

[0090] The solution configuration file is characterized by the following key attributes, which are described in detail herein after describing the contents of the solution configuration file: The solution configuration file is unified such that machine learning methods such as feature engineering and model algorithms from multiple machine learning libraries can be specified in a single solution configuration file. The solution configuration file is extensible such that solutions can be easily extended with state-of-the-art techniques by adding sections to the solution configuration file. The solution configuration file is customizable. Optimized generic solution templates are provided for each solution type, including but not limited to classification, regression, clustering, anomaly detection, recommendation engine, remaining useful life prediction. For the purpose of hyperparameter optimization, the templates specify all possible feature engineering techniques and associated parameter grids, model algorithms and associated parameter grids, and evaluation techniques. Users can use the configuration specified in the solution template or modify the configuration and customize it to fit their needs.

[0091] FIG. 5(a) shows an example flow diagram for automated and autonomous self-learning machine learning model generation and deployment according to an example implementation. In an example implementation, the flow generates and executes the description component 100, exploration component 110, prediction component 120, and prescription component 130 from the solution configuration file shown in FIG. 4 in 501 to generate predictive and prescription models in a first offline system. Then, in 502, the flow deploys the predictive and prescription models to a second online system to control the operation of the second system from prescription actions generated by the prescription model in response to real-time data provided from the second system. In 503, the flow autonomously updates the predictive and prescription models based on real-time data, errors, and feedback from the second system, as shown in FIG. 2 and FIG. 5(b).

[0092] FIG. 5(b) illustrates an example flow of an online process configured to autonomously update the predictive and prescriptive models from real-time data, error, and feedback from the second system while the models are deployed, according to an example implementation. At 510, the flow receives feedback from the second system. At 511, the flow determines the error between the prediction from the predictive model and the feedback from the second system. At 512, the flow retrains the predictive and prescriptive models based on the new data, error, and feedback, as shown in FIG. 3. At 513, a determination is made whether the retrained predictive and prescriptive models have higher accuracy than the currently deployed predictive and prescriptive models based on using reserved test data for the retrained models compared to the current models. If they do (if yes), the flow proceeds to 514 to replace the deployed models with the retrained models, and if they do not have higher accuracy (if no), the flow ends. For another example implementation of comparing an existing model with a retrained model, see the example implementation including how the retrained model can be deployed in an online system in parallel with an existing model.

[0093] Through the flow as shown in Figures 5(a) and 5(b), the offline training system for generating the model can be integrated with the online system that deploys the model, thereby enabling seamless update of the model for controlling the online system without requiring the online system to stop operation. After the offline model is deployed to the online system, it is improved in parallel based on historical data, and then, if the performance improvement from the offline model (compared to the current version of the online model) meets a predefined criterion, the online model can be seamlessly replaced without interrupting the normal operation of the online system.

[0094] 6 illustrates a plurality of systems and a management device according to an exemplary implementation. One or more second systems 601-1, 601-2, 601-3, and 601-4 are communicatively coupled to a network 600 connected to a management device 602 that receives models and deploys them online to their underlying devices, and controls offline models and assists the first system in processing and deploying the offline models to one or more second systems 601-1, 601-2, 601-3, and 601-4. The management device 602 manages a database 603 that includes aggregated data feedback from the systems in the network 600. In an alternative exemplary implementation, data feedback from the second systems 601-1, 601-2, 601-3, and 601-4 may be aggregated in a central repository or database, such as a proprietary database that aggregates data from systems such as enterprise resource planning systems, and the management device 602 may access or retrieve data from the central repository or database. Such systems include equipment that will be operated by an operator for manufacturing or other processes, such as lathes, presses, mills, grinders, as well as mobile equipment such as forklifts, cranes, trucks, and any other equipment that will be operated by an operator for manufacturing, machining, or other process purposes. To deploy the predictive and prescriptive models online, each of the one or more second systems 601-1, 601-2, 601-3, and 601-4 may comprise one or more programmable logic controllers (PLCs) that directly control the operation of the underlying devices, or an Internet of Things (IoT) gateway configured to issue instructions to the underlying devices that control their operation. The PLCs or IoT gateways may be configured to receive and execute the models from the management device 602, deploy the models to the underlying systems, and provide feedback to the management device 602.The management device 602 may be operated separately from the second systems 601-1, 601-2, 601-3, and 601-4, so as to operate offline and independently from the systems. The proposed automated system may be implemented as follows.

[0095] In an exemplary use case of an exemplary implementation, the system may be provided as a standalone system that includes a machine learning library having software components for automatically performing various tasks of a data science process. In such an implementation, a user writes code that calls functions in the library and manages the computation and data storage.

[0096] In an exemplary use case of an exemplary implementation, the system may be implemented as on-premise software with a graphical user interface (GUI) that presents functions provided in a machine learning library, where a user executes automated data science processes and manages computation and data storage.

[0097] In an exemplary use case of an exemplary implementation, a Software as a Service (SaaS) based solution may be utilized through a web-based user interface that presents machine learning libraries. Users subscribe to the service, upload data to the service provider, and run automated data science processes using the web-based user interface. The service provider manages the computation and data storage.

[0098] In an example use case of an example implementation, the system may be embedded as a microservice or microapplication to present a machine learning library.

[0099] Figure 7 illustrates an exemplary computing environment including an exemplary computer device suitable for use in some exemplary embodiments, such as the management device 602 as shown in Figure 6, or the PLC / IoT gateways of the second systems 601-1, 601-2, 601-3, 601-4 as shown in Figure 6. The functionality described herein may be implemented in the management device 602 or supported throughout the system based on some combination of elements, depending on the desired implementation.

[0100] A computing device 705 in computing environment 700 may include one or more processing units, cores, or processors 710, memory 715 (e.g., RAM, ROM, and / or the like), internal storage 720 (e.g., magnetic storage, optical storage, solid-state storage, and / or organic storage), and / or I / O interface 725, any of which may be coupled over a communication mechanism or bus 730 for communicating information or may be embedded within computing device 705. I / O interface 725 is also configured to receive images from a camera or provide images to a projector or display, depending on the desired implementation.

[0101] The computing device 705 may be communicatively coupled to an input / user interface 735 and an output device / interface 740. Either one or both of the input / user interface 735 and the output device / interface 740 may be a wired or wireless interface and may be removable. The input / user interface 735 may include any device, component, sensor, or interface, physical or virtual, that may be used to provide input (e.g., buttons, touch screen interfaces, keyboards, pointing / cursor controls, microphones, cameras, Braille, motion sensors, optical readers, and / or the like). The output device / interface 740 may include displays, televisions, monitors, printers, speakers, Braille, and the like. In some example implementations, the input / user interface 735 and the output device / interface 740 may be embedded in the computing device 705 or may be physically coupled to the computing device 705. In other example implementations, other computing devices may function as or provide the functionality of the input / user interface 735 and the output device / interface 740 for the computing device 705 .

[0102] Examples of computing devices 705 include, but are not limited to, highly mobile devices (e.g., smart phones, devices in vehicles and other machines, devices carried by humans and animals, etc.), mobile devices (e.g., tablets, notebooks, laptops, personal computers, portable televisions, radios, etc.), and devices not designed for mobility (e.g., desktop computers, other computers, information kiosks, televisions with one or more processors embedded therein / coupled thereto, radios, etc.).

[0103] Computing device 705 may be communicatively coupled (e.g., via I / O interface 725) to external storage 745 and a network 750 for communicating with any number of networked components, devices, and systems, including one or more computing devices of the same or different configurations. Computing device 705, or any connected computing device, may function as, provide services to, or be referred to as, a server, a client, a thin server, a general purpose machine, a special purpose machine, etc.

[0104] I / O interface 725 may include, but is not limited to, wired and / or wireless interfaces using any communications or I / O protocol or standard (e.g., Ethernet, 802.11x, Universal System Bus, WiMax, modem, cellular network protocols, etc.) for communicating information to or from at least all connected components, devices, and networks in computing environment 700. Network 750 may be any network or combination of networks (e.g., the Internet, a local area network, a wide area network, a telephone network, a cellular network, a satellite network, etc.).

[0105] The computing device 705 may use and / or communicate using computer usable or computer readable media, including transitory and non-transitory media. Transitory media include transmission media (e.g., metallic cables, optical fibers), signals, carrier waves, etc. Non-transitory media include magnetic media (e.g., disks and tapes), optical media (e.g., CD-ROMs, digital video disks, Blu-ray disks), solid-state media (e.g., RAM, ROM, flash memory, solid-state storage), and other non-volatile storage or memory.

[0106] The computing device 705 may be used to implement techniques, methods, applications, processes, or computer-executable instructions in some exemplary computing environments. The computer-executable instructions may be obtained from a transitory medium, may be stored on a non-transitory medium, or may be obtained from a non-transitory medium. The executable instructions may be from one or more of any programming, scripting, and machine language (e.g., C, C++, C#, Java, Visual Basic, Python, Perl, JavaScript, etc.).

[0107] The processor 710 can run under any operating system (OS) (not shown) in a native or virtual environment. One or more applications can be deployed, including a logic unit 760, an application programming interface (API) unit 765, an input unit 770, an output unit 775, and inter-unit communication mechanisms 795 for different units to communicate with each other, with the OS, and with other applications (not shown). The described units and elements may differ in design, function, configuration, or implementation and are not limited to the provided description.

[0108] In some example implementations, when information or instructions to execute are received by the API unit 765, the information or instructions to execute may be communicated to one or more other units (e.g., logic unit 760, input unit 770, output unit 775). In some examples, logic unit 760 may be configured to control information flow between units and direct services provided by API unit 765, input unit 770, output unit 775 in some example implementations described above. For example, the flow of one or more processes or implementations may be controlled by logic unit 760 alone or in cooperation with API unit 765. Input unit 770 may be configured to obtain inputs for the calculations described in the example implementations, and output unit 775 may be configured to provide outputs based on the calculations described in the example implementations.

[0109] Processor 710 may be configured to generate predictive and prescriptive models through an offline learning process, control the operation of the second system through deploying the predictive and prescriptive models to the second system, and autonomously update the prescriptive and predictive models from feedback from the second system through an online learning process once the prescriptive and predictive models are deployed to the second system, as shown in Figures 1(a), 5(a), and 5(b).

[0110] The processor 710 is configured to control the operation of the second system through deploying the predictive model and the prescriptive model to the second system and controlling the operation of the second system by performing automatic application of prescriptive actions generated from the prescriptive model that modify the operation of the second system, as shown in Figures 1(a) to 3.

[0111] The processor 710 may be configured to generate predictive and prescriptive models through an offline learning process in the first system by generating, from the solution configuration file, a descriptive component configured to perform a descriptive analysis; generating, from the solution configuration file and the descriptive analysis, a discovery component configured to perform an exploratory analysis; generating, from the solution configuration file and the exploratory analysis, a predictive component configured to incorporate one or more machine learning libraries specified in the solution configuration file to generate a predictive model; and generating, from the solution configuration file and the predictive model, a prescriptive component configured to map prescriptive actions to results from the predictive model to generate a prescriptive model, as shown in Figures 1(a) to 4.

[0112] Processor 710 may be configured to control operation of the second system through deploying the predictive model and the prescriptive model to the second system, as shown in FIG. 1(a), by deploying the predictive model and the prescriptive model to be online and configured to incorporate real-time data from the second system, generating predictions from the real-time data through the predictive model, generating prescriptive actions from the predictions through the prescriptive model, and controlling operation of the second system according to the prescriptive actions.

[0113] The processor 710 may be configured to determine an error based on a difference between the feedback from the second system and a prediction from the predictive model associated with controlling the operation, as shown in FIG. 5(b); the processor 710 for the second system may be configured to retrain the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through a continuous learning process with the predictive model and the prescriptive model deployed to the second system; and autonomously update the predictive model and the prescriptive model from the feedback from the second system through an online learning process with the predictive model and the prescriptive model deployed to the second system by deploying the retrained prescriptive model and the retrained predictive model to the second system for the retrained predictive model and the retrained prescriptive model that have better performance than the predictive model and the prescriptive model.

[0114] The processor 710 may be configured to retrain the predictive and prescriptive models in the second system based on real-time data, errors, and feedback through the continuous learning process with the predictive and prescriptive models deployed in the second system by distributing a machine learning process for generating a retrained predictive model and a retrained prescriptive model to a plurality of local models associated with the subsystems of the second system, iteratively ensembling the plurality of local models to generate a plurality of retrained predictive models and a plurality of retrained prescriptive models, selecting one of the generated plurality of retrained predictive models and one of the generated plurality of retrained prescriptive models for distribution back to the machine learning process, and ensembling the plurality of local models to generate a retrained predictive model and a retrained prescriptive model, as shown in FIG.

[0115] The processor 710 may be configured to retrain the predictive model and the prescriptive model based on the real-time data, error, and feedback through a continuous learning process in the second system with the predictive model and the prescriptive model deployed to the second system by retraining the predictive model and the prescriptive model based on the real-time data, error, and feedback through one or more of a reinforcement learning process or a transfer learning process, as shown in FIG.

[0116] Some portions of the detailed description are presented in terms of algorithms and symbolic representations of operations within a computer. These algorithmic descriptions and symbolic representations are the means used by those skilled in the data processing arts to convey the substance of their innovations to others skilled in the art. An algorithm is a sequence of defined steps leading to a desired end state or result. In the exemplary implementation, each step performed requires physical manipulations of tangible quantities to achieve a tangible result.

[0117] Unless otherwise specifically stated, as is apparent from the above discussion, throughout the description, discussion utilizing terms such as "processing," "computing," "calculating," "determining," "displaying," and the like may include the actions and processes of a computer system or other information processing device, which will be understood to manipulate data represented as physical (electronic) quantities in the computer system's registers and memory, and to convert it into other data similarly represented as physical quantities in the computer system's memory or registers or other information storage, transmission, or display device.

[0118] Exemplary implementations may also relate to an apparatus for performing the operations herein. This apparatus may be specially constructed for the required purposes or may comprise one or more general-purpose computers selectively activated or reconfigured by one or more computer programs. Such computer programs may be stored in a computer-readable medium, such as a computer-readable storage medium or a computer-readable signal medium. Computer-readable storage media may include tangible media, such as, but not limited to, optical disks, magnetic disks, read-only memory, random access memory, solid-state devices and solid-state drives, or any other type of tangible or non-transitory medium suitable for storing electronic information. Computer-readable signal media may include media such as carrier waves. The algorithms and displays presented herein are not inherently related to any particular computer or other apparatus. A computer program may include a pure software implementation that includes instructions for performing the operations of a desired implementation.

[0119] Various general-purpose systems may be used with the programs and components according to the examples herein, or it may prove convenient to construct more specialized apparatus to perform the desired method steps. Additionally, the exemplary implementations are not described with reference to any particular programming language. It should be understood that a variety of programming languages ​​may be used to implement the teachings of the exemplary implementations described herein. Instructions in the programming language may be executed by one or more processing devices, such as a central processing unit (CPU), processor, or controller.

[0120] As known in the art, the operations described above may be performed by hardware, software, or some combination of software and hardware. Various aspects of the exemplary implementations may be implemented using circuits and logic devices (hardware), and other aspects may be implemented using instructions stored on a machine-readable medium (software) that, when executed by a processor, cause the processor to perform a method for performing an implementation of the present application. Furthermore, some exemplary implementations of the present application may be performed exclusively in hardware, and other exemplary implementations may be performed exclusively in software. Furthermore, the various functions described may be performed in a single unit or distributed across several components in any number of ways. When performed by software, the method may be executed by a processor, such as a general-purpose computer, based on instructions stored on a computer-readable medium. If necessary, the instructions may be stored on the medium in compressed and / or encrypted form.

[0121] Moreover, other implementations will occur to those skilled in the art upon consideration of this specification and practice of the teachings disclosed herein. Various aspects and / or components of the described exemplary implementations may be used alone or in any combination. It is intended that the specification and exemplary implementations be considered as examples only, with a true scope and spirit of the present application being indicated by the following claims.

Claims

1. 1. A method for automated, autonomous, self-learning machine learning model generation and deployment, the method comprising: generating predictive and prescriptive models through an offline learning process in a first system; Controlling an operation of a second system through deploying the predictive model and the prescription model to the second system; autonomously updating the prescriptive model and the predictive model from feedback from the second system through an online learning process while the prescriptive model and the predictive model are deployed to the second system; Including, generating the predictive model and the prescriptive model through the offline learning process in the first system; generating a descriptive component from the solution configuration file, the descriptive component being configured to perform a descriptive analysis; generating, from the solution configuration file and the descriptive analysis, an exploration component configured to perform an exploratory analysis; generating, from the solution configuration file and the exploratory analysis, a predictive component configured to incorporate one or more machine learning libraries specified in the solution configuration file to generate the predictive model; generating, from the solution configuration file and the predictive model, a prescription component configured to map prescription actions to outcomes from the predictive model to generate a prescription model; A method comprising:

2. 2. The method of claim 1, wherein controlling the operation of the second system through deploying the predictive model and the prescriptive model to the second system includes performing automated application of prescriptive actions generated from the prescriptive model to modify the operation of the second system.

3. controlling an operation of the second system through deploying the predictive model and the prescriptive model to the second system; deploying the predictive model and the prescriptive model to be online and configured to ingest real-time data from the second system; generating a prediction from the real-time data through the predictive model; generating a prescription action from the prediction through the prescription model; controlling the operation of the second system in accordance with the prescribed action; The method of claim 1 , comprising:

4. and autonomously updating the predictive model and the prescriptive model from the feedback from the second system through the online learning process, with the predictive model and the prescriptive model deployed to the second system. determining an error based on a difference between the feedback from the second system and a prediction from the predictive model associated with the controlling of the operation; retraining, in the second system, the predictive model and the prescriptive model based on real-time data, the error, and the feedback through a continuous learning process with the predictive model and the prescriptive model deployed in the second system; For the retrained predictive model and the retrained prescription model that have better performance than the predictive model and the prescription model, deploying the retrained prescription model and the retrained predictive model to the second system; The method of claim 1 , comprising:

5. retraining, in the second system, the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through the continuous learning process, with the predictive model and the prescriptive model deployed in the second system; distributing a machine learning process to generate a retrained predictive model and a retrained prescriptive model to a plurality of local models associated with subsystems of the second system; iteratively ensembling the local models to generate a plurality of retrained predictive models and a plurality of retrained prescriptive models, and selecting one of the generated retrained predictive models and one of the generated retrained prescriptive models for distribution back to the machine learning process; ensembling the local models to generate the retrained predictive model and the retrained prescriptive model; The method of claim 4 , comprising:

6. retraining, in the second system, the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through the continuous learning process, with the predictive model and the prescriptive model deployed in the second system; retraining the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through one or more of a reinforcement learning process or a transfer learning process. The method of claim 4 , comprising:

7. 1. A system for automated, autonomous, self-learning machine learning model generation and deployment, the system comprising:

1. A first system, comprising: Generate predictive and prescriptive models through an offline learning process; Controlling operation of a second system through deploying the predictive model and the prescription model to the second system; With the prescriptive model and the predictive model deployed to the second system, autonomously updating the prescriptive model and the predictive model from feedback from the second system through an online learning process. A first system including a processor configured to Equipped with The processor, generating a descriptive component from the solution configuration file, the descriptive component being configured to perform a descriptive analysis; generating, from the solution configuration file and the descriptive analysis, an exploration component configured to perform an exploratory analysis; generating, from the solution configuration file and the exploratory analysis, a predictive component configured to incorporate one or more machine learning libraries specified in the solution configuration file to generate the predictive model; generating, from the solution configuration file and the predictive model, a prescription component configured to map prescription actions to outcomes from the predictive model to generate a prescription model; generating the predictive model and the prescriptive model through the offline learning process in the first system by

8. 8. The system of claim 7, wherein the processor is configured to control operation of the second system through deploying the predictive model and the prescriptive model to the second system by performing automated application of prescriptive actions generated from the prescriptive model that modify operation of the second system.

9. The processor, deploying the predictive model and the prescriptive model to be online and configured to ingest real-time data from the second system; generating a prediction from the real-time data through the predictive model; generating a prescription action from the prediction through the prescription model; controlling the operation of the second system in accordance with the prescribed action; 8. The system of claim 7, configured to control an operation of the second system through deploying the predictive model and the prescriptive model to the second system by:

10. The processor, determining an error based on a difference between the feedback from the second system and a prediction from the predictive model associated with the controlling of the operation; the second system is configured to retrain the predictive model and the prescriptive model based on real-time data, the error, and the feedback through a continuous learning process with the predictive model and the prescriptive model deployed to the second system; For the retrained predictive model and the retrained prescription model that have better performance than the predictive model and the prescription model, deploying the retrained prescription model and the retrained predictive model to the second system; 8. The system of claim 7, configured to autonomously update the predictive model and the prescriptive model from the feedback from the second system through the online learning process once the predictive model and the prescriptive model are deployed to the second system.

11. The second system comprises: distributing a machine learning process to generate a retrained predictive model and a retrained prescriptive model to a plurality of local models associated with subsystems of the second system; iteratively ensembling the local models to generate a plurality of retrained predictive models and a plurality of retrained prescriptive models, and selecting one of the generated retrained predictive models and one of the generated retrained prescriptive models for distribution back to the machine learning process; ensembling the local models to generate the retrained predictive model and the retrained prescriptive model; 11. The system of claim 10, configured to retrain the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through the continuous learning process with the predictive model and the prescriptive model deployed to the second system.

12. The second system comprises: retraining the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through one or more of a reinforcement learning process or a transfer learning process.

12. The system of claim 11, configured to retrain the predictive model and the prescriptive model based on the real-time data, the error, and the feedback through the continuous learning process with the predictive model and the prescriptive model deployed to the second system.

13. 1. A non-transitory computer-readable medium storing instructions for automated, autonomous, self-learning machine learning model generation and deployment, the method being performed by a computer executing the instructions, the method comprising: generating predictive and prescriptive models through an offline learning process in a first system; Controlling an operation of a second system through deploying the predictive model and the prescription model to the second system; autonomously updating the prescriptive model and the predictive model from feedback from the second system through an online learning process while the prescriptive model and the predictive model are deployed to the second system; Including, generating the predictive model and the prescriptive model through the offline learning process in the first system; generating a descriptive component from the solution configuration file, the descriptive component being configured to perform a descriptive analysis; generating, from the solution configuration file and the descriptive analysis, an exploration component configured to perform an exploratory analysis; generating, from the solution configuration file and the exploratory analysis, a predictive component configured to incorporate one or more machine learning libraries specified in the solution configuration file to generate the predictive model; generating, from the solution configuration file and the predictive model, a prescription component configured to map prescription actions to outcomes from the predictive model to generate a prescription model; A non-transitory computer readable medium comprising:

Citation Information

Patent Citations

  • Stormwater drainage controlling device

    JP2004353220A

  • Distributed machine learning system, apparatus, and method

    JP2019526851A

  • Product state estimation device

    JP2020075385A

  • Machine learning based analytics platform

    US20180012145A1

  • Systems and methods for automating data science machine learning analytical workflows

    US20180165604A1