Research Projects

Prasad Tadepalli

Active Transfer Learning (with Alan Fern)

Active Transfer Learning studies the problem of actively interacting with an expert to transfer knowledge from one domain to another domain. We are exploring active transfer learning in the context of Markov Decision Processes, as applied to a number of domains including real-time strategy games like Wargus. We consider a variety of interactions including demonstrations, expert critique, and trajectory queries, and a variety of kinds of knowledge to transfer including domain models, task hierarchies, policies, and value functions. [ONR]

Deep Reading and Learning (with Tom Dietterich and Xiaoli Fern)

Getting machines to read and understand text has enormous benefits and is tremendously challenging. In this project we expore the problem of natural language understanding through machine learning. In particular we are applying a search-based structured prediction approach to learn to map natural langauge text to meaning representations. Natural language texts are radically incomplete in that only a small part of the whole truth is ever mentioned. They are also systematically biased by factors such as novelty and salience. We are exploring an approach based on modeling the systematic biases and bootstrapping from them so that it is possible to infer fairly robust and complete meaning representations from radically incomplete texts. [DARPA]

Integrating Learning and Search via Structured Prediction (with Alan Fern)

Structured prediction refers to the problem of learning to map structured inputs to strcutured outputs. A variety of problems in natural langauge processing and machine learning fall under the category of structured prediction. In this project we are exploring an a search-based approach to structured prediction. The problem is formulated as learning a cost metric that evaluates the match between the input and the output, and a search heuristic that learns to guide the search toward finding the best least cost output for a given input. The former problem can be viewed as cost-based structured prediction, and the latter problem can be viewed as speedup learning. We are applying this work to a variety of settings including coreference resolution in natural language understanding and object recognition in computer vision. [NSF]

Learning Service Policies in Complex Domains (with Alan Fern and Roni Khardon)

In any complex real-world problems such as fire-fighting in a city, there are many choices such as allocating zones for each fire station, staffing them appropriately, and allocating resources and personnel to a task. We are exploring relational reinforcement learning and decision-theoretic planning to optimize these decisions to particular cities with resource constraints. These methods search the space of policies intelligently to change the policies in the direction of improved performance. We are collaborating with the Corvallis Fire Department and Coelo, Company of Design on this project. [NSF]