Microsoft Technology Licensing, LLC

Hewlett-packard Development Company, L.P.

Apple Inc.

Zynga Inc.

Moises Goldszmidt

Ira Cohen

Timothy S Paek

Terence Patrick Kelly

Juan Carlos Garcia

Alexandros Ntoulas

Vishwas Kulkarni

Anatoly D Adamov

Julia R Reisler

Yu-Chung Hsiao

Pavan Chitta

Daniel R Delling

Andrew V Goldberg

Paul Barham

John Philip MacCormick

Peter Bodik

Julie Anna Symons

Jason Bucher

Brett Bauleke

Samer Ead

Aleksandr Simma

Xinxian Huang

Eric J Horvitz

Kannan Achan

John D Davis

John Charles Krumm

Yue Hua Zhang

Kimberly Keeton

Richard John Black

Edith Cohen

Renato F Werneck

Mihai Budiu

Paul B Koch

Hans Christian Andersen

Rebecca Isaacs

David L Racz

Richard Mortier

Aleksandrs Slivkins

Steve Y Zhang

Ittai Abraham

Michael Pechuk

Xuyang Tan

Hongzhong Jia

David Kempe

Suzanne M Rivoire

Lev Ratinov

Juergen Wiest

Yuanli Pei

Tobias Martin Gindele

Dawn Woodard

Galina Nogin

Fabian Fuchs

Renato Fonseca Furquim Werneck

Joseph K Bradley

Shiri Chechik

Systems and methods are disclosed for triggering an update to a machine-learning model upon detecting that a distribution of particular (e.g., recently collected) input data set is sufficiently different from a distribution training input data set used to train the model. The distributions may be determined to be sufficiently different when a classifier can identify to which distribution individual data elements belong (e.g., to at least a predetermined degree). An update to the machine-learning model can include morphing weights used by the model and/or retraining the model.

Automated input-data monitoring to dynamically adapt machine-learning techniques

An encoding of an environment for operating vehicles is obtained, comprising a combination of at least a representation of moving entities with a graph representation of static infrastructure elements. Using the encoding and a set of one or more observations of the environment state, a machine learning model is trained to produce a probabilistic representation of a set of predicted states of the environment. A trained version of the machine learning model is stored and deployed at one or more vehicles to help plan and control vehicle movements.

Generating environmental input encoding for training neural networks

A system for automated tuning of a computer-implemented game is configured to enable definition of a performance metric indicative of player performance in a computer-implemented game that has tunable gameplay parameters. A performance target is defined that represents target values for the performance metric during progress in the game. The system executes a gameplay simulation using an automated player, and performs an iterative tuning operation based on results of the simulation. The tuning operation automatically determines a suggested value set for the tunable parameters.

Automated tuning of computer-implemented games

A system, a machine-readable storage medium storing instructions, and a computer-implemented method are described herein for a Clustering Engine that determines that respective actions, performed in a first instance of a virtual environment by a first user during a first time range, correspond with a first latent state. The Clustering Engine determines that respective actions, performed in a second instance of the virtual environment by a second user during the first time range, correspond with a second latent state. The Clustering Engine triggers a first virtual environment feature based on a first latent state parameter space for the first user. The Clustering Engine triggers a second virtual environment feature based on a second latent state parameter space for the second user.

Dynamic virtual environment customization based on user behavior clustering

A system for predicting variability of travel time for a trip at a particular time may utilize a machine learning model including latent variables that are associated with the trip. The machine learning model may be trained from historical trip data that is based on location-based measurements reported from mobile devices. Once trained, the machine learning model may be utilized for predicting variability of travel time. A process may include receiving an origin, a destination, and a start time associated with a trip, obtaining candidate routes that run from the origin to the destination, and predicting, based at least in part on the machine learning model, a probability distribution of travel time for individual ones of the candidate routes. One or more routes may be recommended based on the predicted probability distribution, and a measure of travel time for the recommended route(s) may be provided.

Predicting and utilizing variability of travel times in mapping services

One or more all-distances sketches are generated for nodes in a graph. An all-distances sketch for a node includes a subset of the nodes of the graph, and a shortest distance between the node and each of the nodes in the subset of nodes. The generated all-distances sketches are used to estimate the closeness similarity of nodes. The estimated closeness similarity can be used for targeted advertising or for content item recommendation, for example.

Estimating similarity of nodes using all-distances sketches

One or more techniques and/or systems are provided for controlling a travel route planning module associated with a user device. Travel related data, for a user and regarding previously traveled routes by the user, may be indicative of user travel preferences and/or behaviors. The travel related data is evaluated against computed routes derived from different weighting values applied to travel metrics (e.g., a cost associated with a U-turn, a highway, an industrial zone, etc.). For example, weighting values may be iteratively adjusted to generate a plurality of modified computed routes that may be evaluated to identify a target computed route having a similarity to a previously traveled route of the user above a threshold. User preference weighted travel metrics, generated based upon weighted travel metrics of the target computed route, are used to control a travel route planning module to generate a customized travel route for the user.

Controlling travel route planning module based upon user travel preference

A graph of a social network is received. The graph may include a node for each user account and an edge between nodes that represent social networking relationships such as messages between the user accounts or a friend relationship. The graph is transformed into a transformed graph where nodes have direct edges depending on a local test among its neighbors in the original graph. Small subsets of the transformed graph are categorized. The categories are used to identify subgraphs in the transformed graph. Each subgraph is grown by adding an edge from the transformed graph to the subgraph depending on local tests among nodes associated with the edge that have at least one edge that is already in the subgraph. The categorized subgraphs are used to provide targeted advertising, suggest new connections, identify different personalities and interests of users, or to provide other services.

Identifying subgraphs in transformed social network graphs

Power consumption of computing devices are monitored with performance counters and used to generate a power model for each computing device. The power models are used to estimate the power consumption of each computing device based on the performance counters. Each computing device is assigned a power cap, and a software-based power control at each computing device monitors the performance counters, estimates the power consumption using the performance counters and the model, and compares the estimated power consumption with the power cap. Depending on whether the estimated power consumption violates the power cap, the power control may transition the computing device to a lower power state to prevent a violation of the power cap or a higher power state if the computing device is below the power cap.

Estimating and managing power consumption of computing devices using power models

In a distributed system a plurality of devices (including computing units, storage and communication units) are monitored by an automated repair service that uses sensors and performs one or more repair actions on computing devices that are found to fail according to repair policies. The repair actions include automated repair actions and non-automated repair actions. The health of the computing devices is recorded in the form of states along with the repair actions that were performed on the computing devices and the times at which the repair actions were performed, and events generated by both sensors and the devices themselves. After some period of the time, the history of states of each device, the events, and the repair actions performed on the computing devices are analyzed to determine the effectiveness of the repair actions. A statistical analysis is performed based on the cost of each repair action and the determined effectiveness of each repair action, and one or more of the policies may be adjusted, as well as determining from the signals and events from the sensors whether the sensors themselves require adjustment.

Repair-policy refinement in distributed systems

In a distributed storage system such as those in a data center or web based service, user characteristics and characteristics of the hardware such as storage size and storage throughput impact the capacity and performance of the system. In such systems, an allocation is a mapping from the user to the physical storage devices where data/information pertaining to the user will be stored. Policies regarding quality of service and reliability including replication of user data/information may be provided by the entity managing the system. A policy may define an objective function which quantifies the value of a given allocation. Maximizing the value of the allocation will optimize the objective function. This optimization may include the dynamics in terms of changes in patterns of user characteristics and the cost of moving data/information between the physical devices to satisfy a particular allocation.

Data allocation and replication across distributed storage system

Methods for automatically identifying and classifying a crisis state occurring in a system having a plurality of computer resources. Signals are received from a device that collects the signals from each computer resource in the system. For each epoch, an epoch fingerprint is generated. Upon detecting a performance crisis within the system, a crisis fingerprint is generated consisting of at least one epoch fingerprint. The technology is able to identify that a performance crisis has previously occurred within the datacenter if a generated crisis fingerprint favorably matches any of the model crisis fingerprints stored in a database. The technology may also predict that a crisis is about to occur.

Automated identification of performance crisis

Growing community of inventors

Palo Alto, CA, United States of America

Moises Goldszmidt