Coactive Learning

Researchers: Lee Becker, Dan Grecu and Shawn Nicholson

Project:

Coactive learning is a multi-agent learning paradigm that we have developed at WPI to investigate collaborative learning schemes. The coacting approach models collaborative learners focusing on the same task that support each other in their learning processes. The learners are assumed to be incremental and they provide each other with feedback and/or support during learning. For example, an agent may consult another agent about how to classify a given training instance, or ask the other agent what knowledge (e.g., rule or training instance) it would use to classify a new instance.

Coactive learning has been successfully used to achieve an emergent noise filtering effect between learners that do not possess noise filtering algorithms, to reduce the number of instances needed by an agent for training, to parallelize learning processes, and to compensate for class distribution biases in the training sets of the agents. The coactive learning model allows to run simulations that combine learners in various interaction schemes, and provides insight into the impact of interaction structure and density on learning.

Publications:

• D. L. Grecu & L. A. Becker, "Coactive Learning for Distributed Data Mining", Fourth International Conference on Knowledge Discovery in Databases - KDD-98, 1998, Menlo Park, CA: AAAI Press, pp. 209-213.

• D. L. Grecu & L. A. Becker, "Agent Experiments for Strategies for Human Collaborative Learning", 3rd International Conference on the Learning Sciences, Atlanta, GA, December 1998, pp. 118-124.