Back to results
US 12,586,683 B2Granted
Decision-making under selective labels
AssigneeInternational Business Machines Corporation
InventorDennis Wei
Life of this patent
Estimating the term…
Abstract
A computer-implemented method of decision-making using selective labels, includes receiving a conditional success probability value of a feature associated with an entity. A confidence value of the received success probability value is received. A parameter value that is a trade-off between a short-term learning and a long-term utility is selected. A decision is rendered to accept or reject the feature associated with the entity according to a machine learning policy.
Claims17
A computer-implemented method of automatic decision-making using selective labels, the method comprising:
- receiving, at a success probability model of a computing device, a feature vector associated with an entity and generating, via the success probability model analyzing the feature vector, a conditional success probability value representing a likelihood of success of a feature associated with the entity if the entity is accepted;
- receiving, at a confidence model of the computing device, a sample size associated with prior accepted entities having similar features and generating, via the confidence model analyzing the sample size, a confidence value associated with the conditional success probability value;
- determining, at a discount factor model of the computing device, a tuning parameter value that is <1 and that represents a trade-off between a short-term learning cost and a long-term utility;
- rendering a decision, using a policy model of the computing device, to accept the entity based on a threshold function of the conditional success probability value, the confidence value, and the tuning parameter value;
- and
- iteratively updating the success probability model and the confidence model using an observed outcome that results from accepting the entity.
3 more elements
A computing device configured for decision-making using selective labels, the computing device comprising:
- a processor,
- a memory coupled to the processor, the memory storing instructions to cause the processor to perform acts comprising:
- receiving, at a success probability model of a computing device, a feature vector associated with an entity and generating, via the success probability model analyzing the feature vector, a conditional success probability value representing a likelihood of success of a feature associated with the entity if the entity is accepted;
- receiving, at a confidence model of the computing device, a sample size associated with prior accepted entities having similar features and generating, via the confidence model analyzing the sample size, a confidence value associated with the conditional success probability value;
- determining, at a discount factor model of the computing device, a tuning parameter value that is <1 and that represents a trade-off between a short-term learning cost and a long-term utility;
- rendering a decision, using a policy model of the computing device, to accept the entity based on a threshold function of the conditional success probability value, the confidence value, and the tuning parameter value;
- and
- iteratively updating the success probability model and the confidence model using an observed outcome that results from accepting the entity.
5 more elements
A computing device configured to perform decision-making using selective labels, the computing device comprising:
- one or more processors including a dialog processor configured to process extracted text from a plurality of participants of a collaborative query;
- a memory coupled to the one or more processors;
- a plurality of models configured in the one or more processors, the plurality of models comprising:
- a success probability model configured to provide a conditional success probability value representing an empirical success rate of a feature associated with an entity;
- a confidence model configured to provide a confidence value of the conditional success probability value, based at least on a sample size of entities;
- a discount factor model configured to determine a tuning parameter value that is <1 and that represents a trade-off between a short-term learning cost and a long-term utility;
- and
- a policy model configured to render a decision to accept the feature associated with the entity based on a threshold function of the conditional success probability value, the confidence value, and the tuning parameter value;
- and
- iteratively update the success probability model and the confidence model with a resultant outcome when the decision is to accept.
7 more elements
Dependencies are inferred from the claim text.