A.I. is concerned with computational models of the human mind and the synthesis of methods to accomplish human mental tasks. Areas of study range from subconscious mental activities to conscious ones and include image interpretation, speech recognition, language understanding, game playing, expert advice-giving, and theorem proving. Core mental activities such as knowledge representation, reasoning, learning, and planning are under intense investigation for use in robotics, intelligent information retrieval, robust and reliable diagnosis automata, and other high impact applications. An estimated 10,000 expert systems that give expert-level advice in various highly constrained areas have been developed using methods developed in the A.I. field. A.I. research has also given rise to widely used high-level programming languages for symbolic computation, such as LISP and Prolog.
Current A.I. research in our Department includes natural language understanding and dialog modeling, automated reasoning, planning, statistical models and more powerful programming languages for A.I. and other systems development. Research in neural networks and genetic algorithms is done by joint faculty in the Departments of Physics and Electrical and Computer Engineering. The group has a strong analytical focus, but has also implemented important leading-edge systems in each of the areas mentioned above. We also have experience in components of other A.I. subfields such as program synthesis, robotics, machine learning, expert systems, search analysis, and game playing.
The faculty in the A.I. group are listed below. More detailed information appears on subsequent pages.
In the area of natural language processing systems, we seek to develop a class of task-oriented natural language processors that enable a human and machine to collaborate efficiently in problem solving. We emphasize the importance of fast, convenient communication with many interactions per minute using multimode input (voice, typing, and touch screen) and output (synthesized voice and video screen). Our work primarily concentrates on the development of theories of semantics appropriate for such systems with particular emphasis on mechanisms for handling complex noun-phrase resolution, focusing mechanisms, dialogue comprehension, and user models.
In order to test theories, we continuously build natural language interactive systems and study their performance in typical problem-solving situations with human subjects. This experimental work provides examples of human-machine dialogue for study and yields data related to the achieved level of efficiency.
As a second area of research, we seek theories for the extraction of meaning from natural language text. Here we study algorithms that can scan text at high speed and compile meaning structures to answer specific queries.
Selected
Publications
In a nutshell, I study ways of using probability and statistics to make computers better at solving problems. There are two primary reasons that I think probability and statistics are an important topic in artificial intelligence. First, human beings routinely reason about the likelihood of uncertain events and computer programs that can't are often brittle and unreliable. Second, people reason extremely poorly about probability and statistics, so there are many applications for creating software that can do a better job at this than humans.
My recent work examines algorithms for efficiently making optimal decisions under uncertainty in sequential domains. Potential applications of this work include robot navigation and automatic network monitoring and control. Of particular interest to me is the problem of how an autonomous system should best act to gain information; although this problem is solved in principle, its computational complexity makes it necessary to derive special-case solutions for particular instances of the problem.
I have also worked on projects in machine learning (making a machine learn from its mistakes), reinforcement learning (improving performance in sequential tasks such as games via weak feedback), statistical natural language processing (finding relationships between words via statistical analysis of texts), multilingual information retrieval (finding relationships between words in different languages), information visualization (presenting relationships between objects using 3-D graphics), user interface design/evaluation, and parallel computer languages.
Selected
Publications
In the area of automated theorem proving we are interested in the use of computers to discover the truth of proposed theorems. Although standard areas of mathematics are of special interest, proving theorems in any theory that can be formally defined is also of great interest. Currently, we are studying the Model Elimination procedure, a proof procedure for first-order logic that uses the implementation techniques of the logic programming language Prolog. The implementation is impressively fast but computes much redundant information. We seek to reduce this redundancy and add an automated lemma capability.
We have recently explored an extension to Prolog that permits an honest negation but offers much of the computational speed of Prolog potentially. The extension appears to be particularly effective when the input is nearly in Horn clause form, the latter being the format for Prolog.
We have also had an interest in the problem of updating knowledge bases where the new information being added may well interact with the existing information. Subtle interactions can occur because many facts or conjectures are inferred and not explicit in the knowledge base. This is part of our general interest in expert systems.
We have studied algorithms for determining near-optimal test procedures and test-and-treatment procedures. Algorithms for optimal results are too computationally expensive, so approximate algorithms are sought.
Selected
Publications
My research revolves around various types of complex systems and complex adaptive systems, using methods of statistical mechanics, stochastic processes, dynamical systems theory, graph theory, and computer simulation.
From my perspective, complex systems are systems with significant internal structure and memory, so that the external conditions do not adequately specify the internal state; this is natural in computer science but not (until recently) in statistical physics. My main interest under this heading concerns glasses, where I construct general models of highly constrained dynamical systems. The mathematical and computational problems chiefly concern diffusion on large random graphs and eigenanalysis of large random matrices.
Complex adaptive systems are complex systems that learn or adapt to their environment. I'm particularly interested in neural networks and genetic algorithms, and am currently working on a theory of genetic algorithms (and coevolution in general) based on statistical mechanics. I also apply learning algorithms to interacting-agent simulations of economic markets. More generally, I'm trying to construct mathematical theories of interacting- agent systems.
Selected
Publications