Google Inc.

Christopher Kenneth Harris

Samuel Sze Ming Ieong

Chenyu Zhao

Di Wang

Vinod Kumar Ramachandran

Shibani Sanan

Yifang Liu

Sissie Ling-le Hsiao

Konstantinos Katsiapis

Nicola Rettke

Anthony Chavez

Aranyak Mehta

Sissie Ling-Ie Hsiao

Yefei Peng

Xi Tan

Gabriel Florin Constantin

Sergey Levine

Kanury Kanishka Rao

Samuel Ieong

Konstantinos Bousmalis

Bin Liu

Timothy S Lipus

Julian Ibarz

Patrick Hummel

Randolph Preston McAfee

Alexander Irpan

Joshua V Dillon

Methods, systems, and apparatus include computer programs encoded on a computer-readable storage medium, including a system that controls content distribution using a feedback loop. Content is distributed over multiple different online channels using a same initial selection value for distribution over each different online channel. An observed user actions required for distribution of the content over the multiple different online channels is received through a feedback loop and for multiple different distributions of the content. Based on the observed user actions received through the feedback loop, a predicted user action rate is determined for the multiple different distributions across the multiple different online channels. The selection value is adjusted based on a difference between the predicted user action rate and a reference distribution amount specified by a provider of the content. The content is distributed over the multiple different online channels using the adjusted selection value.

Optimization of a multi-channel system using a feedback loop

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for off-policy evaluation of a control policy. One of the methods includes obtaining policy data specifying a control policy for controlling a source agent interacting with a source environment to perform a particular task; obtaining a validation data set generated from interactions of a target agent in a target environment; determining a performance estimate that represents an estimate of a performance of the control policy in controlling the target agent to perform the particular task in the target environment; and determining, based on the performance estimate, whether to deploy the control policy for controlling the target agent to perform the particular task in the target environment.

Off-policy control policy evaluation

When a content item is initially served to a client device, the content item may result in an impression effect. As time elapses, the initial impression may fade. Such a decay of the impression effect may be predicted through the use of a predictive model. In some implementations, one or more impression effect parameters may be accessed and used with the predictive model to determine a decay factor or predicted value that incorporates the impression effect decay for a content item. A value, such as a score, may be determined based on the decay factor or the predicted value and a bid associated with a content item. A content item may be selected based on the determined value and data to effect presentation of the content item may be provided.

Content item impression effect decay

Methods, systems, and apparatus include computer programs encoded on a computer-readable storage medium, including a method for providing content. Data specifying a post-install activity is received from a provider of an application. An opportunity is identified to provide third-party content to a user. A likelihood is determined that the user will perform the specified post-install activity based on one or more attributes of the user and attributes of users that have previously performed the specified post-install activity in the application. A selection value is adjusted for third-party content that identifies the application based on the determined likelihood, wherein the selection value increases as the likelihood increases. The third-party content identifying the application is selected based on the adjusted selection value. The third-party content identifying the application is distributed to a client device of the user.

Post-install application interaction

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for optimizing digital components. In one aspect, the system includes a digital component server that identifies a digital component to be presented in an electronic document. The system also includes a digital component distribution server that interacts with the digital component server to determine that the electronic document includes a particular item that identifies a same entity as the digital component that was identified to be presented in the electronic document. The system combines at least a portion of the digital component with content from the particular item to form an optimized digital component. The optimized digital component is integrated into the electronic document rather than presenting either of the particular item or the digital component individually.

Optimized digital components

Methods, systems, apparatus, including computer programs encoded on a computer storage medium, for determining whether to execute a query based on a predicted computerized rendering period The method may include actions of receiving a query, determining a set of one or more features based on the query, generating a query vector, providing the query vector to a machine learning model, receiving first data, based on the output of the machine learning model, that is indicative of whether the computerized rendering period associated with the query is likely to satisfy the predetermined threshold, determining based on the first data that the computerized rendering period available for displaying a set of one or more candidate content items that are responsive to the query is not likely to satisfy the predetermined threshold, and determining to not use a search engine to execute the received query.

Content search engine

Systems and methods of a balancing data requests over a computer network are provided. The system can receive a request for content to be rendered in a content slot on a computing device. The content slot can be coded with a default timer value. The system can execute a real-time content selection process to identify content items that each have a magnitude. The system can determine a timer value based on the magnitudes, and transmit the timer value to the computing device to override the default timer value to cause the computing device to replace a first content item rendered in the content slot upon expiration of the timer.

Balancing data requests over a network

Systems and methods of the present disclosure are directed generally to facilitating content selection by identifying low impact criteria. In some implementations, a data processing system accesses a data structure storing, in a memory element, a plurality of impression records. Each impression record can include one or more features and an indication of user interest corresponding to a content impression. The data processing system can identify a combination feature based on at least two of the features. The data processing system can execute a statistical model (e.g., logistic regression model) using the impression records and the combination feature. The data processing system can determine a weight for the combination feature. Responsive to the weight being less than a threshold, the data processing system can transmit an indication to disable the combination feature for selecting content associated with the plurality of impression records.

Content Selection

Methods, systems, and apparatus include computer programs encoded on a computer-readable storage medium, including a system that controls content distribution using a feedback loop. Content is distributed over multiple different online channels using a same initial maximum selection value for distribution over each different online channel. An observed distribution amount required for distribution of the content over the multiple different online channels is received through a feedback loop and for multiple different distributions of the content. Based on the observed distribution amount received through the feedback loop, a realized distribution amount is determined for the multiple different distributions across the multiple different online channels. The maximum selection value is adjusted based on a difference between the realized distribution amount and a reference distribution amount specified by a provider of the content. The content is distributed over the multiple different online channels using the adjusted maximum selection value.

Growing community of inventors

Santa Clara, CA, United States of America

Christopher Kenneth Harris