a report by Ralph Kenyon EXTRAPOLATOR Aug 25, 1987
This report provides an overview of artificial intelligence. It describes the state of the field as of July 1987 and explains what the term really means. It covers the sources, the organization of the field into sub-fields, the limitations and problems, successful applications, exciting new directions, and projects likely near-term applications. Included is a description of topics in sub-fields within the field of artificial intelligence.
© Copyright 1987 by Ralph E. Kenyon, Jr.
With degrees in Mathematics, Management, Computer Science, and Philosophy, and experience in Intelligence Analysis, Physical and Personnel Security, Nuclear Power, Marine Engineering, Contract Administration, and Systems Design and Programming, Ralph Kenyon is uniquely qualified to provide a broad synthesis across diverse fields.
The term 'artificial intelligence' doesn't mean, to those who use it most, what one might expect from combining the two separate words 'artificial' and 'intelligence'. As a field of academic and economic focus, artificial intelligence (AI) has several sub-fields. (Which also don't necessarily mean what one might infer from their terms).
The earliest sources of artificial intelligence derive from four areas. Efforts to automate mathematical theorem proving flows from the philosophical Logicist tradition. Attempt to translate natural languages derives from military intelligence needs Vision systems and robotics arise from industrial frugality. Learning systems arise in education.
The field has become diverse, but there remains a focus on six areas.
The sixth area, expert systems, arises more recently from efforts to devise more sophisticated computer programs for dealing with specialized areas of knowledge.
Most recently on the scene, and not yet forming its own niche, is the neural net genre. Many are viewing this technique (hardware or simulated in serial software) as a panacea which will solve previously intractable problems. Relaxing a network seems to overcome the local maximum problem and allows solving the traveling salesman problem (np-complete). However, this initial mystique is un-justified.
There has arisen some tension between the foci as they have come into being and the somewhat more flashy interpretations of the 'science fiction genre'. Much research is driven by individuals who are more sympathetic to these more fanciful wishes.
The reality of current (1987) progress is much less optimistic. People doing work in AI fall into a few specific categories and are working on rather limited areas. However, media hype has created a demand for artificial intelligence, most often without a clear idea of what it is that is desired. The two most salable products available are expert system shells, which allow a capturing some specific types of expertise, and industrial robots.
Heavy research in vision systems will pay dividends in the form of improved versatility of industrial robots. In addition, these systems will make possible more general purpose types of robots.
Combining the improved sensor capabilities with expert systems and general purpose arms will make possible a generation of robots for sorting and selecting. Applications will include fruit & vegetable sorting and grading, parts selection and sorting, trash separation for recycling, and other kinds of simple sorting and selecting which must currently be done by human beings.
We have already seen the flurry of activity in applying expert systems in a diagnostic capacity. Another product area will be in interfacing expert systems to large data bases of specialized as well as general information, including libraries. A natural adjunct to these expert system front ends on data bases is their use for training new users of the data bases as well as in preliminary training of new experts. How long will it be before someone puts salesmanship expertise into an expert system front end on a large data base of consumer products? -- The unwary inquirer would find himself maneuvered into a buying position.
Dreamed about capabilities include common sense conversations in natural language using voice input/output. However, this goal is a lot further away than most people would like to admit. Research at Bell Labs is showing that the spoken language generation problem has many more levels of structure than heretofore imagined. As more and more level of structure emerge, the question of how many more levels will eventually emerge arises. And, spoken language recognition is even more intractable than its generation.
The main problems that need to be solved are an explication of just what are "knowledge", "intelligence", "understanding", and "meaning". We can't create artificial intelligence unless we can know or decide what natural intelligence is. Similarly, we cannot create machines or programs which "understand" unless we know what understanding entails.
The philosophers haven't been able to characterize "knowledge", "intelligence", or "understanding" in 2500 years. We need to "de-mystify" these terms. What is required is a paradigm shift which entails our thinking of "knowledge", "intelligence", "understanding", etc., not as "things", but as classifications into which many "things" fit. We do not use these terms univocally. Until we can shift our focus to the level in which the many different things or processes under which we use these terms are differentiated, we won't be able to say what they are.
Marvin Minski's new book "The Society of Mind" presents a new model of intelligence. He proposes a bureaucracy model with agents and agencies all interconnected. This distributed form of intelligence allows many different processes to act from a central agency. If we allow intelligence as a term which encompasses many different processes, then such a model fits well. But, we are on a wild homunculus chase when we search for one thing to explain intelligence.
There is need for an overall 'big picture' organizing scheme for activities focusing on intelligence and artificial intelligence. My six stage model of information processing provides the basic paradigm. Adapting that model to the state artificial intelligence is in requires providing a characterization of intelligence. First, it is not a single 'thing'; intelligence is at least an information handling process which has sub-processes. As such, it cannot be said to be a thing.
The level of structure in which intelligence functions can be characterized by referring to sub-processes. Intelligence inheres in the processes of sensing, abstracting, representing, planning, deciding, and acting, all in order to resolve motivations. The AAAI-87 conference provided a 13 way classification scheme for AI efforts. The thirteen AAAI classifications (in which there is considerable overlapping) relate to the basic sub-processes in intelligence as follows.
Planning maps representations to potential actions in order to resolve motivations. If each of the sub-processes is thought of as a central region with nebulous extensions extending into the other processes and influencing their structure and function, it can be seen that developments in any one area significantly impact the other areas. The amorphous nature of this classification suggests that it will be some time before intelligence is fully realized in artificial systems. A viable theory will be heralded by the arrival of a classification scheme which is less intertwined. It is partly a matter of complexity and of finding a lower level classification scheme which provides greater structural definition. We need a "periodic table" of "elements" that combine in compound, but regular ways, to form the polymer "intelligence", and, we need to stop asking "but which one is the rubber atom?" To change the metaphor, we are mapping out much of the geography of the country of intelligence, but we don't equate the country to one of its provinces. We need to discover the infrastructure of the country, and note that the country accomplishes things by using and modifying its infrastructure.
Where we go from here will depend upon what are seen as possibilities, and that depends upon the paradigm implicit in our metaphysics. We have long been driven by the search for "essences", which when applied to intelligence leads directly to the Cartesian homunculus and infinite regression. It is clear that we need to look for structure and function, and not equate intelligence with any particular sub-structure.
One of the latest emerging views in philosophy, evolutionary epistemology, would have it that there really is no intelligence per-se. The mechanism of evolutionary epistemology is quite simple; it includes variation and selective retention. Loosely, this means that different methods are tried and the ones that work are retained, while those that do not work are discarded from the repertoire. Darwin's revolution applies not only to the evolution of the species, and to the growth of scientific ideas, but also to the very workings of our individual minds. We "brainstorm" to come up with different ideas (variation) and keep those that seem to work (selective retention). Of course, some of those methods which have been retained for long periods of time include classical logic, set theory, law, etc. We can think of intelligence as an eclectic combination of methods that have worked in the past. The combination is dynamic in that new methods are being added and old ones are occasionally being dropped. Part of intelligence includes as yet unspecified methods for deciding which methods map to which problems in which contexts for what purposes.
The really exciting items include work on a visual graphics editor, a robot navigation system that builds internal maps based upon seen landmarks, and implementing Piagetian style learning.
The thirteen classifications of AAAI 87 and their topics.
Butterfly Lisp; Concurrent Common Lisp; parallel resolution based on connection graph; generalized Blackboard; forward chaining logic; goal directed reasoning in blackboards; multiprocessor parallel production system matching; conflict set support in production system matching; syntactically uniform access to heterogeneous knowledge bases; concurrent, controllable constraint systems; non-deterministic LISP with dependency-directed backtracking.
AI & Education:
Intelligent tutoring with misconception models applied to psychotherapy; intelligent tutoring applied to satellite ground tracks; student modeling as plan inferencing; need for community memory for multiple experts.
Heuristic evaluation function for two player games; syntactic analogies between proofs with second order pattern matching; comparing minimax and product backup rules; removing redundancy in constraint networks; cost-benefit assessment in blackboard environments; interpreting sensor data with probalistic type constraints; rule-based systems limited in uncertain reasoning; inferring formal software specifications from episodic descriptions; real-time heuristic search; reasoning with inconsistencies; structural induction with mutual recursion; synthesizing algorithms with performance constraints; imperative lisp program synthesis from specifications; path dissolution is a strongly complete rule of inference; revised dependency directed backtracking for default reasoning; efficiency analysis of multiple-context TMSs in scene representation; a parallel implementation of iterative deepening A*;clause management system in foundations of assumption based truth maintenance systems.
Reasoning about exceptions during plan execution monitoring; a polynomial time algorithm for incremental causal reasoning in partial ordered events; reactive action planning in complex domains; stratified autoepistemic theories; possible worlds and quantification; simple causal minimizations for temporal persistence and projection; using goal interaction to guide planning; plan operators from qualitative process theory; axioms for time intervals; localized representations and planning methods for parallel domains; a model for concurrent actions having temporal extent; consistent labeling problem in temporal reasoning; the satisfiability of temporal constraints networks; validating generalized plans with incomplete information.
Implementing a theory of activity; compare and contrast in legal reasoning; autoassociative module learning in neural networks; reducing indeterminism in modeling user librarian interaction; a mechanism for early Piagetian learning; rules for implicit acquisition of knowledge about users; neural net approach to case based problem solving with a large knowledge base; neural network connectivity and propagation applied to materials handling; asking questions to understand answers; information retrieval below document level; structure mapping in analogical processing; goal based generation of motivational expressions in a learning environment.
Incremental inference in a mixed initiative environment; augmenting first order logic to implement default reasoning; annealing in connectionist nets to implement counterfactual reasoning; inheritance hierarchies with exception; semantic networks with multiple inheritance and exceptions; a formalism describing circumscription policies by axioms included with the knowledge base axioms; causality in formal reasoning; representing dependencies by directed graphs; default reasoning by belief revision; default reasoning by partially ordered theories.
Goal/subgoal plan representation for real-time process monitoring; partial compilation of declarative knowledge into procedures; representing databases in frames; Intension as choice plus commitment; manipulating knowledge as taxonomic representation schemes; complexity in classificatory reasoning; a logic of belief, with semantics, for non-monotonic reasoning; algorithm synthesis through problem reformulation using generic designs as parameterized theories; using truth maintenance systems to cure anomalous extensions in non-monotic logics; semantically sound inheritance for a formally defined frame language with defaults; assimilation as a strategy for implementing self-reorganizing knowledge bases;
Machine Learning & Knowledge Acquisition:
Learning to control a dynamic physical system; improved inference through conceptual clustering; learning conjunctive concepts in structural domains; comparing knowledge engineering to decision analysis; formulating concepts according to purpose in explanation-based learning; defining operationally for explanation-based learning; interactive expert system generation using general knowledge about evaluation tasks; using explanation-based generalization to implement a prolog interpreter that learns; knowledge level learning by acquiring general procedures from goal-based experience; inductive concept learning by reasoning with declarative formulation of biases; dynamic acquisition of appropriate representations (concept generalization) to minimize initial representational bias; generalizing to the Nth case in explanation-based reasoning; optimizing prediction in diagnostic decision rules.
Interpreting clues in restricting arguments and discourse; principle-based description of grammar in machine translation; recovering from erroneous inferences; control strategies for achieving pragmatic goals in language generation by interpreting inputs; word-order variation in natural language generation; porting an extensible natural language interface; inference (interpretation?) in text understanding; acquisition of conceptual structures for the lexicon; procrastination in resolving ambiguity; memory-based reasoning applied to pronunciation; nondestructive graph unification.
Engineering Problem Solving:
Reasoning about fluids via molecular collections; extending the mathematics of qualitative process theory; using order of magnitude reasoning for troubleshooting complex analog circuits; explanation based failure recovery in systems with partially compiled operators; generating function from shape and motion constraints on parts in mechanical devices; establishing critical hyper surfaces in the parameter product space for physical processes; time-scale abstraction: a method for structuring a complex system as a hierarchy of smaller, interacting equilibrium mechanisms; formalizing reasoning with orders of magnitude and approximate relations; making partial choices in constraint reasoning problems; multi-level resolution of constraint propagation failures by prototype modification using physical knowledge in a graph of models; reasoning about discontinuous change; a system for hierarchical reasoning about inequalities; analyzing dynamic systems describable by finite sets of ordinary differential equations; probabilistic semantics for qualitative influences; extracting qualitative dynamics from numerical experiments using phase space.
Intelligent task automation integrating task planning, path planning, vision and robotics for performing autonomous manufacturing tasks in dynamic, unstructured environments; reactive reasoning and planning in autonomous mobile robots including belief, desire, and intension; graphics editor using visual grammars for visual languages; space representation and use by landmark-based path planning and following; insertions using geometric analysis and hybrid force-position control.
Sensitivity of motion and structure computation; linking high and low level image understanding techniques for shape extraction using generic geometric models; developing general techniques for automated mapping and photo interpretation tasks; hypothesis testing in a computational theory of visual word recognition; integrating multiple shape-from-texture algorithms; similitude-invariant pattern recognition using parallel distributed processing; range image interpretation of mail pieces with superquadrics; closed form solution to the structure-from-motion problem from line correspondences; Bounds on translational and angular velocity components from first order derivatives of image flow; regularization of visual data using fractal priors in Bayesian modeling; energy constraints on deformable models in recovering shape and non-rigid motions; locating object boundaries using shadows; color separation by perceptual significance hierarchy; detecting, tracking, and locating 3-D line segments in a mobile robot vision system.
Data validation during diagnosis using expectations; incomplete knowledge (prospective) reasoning; learning by deriving symptom-fault associations in diagnostic environments; planning machining process for numerical control cutting machines from drawings; multiple representation approach to understanding the time behavior of digital circuits; integrating multiple expert systems and databases in computer aided engineering; automated reasoning for providing real time advice in process operations; TEST an application shell that provides a domain-independent diagnostic problem solver with a library of schematic prototypes; script-based reasoning for situation monitoring of complex activity; coping with the problems of a very large rule-base; design as top-down refinement plus constraint propagation.
|This page was updated by Ralph Kenyon on 2009/11/16 at 21:21 and has been accessed 8634 times at 57 hits per month.|