The present technology addresses the problem of quickly and safely improving policies in online reinforcement learning domains. As its solution, an exploration strategy comprising diverse exploration (DE) is employed, which learns and deploys a diverse set of safe policies to explore the environment. DE theory explains why diversity in behavior policies enables effective exploration without sacrificing exploitation. An empirical study shows that an online policy improvement algorithm framework implementing the DE strategy can achieve both fast policy improvement and safe online performance.

Direct measurement and simulation of real-time production rates of chemical products in complex chemical plants is complex. A predictive model developed based on machine learning algorithms using historical sensor data and production data provides accurate real-time prediction of production rates of chemical products in chemical plants. An optimization model based on machine learning algorithms using clustered historical sensor data and production data provides optimal values for controllable parameters for production maximization.

The present disclosure discloses a control method of industrial Internet of Things (IoT) based on abnormal identification. The IoT includes: an obtaining unit, which is configured to obtain a first machining parameter; a detection unit, which is configured to obtain real-time image data when the first machining parameter is abnormal; an extraction unit, which is configured to obtain a keyframe and obtain a second machining parameter; a judgment unit, which is configured to determine an abnormal cause based on the first machining parameter and the second machining parameter; and a communication unit, which is configured to transmit the abnormal cause to a user terminal through a service platform.

Systems and methods manage electrical loads in a grid by applying Robust principal component analysis (R-PCA) to decompose annual load profiles into low-rank components and sparse components; extracting one or more predetermined features; constructing a similarity graph; selecting submodular cluster centers through the constructed similarity graph; determining a cluster assignment based on selected centers; and applying the clustering assignment for load analysis.

Apparatuses, systems, and methods of physical-model based building automation using in-situ regression to optimize control systems are presented. A simulation engine is configured to simulate a behavior or a controlled system using a physical model for the controlled system. A data stream comprises data from a controlled system. A training loop is configured to compare an output of a simulation engine to a data stream using a heuristic so that a physical model is regressed in a manner that the output of the simulation engine approaches the data stream.

Devices, methods, and systems for data-driven acceleration of deployed services are described herein. One system includes a database configured to store a plurality of case pairs that correspond to previously calculated models, wherein the previously calculated models are based on a number of features, and wherein each of the plurality of case pairs comprise a first value representative of the number of features and a second value representative of deployed heating, ventilation, and air conditioning (HVAC) resources, a ranking engine configured to rank each of the plurality of case pairs based on a performance of the deployed services, and a deployment engine configured to: receive actual feature values, and deploy HVAC resources based on a comparison between the actual feature values and the plurality of ranked case pairs.

Disclosed is a model predictive control. The model predictive control method includes: receiving an input system dynamics model for a character; performing numerical differentiation to the input system dynamics model by using at least one technique selected from the group consisting of a differentiation information interpolation technique, a physical quantity reuse technique and a contact-space inverse mass matrix (CIMM) technique; and outputting differentiation information for the system dynamics model.

Techniques are described for implementing automated control systems to control operations of specified physical target systems. In some situations, the described techniques include forecasting future values of parameters that affect operation of a target system, and using the forecasted future values as part of determining current automated control actions to take for the target system—in this manner, the current automated control actions may be improved relative to other possible actions that do not reflect such forecasted future values. Various automated operations may also be performed to improve the forecasting in at least some situations, such as by combining the use of multiple different types of forecasting models and multiple different groups of past data to use for training the models, and/or by improving the estimated internal non-observable state information reflected in at least some of the models.

A method for controlling an operation of a machine determines a feasible region for states of the machine and states of the reference trajectory defined by constraints of the machine, constraints on a reference trajectory and constraints on bounds of a tracking error and selects a subset of the feasible region, such that for any state of the machine and any state of the reference trajectory within the subset, there is an admissible control maintaining the state of the machine within the subset for all admissible future states of the reference trajectory determined by the model and the constraints of the reference trajectory. Next, an admissible control action for controlling the operation of the machine is selected such that the state of the machine remains in the subset for all admissible future states of the reference trajectory.

An apparatus, method and computer program is described comprising detecting a first object in a first image of a sequence of images using a neural network (22), wherein the means for detecting the first object provides an object area indicative of a first location of the first object; and tracking the first object (24), wherein the means for tracking the first object further comprises generating a predicted future location of the first object and generating an updated location of the first object using the neural network. The means for generating the predicted future location of the first object may, for example, receive said object area indicative of a first location of the first object and may receive said updated location information of the first object.