Transfer Learning refers to the problem of transferring knowledge learned from a task, domain, or distribution to learn faster in a new similar task, domain, or distribution. We are exploring transfer of skills learned in playing different versions of real-time strategy games like Wargus. We consider transferring domain models, task hierarchies, policies and value functions across different games that share qualitative dynamics, reward functions, and other features. [DARPA]
How is it possible to learn from a single demonstration of a complex task such as scheduling meetings or managing the air-space? In this project we are exploring the problem of learning the utility function of an expert from a small number of demonstrations given by the expert. We treat the expert as solving a constrained optimization problem, and seek to learn the objective function he or she is maximizing. We are working with researchers in Lockheed-Martin and other universities including Rensselaer Polytechnic, Arizona State, Georgia Tech, and others. [DARPA]
Today's learning algorithms require machine learning experts to develop them and maintain them. We are exploring architectures to make the application of learning much easier and can be initiated by the user. The user will be able to instruct the machine to learn a new task, and the system will collect relevant training data, use an appropriate learning algorithm, and use its predictions to help the user when needed. For example, the computer may be asked to learn to remind the user when he or she forgets to attach a document before hitting the send button when sending email. [DARPA]
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 using modern learning methods such as reinforcement learning and search 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.