The University of Arizona

Complete Intelligence



Overview

The purpose of this project is to develop a cognitive architecture for robots that implements as many perceptual and cognitive functions as possible. Artificial Intelligence is drunk on performing hard tasks at high levels. Given a choice between power and generality, most of us choose power. Our programs depend on designed exploits, or on designed search spaces in which programs can learn exploits. Divide-and-conquer, specific function, power over generality, and exploits are valuable engineering methods in many disciplines. They are apt to build machines that do one thing well. Human intelligence isn't that kind of machine.

This project focuses not on the design of new algorithms or problem solving methods but on the architecture that integrates extant algorithms and methods to produce intelligent behavior.

The lessons to be learned in this project are about the interactions of abilities -- planning, language, perception, motor control, social interaction, learning and so on. Independent of the implementations of these abilities, we expect to discover truths about how they interact. For example, it has been known for decades that humans interleave top-down, or expectation-driven processing with bottom-up, or data-driven processing; but few researchers have tried this kind of integration in computational systems. Is it a good idea? In which kinds of tasks does it confer advantage? Another example, which drives much of our research, is the conjecture that concepts form in infancy and arise from sensorimotor activity yet serve throughout life and extend to nonphysical situations. No-one has given a formal account (much less one that runs on a computer) of these kinds of concepts.

In sum, this project asks, What empirical laws can we discover when computational systems (and, particularly, robots) are designed not for high performance on a single task but for adequate performance on a wide array of tasks?

People
Publications
Howe, Adele E. and Paul R. Cohen. 1995. Understanding Planner Behavior. Complete Artificial Intelligence, Special Issue on Planning Systems, vol. 76, nos. 1 & 2, pp. 125-166.
Aram Galstyan and Paul Cohen (2007) Empirical Comparison of "Hard" and "Soft" Label Propagation for Relational Classification The 9th International Conference on Inductive Logic Programming, ILP-07, Corvalis, Oregon.
Anderson, Scott D., Adam Carlson, David L. Westbrook, David M. Hart, and Paul R. Cohen . 1994. Tools for Experiments in Planning. Proceedings of the 6th International IEEE Conference on Tools with Artificial Intelligence, pp. 615-623. Also in Workshop Proceedings of the ARPA/Rome Laboratory Knowledge-Based Planning and Scheduling Initiative, Tucson Arizona. Morgan Kaufmann, pp. 423-432. Technical Report 95-01, Dept. of Computer Science, University of Massachusetts/Amherst.
Funding
Battle Smarts: Lifelong, Personalized Learning and Training
DARPA.
November 14 2008 to February 2010
(Paul Cohen (PI))
MABLE: Modular Architecture for Bootstrapped Learning Experiments
SRI.
February 1 2009 to March 31 2010
(Paul Cohen (PI))

Webmaster: Andrey Kvochko