公开(公告)号：US09245232B1

公开(公告)日：2016-01-26

申请号：US13774767

申请日：2013-02-22

Applicant: Amazon Technologies, Inc.

Inventor： Tyson Christopher Trautmann ,  Peter Varnum Commons ,  Diwakar Chakravarthy ,  Michael Luis Collado ,  Thomas Lowell Keller ,  Benjamin Warren Mercier ,  Zachary Jared Wiggins

IPC: G06N5/00 ,  G06F1/00 ,  G06N99/00

CPC classification number: G06N99/005 ,  G06F17/30902

Abstract: A machine generated service cache that utilizes one or more machine learning classifiers is trained using service requests directed to a human-generated service and service responses generated by the human-generated service in response to the service requests. Once the machine generated service cache has been trained to a predetermined level of performance, the machine generated service cache can be utilized to process actual service requests directed to the human-generated service. The machine generated service cache might be utilized to process service requests for which it is not essential that the returned service response be identical to a response that would be generated by the human-generated service.

Abstract translation: 利用一个或多个机器学习分类器的机器生成的服务高速缓冲存储器是使用针对人类产生的服务的服务请求和由人类生成的服务响应于服务请求生成的服务响应进行训练的。 一旦机器生成的服务高速缓存已经被训练到预定的性能水平，则机器产生的服务高速缓存可以用于处理针对人造服务的实际服务请求。 机器生成的服务高速缓存可以用于处理服务请求，其中所返回的服务响应与由人为生成的服务生成的响应相同并不是必需的。

公开(公告)号：US10936955B1

公开(公告)日：2021-03-02

申请号：US15406282

申请日：2017-01-13

Applicant: Amazon Technologies, Inc.

Inventor： Pramod Muralidharan ,  Ray Chen ,  Donald Kane ,  Vignesh Kannappan ,  Hancao Li ,  Ajay Kant Singh ,  Robert Andrew Kreek ,  Richard Bargreen ,  Rui Chen ,  Michael Luis Collado ,  Kevin Donald Kelly ,  Shen Li ,  Sameer Rajendra Rajyaguru ,  Serguei B. Stepaniants ,  Murali Brahmadesam

IPC: G06N5/02

Abstract: Technologies are disclosed for determining network-accessible content changes based on computed models and providing a long term forecast of user interaction at a network accessible site based upon a short term experiment at the site. A forecast model for a period of time is generated based upon historical data of user interactions at the site. An experiment is run for a short term at the site based upon a potential change at the site. Based upon data obtained during the experiment, scores are generated for a control group (no change) and a treatment group (potential change) and compared. If there are statistically significant differences between the control group and the treatment group scores, the long term forecast may be used to forecast what the long term impact of the experiment would be based upon the short term experiment.

公开(公告)号：US10152458B1

公开(公告)日：2018-12-11

申请号：US14661981

申请日：2015-03-18

Applicant: AMAZON TECHNOLOGIES, INC.

Inventor： Serguei B. Stepaniants ,  Kyle Leon Battisti ,  Rui Chen ,  Michael Luis Collado ,  Kevin Donald Kelly ,  Sebastian Kohlmeier ,  Kevin McAlister ,  Daniel Parshall

IPC: G06F17/18 ,  G06F17/10 ,  G06F17/30

Abstract: Described are techniques for determining long-term effects of an experimental change to a user experience after the end of the experiment. A control state and a treatment state of a statistical hypothesis experiment may be assigned to first and second client devices, respectively, during an experiment time period. Subsequent to the end of the experiment, presentation of the control state may be discontinued. Result data corresponding to the treatment state may be determined during the experiment time period and for a length of time subsequent to the experiment time period. Result data corresponding to the control state may be determined during the experiment time period, and for a length of time prior to assignment of the control state to the first client device.