Thursday, February 10, Volen 101, 2:00-3.00 pm
Today's Internet is seen as a virtual library, classroom, post office and shopping mall. Many groups collect statistics on web usage and use this data for commercial purposes. In this talk, I will discuss the Internet as a "virtual laboratory," showing how the mass of clickstream data, which the Internet uniquely enables us to gather, can be used to address issues in both human and machine learning.
One of the most difficult problems facing the field of machine learning has been finding a practical way to inject the knowledge of humans into an artificial system. With the Internet as our laboratory, we can observe the behavior of thousands of humans and train our artificial system based on these interactions. In our early work with these ideas, we found that the humans interacting with our web site were learning. This observation led us to develop CEL (the Community of Evolving Learners), an Internet learning community where humans play educational games, with each other and with software agents, and both populations of learners -- human and machine -- advance.
CEL was designed as a framework to enable broad experimentation in machine and human learning. The system is available to anyone with Internet access and a Java-enabled web browser. I have been using CEL as a laboratory in which to study evolutionary computation with humans in the loop. The latter half of this talk will discuss an experiment in which two educational games were used in a pilot study at a primary school, and both games were controlled by an evolutionary algorithm.
Host: Jordan Pollack