Jean embodies many ideas about human cognitive development and concept learning, integrating them in an agent architecture that interacts with simulated 3-D virtual environments. Jean perceives, acts, learns, and forms memories. The ideas and inspiration behind these systems comes from many sources:
From Piaget we borrow the idea that children learn some of what they know by repeatedly executing schemas. Executing schemas is in a sense rewarding, and some new schemas are modifications or amalgamations of old ones.
From the Image Schema theorists, we borrow the claim that primitive schemas are encodings or redescriptions of sensorimotor information; and these schemas are semantically rich, general, and extend or transfer to new situations, some of which have no salient sensorimotor aspects.
Another idea, originated from our own group, posits that semantic distinctions often depend on dynamics --- how things change over time --- and so schemas should have a dynamical aspect. Jean acts and learns in a continuous space, and thus she must be able to represent the world in terms of its changes over time.
Finally, we extend another idea from our group to design a learning algorithm driven by the desire to reduce the entropy of the learned model. In the past we've explored the idea of using entropy to discover and define "chunks" of data or experience. Here the discovery of these chunks becomes the basis for Experimental State Splitting, which ties together schemas, dynamics, and action by learning to combine them together to form gists or memories.
The novel algorithms and technologies that comprise Jean are many. The major components include Experimental State Splitting, representation of and reasoning over schema-based dynamic state machines, the extraction and organization of gists in memory, and the learning of behavioral models of other entities in Jean's environment. The research touches upon many other aspects of cognition as well, including motivation, intention, attention, expectation, prediction, and explanation.
The project currently uses two different environments as learning domains. Each provides unique features and unique opportunities for learning.
This work was supported by the Defense Advanced Research Projects Agency (DARPA) contract through the Transfer Learning program via contract N00173-05-1-G035 and the Biologically-Inspired Cognitive Architectures (BICA) program via contract FA8650-05-C-7266.
© 2006, Learning and Development Center, USC ISI. All rights reserved.