Coactive Learning

Dan Grecu
Department of Computer Science
WPI

Friday, April 2, 1999
11 a.m.
Fuller Labs 311

This talk will describe coactive learning, a new method of multi-agent learning. Coacting agents perform individual and similar learning tasks. The agents exchange training information with each other, such as relevant training examples and/or information about example processing. When the coactive learning process is completed each agent produces an independent learning result.

As a demonstration of the approach and its benefits, coacting is illustrated in a cooperative induction scheme, where the agents use instance-based learning algorithms. Coaction can result in an emergent functionality that is not present in either of the individual agents. For example, coacting agents using versions of instance-based learning algorithms with no noise filtering capabilities achieve a noise reduction performance which surpasses that of stronger algorithms, which have noise filtering functionality. Coacting learners can use a variety of interaction protocols to optimize various measures of learning quality, such as the storage requirements for a learned concept representation and the sensitivity to training class bias. Coactive learning schemes are also able to reduce the number of instances needed by an agent for training, and support the parallelization of learning processes.

The talk will conclude by discussing further possibilities to combine learners in coactive interaction schemes, and the potential impact of interaction characteristics on the learning process and result.

Host

Micha Hofri

Colloquium Coordinator

Carolina Ruiz

Maintained by webmaster@wpi.edu
Last modified: Sep 27, 2006, 16:05 EDT