I am interested in artificial intelligence, especially in machine learning. Below are some research projects with opportunities for new students.
Personal learning apprentices. This project aims to develop personalized software agents that help users perform specific tasks (e.g., manage their calendar, notice upcoming seminars of interest, reserve a meeting room), and that learn to improve throughout their lifetime by observing their users. We have recently developed one such learning apprentice: a personal calendar manager that helps schedule meetings, and that learns to provide advice about meeting dates, locations, durations, etc. This system is presently in routine use by one secretary, learns automatically each evening, and has successfully learned hundreds of simple rules for recommending appropriate durations and locations for meetings. New research is now needed to improve the underlying learning mechanisms, test the system's ability to learn from multiple users (we hope to distribute the system throughout SCS), and develop approaches for collaboration and colearning among multiple software agents.
Learning robots. The goal of this work is to develop robots which learn from experience. Much of our recent research has focused on reinforcement learning: an inductive method for learning control strategies based on delayed reward. While we have succeeded in using such learning techniques to acquire control strategies for simple robot tasks (e.g., to dock on the battery charger), the most critical new research issue is to scale up these learning techniques to more complex and realistic robot learning scenarios. This will require new methods for learning to control robot sensors, adding hierarchical organization to the robot system, learning and utilizing background knowledge, etc. One active line of work in this area is a new learning method, described in the following paragraph.
Explanation-based neural network learning. There are fundamentally two types of approaches to machine learning: inductive and analytical. Inductive learning methods (such as neural network backpropagation learning) find regularities by examining large numbers of training examples. Unfortunately, this approach requires huge numbers of training examples to learn complex functions. Analytical learning (such as symbolic explanation-based learning) uses prior knowledge of the learner to produce more correct generalizations from fewer examples. Unfortunately, this analytical approach requires substantial background knowledge, and is sensitive to errors in this background knowledge. We have recently developed an approach to unifying neural network and explanation-based learning approaches, in an attempt to combine the advantages of these complementary methods. New research is now needed to explore and extend this approach, and to try it out on problems such as robot learning.
Learning to design. Within CMU's Engineering Design Research Center, a variety of intelligent design aids have been developed. Within this context, there are many opportunities to study the role of learning in design. For example, it may be possible to organize interactive design aids like MICON so that they learn new design techniques by observing their users and/or exploring large design spaces and evaluating resulting designs. I would like to find a student interested in learning and design, to start up a new project in this area.
|The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.|
Contact Us | Update Instructions