Current explanation-based learning systems acquire plans using a model of the world that is assumed to be perfect. Such systems perform well in executing these acquired plans in their perfect simulated world, but encounter difficulties in the real world. This is because current systems implicitly assume their model to be correct when it is actually impossible to model the real world perfectly. Our approach, embodied in the GRASPER system, is to make explicit approximations that the system can refine when failures are encountered. GRASPER makes use of its explicit approximation capability to represent uncertainties about objects in the world. This enables it to acquire plans capable of functioning in spite of uncertainties.

In real-world domains, large amounts to knowledge are needed to adequately describe world behavior. With a complex domain theory, complete reasoning becomes a computationally intractable task. This is particularly true of knowledge-intensive applications such as explanation-based learning, planning, and expert systems. This project involves development of an approach to problem-solving and learning designed to deal with computationally ill-behaved domains. By using simplifications, the system avoids the intractability of complete reasoning. However, this introduces the possibility of learning imperfect plans. These plans are refined by detecting and analyzing failures. By using the real world to focus attention on incorrect simplifications, the system avoids the intractability of complete reasoning.

Given a formal specification for a desired program, our goal is to transform it step-by-step into executable code. We are investigating the use of derivatives of the Knuth-Bendix completion procedure, designed for equational theorem proving, to synthesize functional and logic programs from specifications and domain knowledge expressed as rewrite rules, using both deductive and inductive reasoning.

While software reuse promises improved programmer productivity, greater code quality, and reduced developmental costs, it has been notoriously difficult to achieve under past paradigms. This research investigates the hypothesis that software reuse benefits are better realized at higher levels of abstraction than is typically attempted, and that automated analogical reasoning, applied to knowledge-based domain-specific models, can greatly facilitate the reuse process. This work is being conducted within the context of the development of intelligent design assistance for the ISLE programming environment, and in conjunction with the Advanced Collaborative Systems Laboratory (ACSL).

This project aims to build a distributed knowledge acquisition system by which domain experts can independently contribute to the gradual construction of a communal knowledge base. The knowledge base contains entities, such as subdomains, concepts and variables, as well as relations between these entities. The system allows experts to enter information concerning a domain-related topic and to determine how this information affects the existing knowledge structures. We are now studying the use of such theories in proving, completing, or modifying user-specified knowledge segments. We are also developing a new knowledge representation language and domain modeling tools and techniques.

Distributed problem solving has been widely used to provide intelligent support for cooperative activities. While a distinguished feature of hybrid reasoning systems is the use of multiple specialized subreasoners, they generally assume a centralized model of computation and rely on single-agent logics. We believe that the basic ideas of hybrid reasoning can be successfully extended to both distributed problem solving and multiagent logics. We will be developing a hybrid reasoning framework for problem solving in multiagent cooperative systems. This work is being pursued in conjunction with the Advanced Collaborative Systems Laboratory (ACSL).

This research deals with issues of distributed problem solving. In particular, we are looking for mechanisms for providing intelligent support for problem solving. This support could be in the form of reasoning mechanisms for handling interdependencies among tasks and interactions among agents; or it could also be in the form of providing scenario analysis for decision making. We are developing a framework for representation of all relevant aspects of a problem-solving process. The framework provides a language composed of three primitives: agents, actions, and artifacts. This work is being pursued in conjunction with the Advanced Collaborative Systems Laboratory (ACSL).

The knowledge-based programming assistant (KBPA) project has been researching, designing, and constructing the elements of an intelligent programming support system capable of aiding programmers in various facets of program production, such as specification, design, coding, debugging, and testing. In this system, techniques of computer inference, general problem solving, planning, and data management are applied to a rich database of knowledge about different phases of program production. Currently we are studying issues of program synthesis and analogical programming. This work is part of the research being conducted in the Advanced Collaborative Research Laboratory (ACSL).

In the CADRE project, we are exploring ways of capturing the knowledge of expert designers in a computer-aided design system. This knowledge will be used to guide and train less experienced designers. In particular, CADRE provides design critiques for low-slope roofing designs. CADRE's power comes from its problem goal model, which captures the expert's subtask decomposition and uses it to guide the user and the software in jointly creating a design. The user goal model guides inexperienced designers in understanding what tasks need to be addressed next.

Designers not only need task guidance and criticism, they also need constructive suggestions as to how to make the design better. In the Manufacturing Design Adviser project, we capture the ways in which human manufacturing engineers produce redesign suggestions for making products less costly to manufacture. The difficulty in producing effective design suggestions is that the relationship between the design and the cost is not simple. We solve this problem by using a plan-based method to model these relationships. This tool offers an important mechanism for helping designers to quickly produce more cost-effective designs.

A large problem in design and manufacturing research is how to pass information between experts. We are constructing a manufacturing feature extraction program which will automatically extract data representations from a designer's product specification in a form that can be understood by an automated manufacturing planner. We are experimenting with ways of creating a process-based feature extractor, which relies on an understanding of capabilities of the manufacturing processes. We believe that this process-based feature extractor will ultimately provide a more robust and flexible tool than current technology.

One difficulty with existing AI planning techniques is that they have not been demonstrated to scale up well for solving problems of any practical size or complexity. In the LCOS project, we are studying problems from a variety of domains in order to understand what properties planners need to have to solve complex problems. We make use of protocol analysis to build cognitive models of planning. We have formulated a variety of methods that have been demonstrated to produce high-quality plans for complex domains found in manufacturing problems.

We are studying how information infrastructure can facilitate a community of design, manufacturing, and evaluation agents who are jointly contributing to an evolving design. The advantage of this approach is that it allows design problems to be worked out early in the design process, when financial commitment to a particular design is still low. This makes it much easier to produce products that have all around better quality. The goal is to understand what information exchange needs to exist between agents in a design community and to set up a cooperative problem-solving framework.

We propose to represent arbitrary smooth 3-D shapes by a discrete set of HOT curves, where a surface admits High Order Tangents. These curves have two remarkable properties: they determine the structure of image contours and its catastrophic changes, and they project onto contour such features as inflections and bitangents. We are investigating the application of HOT curves to three fundamental problems in model-based vision: automatic object model construction from sequences of range and video images, aspect graph construction from discrete sets of video images, and object recognition using viewpoint-dependent monocular image features as table indexes in large databases of object models.

This project is concerned with planning stable (force-closure) grasps of objects by dextrous robotics hands. We consider the case of a hand equipped with two, three, or more hard fingers and assume point contact with friction. We have developed and implemented algorithms for planning maximal three-finger grasps of polygonal 2-D objects and four-finger grasps of polyhedral 3-D objects. These algorithms rely on linear programming and computational geometry techniques. We are currently extending our approach to the computation of optimal grasps.

We are developing a vision system combining geometric and probabilistic reasoning to interpret video images of polyhedral scenes. The first component of this system is aimed at reconstructing a three-dimensional scene from the geometric (line-drawing) and photometric (intensity) information contained in the image. The second component of our system is aimed at recognizing the polyhedral models that are present in the image. It relies on new probabilistic models of the distribution of edge lengths and angles in the two-dimensional projection of a three-dimensional scene.

We propose a hybrid method using both geometric and intensity information to reconstruct multiple curved object boundaries from sparse sets of x-ray images. We have constructed a preliminary implementation of this approach and conducted experiments using both synthetic and real x-ray images. Applications of this work are in medical imaging (reconstruction of organs and tumors from sparse sets of x-ray images, stereo reconstruction of coronary arteries from angiograms, registration of CT models and x-ray images) and in manufacturing (part acquisition and tolerancing, registration, and nondestructive evaluation).

We are continuing study of various issues relevant to design of a very general-purpose neural-based front-end processor for multichannel signals. The objective is to provide a combination of ``prewired,'' self-organizing, and distally supervised learning capable of learning the low-level aspects of any task in signal interpretation. At present, we are studying the merits of a system combining wavelet transform, self-organizing feature maps, an original temporal sequence processor, and an ARTMAP to develop associative relationships between a higher level expert/symbolic system and the proposed neurally based front end.

Because it allows prediction of future events, inductive learning is a fundamental problem of artificial intelligence. When domains are uncertain, incremental, and badly behaved, two important problems are variable bias (the dynamic choice of inductive constraints) and constructive induction (the formation of new knowledge structures). Like the variable-bias system, our approach to constructive induction also manages uncertainty and dynamic environments. Current work involves the application of knowledge, which may be partial, inaccurate, or conflicting. When knowledge is available, our systems refine it and simultaneously improve induction; when no knowledge is available, induction constructs it.

In applications such as computer-aided design, users would benefit from a truly intelligent aid that would help with various tasks, among them providing relevant advice about specific standards and techniques for CAD. To suggest anything valuable to a user, a computer system needs to know the user's intent, background, and perhaps even some aspects of the user's personality; in other words, an adaptive workstation would grow to ``understand'' the user, acting as an intelligent assistant by monitoring user performance, proposing design alternatives, and offering assistance when (and only when) needed. Our research goal is to mechanize the ability to adapt to individual users.