Abstract: K-nearest multi-agent reinforcement learning for collaborative tasks with variable numbers of agents. Centralized reinforcement learning is challenged by variable numbers of agents, whereas decentralized reinforcement learning is challenged by dependencies among agents' actions. An algorithm is disclosed that can address both of these challenges, among others, by grouping agents with their k-nearest agents during training and operation of a policy network. The observations of all k+1 agents in each group are used as the input to the policy network to determine the next action tor each of the k+1 agents in the group. When an agent belongs to more than one group, such that multiple actions are determined for the agent, an aggregation strategy can be used to determine the final action for that agent.

Abstract: Example implementations described herein are directed to systems and methods for predicting if a conjunction of multiple events will occur within a certain time. It relies on an approximate decomposition into subproblems and a search among the possible decompositions and hyperparameters for the best model. When the conjunction is rare, the method mitigates the problem of data imbalance by estimating events that are less rare.

Abstract: In some examples, a system may generate a plurality of care path patient profile models based on a plurality of care path patterns for a plurality of past patient admissions. For example, each care path patient profile model may include a trained classifier. Further, the system may receive information related to a new patient admission, and may generate features from the received information related to the new patient admission. The system may input the features generated from the received information related to the new patient admission into the plurality of care path patient profile models to obtain a respective probability of being classified in a respective care path based on an amount of similarity to the patients who have gone through each care path. In addition, the system may present, on a display, information related to at least one care path pattern in a graphical user interface.

Abstract: Example implementations described herein involve systems and methods for calculating the importance of each iteration and of each input feature for multi-label models that optimize a multi-label objective function in an iterative manner. The example implementations are based on the incremental improvement in the objective function rather than an application-specific metric such as accuracy.

Abstract: In some examples, a computing device may receive hierarchical data having a hierarchical structure including a plurality of levels. The computing device may determine a plurality of features based at least in part on the hierarchical data, and may select a subset of the features at a first level as candidates for consolidating to a next higher level in the hierarchical structure. The computing device may determine that a predicted loss of information from consolidating the subset of features is less than a threshold, and may revise the hierarchical structure to include a consolidated feature at the next higher level, rather than the subset of features. In some examples, a statistical model may be trained based on the revised hierarchical structure and used at least partially to make a determination, send a notification, and/or control a device.

Abstract: In some examples, a system may generate a plurality of care path patient profile models based on a plurality of care path patterns for a plurality of past patient admissions. For example, each care path patient profile model may include a trained classifier. Further, the system may receive information related to a new patient admission, and may generate features from the received information related to the new patient admission. The system may input the features generated from the received information related to the new patient admission into the plurality of care path patient profile models to obtain a respective probability of being classified in a respective care path based on an amount of similarity to the patients who have gone through each care path. In addition, the system may present, on a display, information related to at least one care path pattern in a graphical user interface.

Abstract: In some examples, a system may train a false positive reduction machine learning model (MLM) for nodule detection. The system may receive training data images including negative images and positive images, along with an indication of nodule locations in the positive images. The system may determine elliptical approximations for nodules in the positive images, and may determine respective binarized contours from the elliptical approximations. Further, the system may determine an elliptical approximation space for the binarized contours, and may determine a subspace angle between individual image samples in the positive images and the elliptical approximation space as at least one feature of the MLM. Subsequently, when applying the MLM during nodule detection, one or more images may be input to the MLM to determine whether an indication of a nodule is correct, and if so, a visualization of a location of the nodule may be provided.

Abstract: In some examples, a computing device may receive sensor data associated with a plurality of patient devices that are associated with a plurality of patients. The computing device may further receive caregiver records corresponding at least partially to the sensor data. At least two groups of indicators may be determined from the caregiver records, such a based on a selected subject. Further, the computing device may determine a plurality of clusters from the sensor data. Based on the plurality of clusters and the at least two groups, the computing device may determine an indication of a discrepancy in care for a patient of the plurality of patients. Based on the indication of the discrepancy, the computing device may send at least one of a notification to a caregiver computing device, a notification to a monitoring computing device, or a control signal to one of the patient devices.

Abstract: In some examples, a computing device may receive sensor data for a target and at least one of: log data for the target, or historical log data and historical sensor data for a plurality of other targets. The computing device may determine at least one event classified as a non-uniform event in at least one of the log data or the historical log data, and may determine combined features, such as a feature vector, based on the sensor data and the non-uniform event(s). The computing device may determine an analysis result from the combined features. Further, based on the analysis result, the computing device may send a control signal to a device associated with the target for controlling the device, and/or may send a communication related to the target to another computing device.

Abstract: In some examples, a system may train a false positive reduction machine learning model (MLM) for nodule detection. The system may receive training data images including negative images and positive images, along with an indication of nodule locations in the positive images. The system may determine elliptical approximations for nodules in the positive images, and may determine respective binarized contours from the elliptical approximations. Further, the system may determine an elliptical approximation space for the binarized contours, and may determine a subspace angle between individual image samples in the positive images and the elliptical approximation space as at least one feature of the MLM. Subsequently, when applying the MLM during nodule detection, one or more images may be input to the MLM to determine whether an indication of a nodule is correct, and if so, a visualization of a location of the nodule may be provided.

Abstract: Example implementations described herein are directed to a decision-support system for maintenance recommendation that uses analytics technology to evaluate the effectiveness of a maintenance action or a group of actions in improving the performance of equipment and its components, and provide recommendations on which maintenance actions or a group of maintenance actions should be pursued and which should be avoided.

Abstract: In some examples, a computing device may receive sensor data for a target and at least one of: log data for the target, or historical log data and historical sensor data for a plurality of other targets. The computing device may determine at least one event classified as a non-uniform event in at least one of the log data or the historical log data, and may determine combined features, such as a feature vector, based on the sensor data and the non-uniform event(s). The computing device may determine an analysis result from the combined features. Further, based on the analysis result, the computing device may send a control signal to a device associated with the target for controlling the device, and/or may send a communication related to the target to another computing device.

Abstract: In some examples, a computing device may receive hierarchical data having a hierarchical structure including a plurality of levels. The computing device may determine a plurality of features based at least in part on the hierarchical data, and may select a subset of the features at a first level as candidates for consolidating to a next higher level in the hierarchical structure. The computing device may determine that a predicted loss of information from consolidating the subset of features is less than a threshold, and may revise the hierarchical structure to include a consolidated feature at the next higher level, rather than the subset of features. In some examples, a statistical model may be trained based on the revised hierarchical structure and used at least partially to make a determination, send a notification, and/or control a device.

Abstract: In some examples, a computing device may receive sensor data associated with a plurality of patient devices that are associated with a plurality of patients. The computing device may further receive caregiver records corresponding at least partially to the sensor data. At least two groups of indicators may be determined from the caregiver records, such a based on a selected subject. Further, the computing device may determine a plurality of clusters from the sensor data. Based on the plurality of clusters and the at least two groups, the computing device may determine an indication of a discrepancy in care for a patient of the plurality of patients. Based on the indication of the discrepancy, the computing device may send at least one of a notification to a caregiver computing device, a notification to a monitoring computing device, or a control signal to one of the patient devices.

Abstract: Example implementations described herein are directed to a decision-support system for maintenance recommendation that uses analytics technology to evaluate the effectiveness of a maintenance action or a group of actions in improving the performance of equipment and its components, and provide recommendations on which maintenance actions or a group of maintenance actions should be pursued and which should be avoided.

Abstract: Example implementations described herein are directed to detection of anomalous events and locations on the transmission power system using phasor management unit (PMU) data, which provides information to grid operators for online decision support. From the high-resolution time synchronized PMU data, the transient abnormal events can be monitored and locations can be disclosed to operators for remedy actions. Utilization of PMU information for such decision support compliments operation practices relying on supervisory control and data acquisition (SCADA) measurements at much lower data resolution. Accurate identification of event locations can further advise grid operators the root cause of disturbances and illuminate possible cascading failures. Implementations of the proposed technology may improve the resilience and reliability of the transmission power systems.

Abstract: Example implementations described herein are directed to detection of anomalous events and locations on the transmission power system using phasor management unit (PMU) data, which provides information to grid operators for online decision support. From the high-resolution time synchronized PMU data, the transient abnormal events can be monitored and locations can be disclosed to operators for remedy actions. Utilization of PMU information for such decision support compliments operation practices relying on supervisory control and data acquisition (SCADA) measurements at much lower data resolution. Accurate identification of event locations can further advise grid operators the root cause of disturbances and illuminate possible cascading failures. Implementations of the proposed technology may improve the resilience and reliability of the transmission power systems.

Abstract: System, including method, apparatus, and computer-readable media, for evaluating client status for a likelihood of churn. Client data may be received, with the client data representing events from a set of different event types performed by clients. Parameters of a statistical model that describes client behavior may be estimated using a computer and based on the client data. A churn type of event may be encoded in the statistical model as an absorbing state of a stochastic process, with a time of transition to the absorbing state modeled as being infinite. At least one of the parameters may correspond to the churn type of event. A likelihood of churn may be calculated for a plurality of the clients at one or more time points using the statistical model and its estimated parameters.

Abstract: Embodiments of the present invention are directed to methods and systems for developing customer retention and loyalty strategies. In one aspect, a method comprises calculating (202) likelihoods of next action taken by customers, based on customer attributes and associated attribute weights stored in a customer data base, and calculating (203) customer churn-risk scores, based on customer attributes that vary over time using the computing device. The methods also determines (207) what-if-scenarios for each customer based on churn-risk scores in order to identify the next-best-action to reduce probability of customer churn, and determines (208) when-to-act time thresholds for each customer based on churn-risk scores in order to identify when a non-high risk customer of churning will likely become a high-risk customer of churning at some later time.

Abstract: System, including method, apparatus, and computer-readable storage media, for evaluating probabilities of next actions by customers to permit selective customer targeting. Customer data (20) may be received (32). The customer data (20) may represent a plurality of actions (14) taken by customers (12). At least a portion of the customer data (20) may be transformed (34) according to action number into stratified data (80) including strata, with each of the strata representing actions for one or more action numbers (74). A conditional proportional hazard function (84) may be estimated (36) from a stratum of the stratified data (80). Likelihoods of a next action may be calculated (38) using the hazard functions. The likelihoods may be the likelihoods of a next action at one or more times by individual customers (12) whose latest action has an action number for which the stratum represents actions.