Abstract: Techniques for agent-assist systems to provide context-aware, subdocument-granularity recommended answers to agents that are attempting to answer queries of users. The agent-assist system may obtain collections of documents that include information for responding to queries, and analyze those documents to identify subdocuments that are associated with different semantics or meanings. Subsequently, any queries received can be analyzed to identify their semantics, and relevant subdocuments can be identified as having similar semantics. When the agent-assist system presents the agent with the relevant documents, it may highlight or otherwise indicate the relevant subdocument within the document for quick identification by the agent. Further, the agent-assist system may collect feedback from the agent and/or user to determine a relevancy of the recommended answers. The agent-assist system can use the feedback to improve the quality of the recommended answers provided to the agents.

Abstract: A computer-implemented Tree Alternating Optimization (TAO) algorithm for learning decision trees to find an approximate minimizer of an objective function over the parameters of the tree. Generally, the method comprises inputting an initial decision tree and a training set of instances, processing the initial decision tree by partitioning nodes into sets of non-descendant nodes, processing the nodes in each set by updating the nodes' parameters at each iteration so that the objective function decreases monotonically, and pruning the tree, which produces a final tree of a size no larger than that of the initial tree. TAO applies to many different types of loss functions, regularization terms and constraints, and types of models at both the decision nodes and the leaves, and makes it possible to learn better decision trees than with traditional algorithms, and to learn trees for problems where traditional algorithms do not apply.

Abstract: A method of reinforcement learning in a processing element, the method including receiving, by a receiving module, one reward. Further, a computing module computes a Q-value for a first dimension at time tn, based on the reward. The Q-value is locally stored. A time-division multiplexing module replaces the computed Q-value for the first dimension with at least one Q-value computed for a second dimension at time tn+1. The second dimension is different than the first dimension.

Abstract: The present disclosure describes techniques for training a model using cross-domain adaptation to classify content requests from client devices having unknown attributes. The system can obtain requests for content from client devices of a first domain, and requests for content from client devices of a second domain. The system can train a model by propagating request attributes of the first domain through the model to generate first internal data from an internal layer of the model and a first output vector an output layer of the model. The system can propagate request attributes of the second domain to generate second internal data, and determine a difference between the first internal data and the second internal data. The system can update the model based on the difference and the output vector, and classify a third client device of the second domain using the model.