CUE Home >> Departments >> Artificial Intelligence: Machine Learning By Trial and Error

CUE 2004

COMPUTER SCIENCE
Artificial Intelligence: Machine Learning By Trial and Error

CU faculty member Greg Grudic studies robot learning tasks on rough terrain in south Boulder.

Artificial Intelligence (AI) is a field of computer science that has recently entered the public consciousness, at least in part due to the film of the same title, as well as the Terminator and Matrix series.

Although AI has been an active area of research for over 40 years, no existing AI system comes close to matching the general-purpose intelligence and reasoning capability of movie androids. A key difficulty has been that vast amounts of knowledge and experience are required to reason, understand, and think like a human. Traditional approaches have tried to program the knowledge directly into the AI system, but that is a laborious effort and often results in brittle systems that fail in unforeseen circumstances.

Instead of programming a system with all the knowledge required to perform a task, AI researchers have turned to designing systems that essentially program themselves by observing an expert or by trial-and-error learning in the world. Research on this topic is called "machine learning."

Greg Grudic, a CU-Boulder assistant professor of Computer Science, studies robot learning tasks. By observing human experts, a robot learns to crudely navigate in its environment. But with subsequent experience, the robot can autonomously refine and optimize its navigation strategies. In conjunction with this work, Grudic develops learning techniques that allow computer systems not only to interpret their environment and make decisions but also to estimate confidence in their conclusions and actions. A robot may consider that a substance it approaches is not hazardous, yet the confidence in this assessment is crucial to determining the appropriate course of action.

Complementing this work on control, Assistant Professor Michael Burl focuses on algorithms for interpreting sensory data, especially as related to vision. In most mammals, a significant portion of the brain's real estate is dedicated to vision, suggesting that vision will be key to artificial systems that interact with their environments. Digital cameras can mimic the data capture functions performed by the retina; however, the real challenge is in extracting meaningful information from the resulting image stream. In addition to robotics applications, Burl develops algorithms that discover geological features in planetary images returned by spacecraft, and algorithms that learn to track people and understand their activities in surveillance video.

Machine learning is also important to the research of other faculty in computer science. Jane Mulligan develops vision-based techniques for automatically estimating human body pose, useful in applications such as activity recognition and monitoring of the elderly. James Martin mines large English-text databases to detect statistical properties of word usage and thereby infer the meaning of words in context. Michael Mozer collaborates with neuroscientists to understand mechanisms of learning in the human brain, and applies ideas from machine learning to understand human learning and vice versa.

Machine learning is a mature field with numerous practical applications. For example, web-based recommendation systems that suggest a product X if you liked another product Y are based on a machine-learning analysis of large data bases of individual preferences. The machine-learning group at CU collaborates with start-up companies that do internet personalization, speech recognition, customer-relationship management, and automatic text summarization. Researchers in other departments of the college also use machine-learning techniques, such as in the control of building energy systems and in developing medical diagnosis equipment that performs automatic detection of irregularities.

www.cs.colorado.edu/

Student Profile: Thomas Strohmann
As a first-year college student, Thomas Strohmann remembers working with a team of students to program a computer to play Mancala, a traditional board game. "It was really fun and none of our group could beat the computer," he recalls. >> More