Richard K. Belew, Director
We are a research group within the AI laboratory. Our name, CCSRG, is meant to indicate that while our approach is primarily computational, we are informed by a cognitive perspective on both artificial and natural systems.
The focus of our research group is the characterization of adaptive knowledge representations. Issues of representation have always played a central role in artificial intelligence (AI), as well as in computer science and theories of mind more generally. We would argue that most of this work has (implicitly or explicitly) assumed that the representational language is wielded manually, by humans encoding an explicit characterization of what they believe to be true of the world. We believe there are fundamental philosophical difficulties inherent in any such approach. Further, there now exist modern machine learning techniques capable of automatically developing elaborate representations of the world. To date, however, the representations underlying this learning have not shown themselves able to "scale up" to the semantically sophisticated task domains often associated with AI expert systems. We believe it is therefore appropriate to reconsider basic notions of what makes for good knowledge representation, with constraints imposed by the learning process considered sine qua non but in conjuction with others (expressive adequacy, valid inference, etc.) more typically considered by AI.
We have found it productive to pursue this general interest through several more specific research projects. The first applies statistical techniques to the problem of free-text information retrieval (IR) and linguistics more generally. Many of our projects use a connectionist (neural network) representation of documents and descriptive keywords that uses relevance feedback as a training signal to a reinforcement learning algorithm. This construction allows an IR system to learn a more effective indexing representation of free-text documents as a simple by-product of the browsing behaviors of its users. Second, we have investigated a wide range of Genetic Algorithm (GAs) applications, ranging from use in "artificial life" models of natural phenomena to use as an artificial inductive method to accomplish an engineering goal like optimizing a function. We believe our work in these two areas allows a ``stereoscopic'' view of cognitive adaptation, encompassing a broad range of fundamental issues from low-level, biological constraints to high-level, symbolic communication.