This paper presents an application of a connectionist control-learning system designed for use on an autonomous mini-robot. This system was formerly shown to form useful two-dimensional mappings rapidly when applied to backing a car with a single trailer. In the current paper the learning system is extended to three dimensions and applied to a similar but significantly more difficult problem. The system is shown to be capable of rapid unsupervised learning of output responses in temporal domains through the use of eligibility traces and inter-neural cooperation within topologically defined neighborhoods.

This paper presents the application of a connectionist control-learning system to an autonomous mini-robot. The system's design is severely constrained by the computing power and memory available on board the mini-robot and the on-board training time is greatly limited by the short life of the battery. The system is capable of rapid unsupervised learning of output responses in temporal domains through the use of eligibility traces and data sharing within topologically defined neighborhoods.

A classifier-based strategy to assist in the maintenance of expert systems is proposed. Substantiated claims about the results can be made by computing the error in probability of classification using established statistical techniques. Expert systems whose input, output, and intermediate state value(s) can be recorded as the values of a finite number of variables are considered. This class includes many expert systems including ETA and ESCA, expert systems developed by the program on surgical control of the hyperlipidemias (POSCH). Algorithmic approaches can provide automated assistance in helping tackle some issues of knowledge debugging. Some of these issues are mentioned. Details of using the algorithmic approach are covered in brief. Three types of algorithms that can be used have been identified. They are: classifier algorithms, similarity measures and generating algorithms.>

The image correspondence problem has generally been considered the most difficult step in both stereo and temporal vision. Most existing approaches match area features or linear features extracted from an image pair. The approach described in this paper is novel in that it uses an expert system shell to develop an image correspondence knowledge-based system for the general image correspondence problem. The knowledge it uses consists of both physical properties and spatial relationships of the edges and regions in images for every edge or region matching. A computation network is used to represent this knowledge. It allows the computation of the likelihood of matching two edges or regions with logical and heuristic operators. Heuristics for determining the correspondences between image features and the problem of handling missing information will be discussed. The values of the individual matching results are used to direct the traversal and pruning of the global matching process. The problem of parallelizing the entire process will be discussed. Experimental results on real-world images show that all matching edges and regions have been identified correctly.

This paper describes our recent results in developing enhanced simulated annealing algorithms using a LISP environment. The application is to use simulated annealing for automatic reconfiguration of multiprocessors in space. Our approach to solving this problem involves a combination of object-oriented programming, search strategies, knowledge based reasoning, and an advanced reconfiguration algorithm. The application was developed and is being enhanced on a LISP workstation (Xerox Dandelion) using LOOPS. This environment played an important role in both the initial success of the prototype [9] and the recent development efforts [3] [16] [22].

A knowledge representation used to capture support knowledge necessary for expert system justification, a kind of explanation, is discussed. A system is also presented that is being designed as a preliminary testbed for a technique of expert system explanation called knowledge-based explanation. Previous support knowledge representations in expert systems are reviewed and contrasted to the partitioned support network (PSN) introduced. PSNs are defined both in graphical and nongraphical notation. The significance and importance of the PSN knowledge representation are discussed and examples of its power are given.>

For pt.1 see ibid., p.33-42; for pt.2 see ibid., p.43-4. To show how neural nets work, experiences in an experiment using them are described. The experiment involves using AI techniques to assist in the discovery of causal relationships between the variables existing in a large clinical trial database. A peripheral vascular disease database was used to acquire a feeling for the complexities involved in developing a distributed encoding scheme and to determine the computational resources required to train a neural net for the type of data used. By testing several models the effects of changes in the encoding scheme and the number of training iterations the system needed to predict the appropriate change needed could be determined. These results were compared to the information available from other analyses of the same data. The generative capabilities of the system were then tested by training it over one sample of cases and applying it to cases it had not encountered before. Some idea of the computational resources needed in terms of time and memory capacity was developed.>

For pt.1 see ibid., p.33-42. A brief history of temporal reasoning in medical applications is given. An expert system called Monitor is described that can perform the following tasks: the incremental evaluation of the data; retrieval of historical data for comparison purposes; determination of the validity of the data cross references over time; assessment of the disease progression by abstracting trends and changes from the time-ordered data; and extraction of duration information from the symptom data. An initial evaluation of Monitor using 111 cases has shown that the system can classify 88% of the cases correctly.>

The authors document their experiences implementing expert systems in an operational environment. They cover the problems they encountered, the procedures used, and the solutions they found. Issues related to large databases are emphasized, as well as situations, problems, and solutions that are apt to be encountered.>

A generalized network-based expert system shell (AGNESS), which has eighteen explanation queries that put more information at the fingertips of the user, the expert and the knowledge engineer, is presented. The system is simple enough to be used in most of today's expert system shells. Other advantages include the use of a computation network to represent expert-defined rules, the ability to handle any well-defined inference method, the increased storage efficiency that comes from sharing constructs whenever possible, and the ability to use both numeric and nonnumeric data. AGNESS' menu-driven interface has several features that make the system friendly and convenient to use. Five expert systems generated by AGNESS are described.>

Expert consultant systems often perform computations on a directed graph of associated propositions. Each proposition is represented by a node. Edges connecting these nodes are associated with rules which organize the propositions into antecedent/consequent relationships. A node may be assigned a value through the edges that bind it to its antecedents. Various strategies are employed to determine assignment sequences that result in efficient computer consultation. One such strategy, the merit system, has been successfully implemented in Battle, an expert consultant system for the Marine Corps. The merit strategy enables Battle to focus the consultation process on the most appropriate questions. The merit system, originally defined for logical functions in the Multiple program, has been extended to the Mycin style of propagation and to the method of subjective Bayesian assignments used by Prospector. A procedure for merit calculations with any differentiable, real-valued assignment function is presented. Our experience has shown that merit values provide an efficient flow of control for expert consultation.

We computer scientists face at least two problems in promoting the use of computerized data-base systems: 1) some important data might be missing; 2) there might be errors in the data. Both of these problems can be quite serious. If they cannot be solved, it will be quite hard to convince potential users that computerized information systems are useful.

A method for the sequential mapping of points in a high-dimensional space onto a plane is presented. Whenever a new point is mapped, its distgnces to two points previously mapped are exactly preserved. On the resulting map, 2M -3 of the original distances can be exactly preserved. The mapping is based on the distances of a minimal spanning tree constructed from the points. All of the distances on the minimal spanning tree are exactly preserved.

To make further progress, resolution principle programs need to make better inferences and to make them faster. Previous papers of the authors have presented a fairly general approach for taking some advantage of the structure of special theories, for example, the theories of equality, partial ordering and sets, and described experiments with a program base on this approach. The object of the approach is to replace some or all of the axioms of the given theory by (refutation) complete, valid, efficient (in time) inference rules. In this paper, the approach is used to develop an improved procedure for ‘building in’ partial ordering and a procedure for total ordering. These results may be stated roughly as follows. 1. If the five (not all independent) partial ordering axioms for {=, ≤, <} are replaced by the irreflexivity rule ri and the transitivity rule rt (for <), by an expansion rule, and by an extension to hyper-resolution, then refutation completeness is preserved. 2. If only the connectivity axiom, {x < y ∨ y] ≤ x}, is retained from the five total ordering axioms for {=, ≤, <} and if the other four are replaced by ri, rt, and an antisymmetry rule, refutation completeness is preserved. A program using total ordering inference rules is then described. Differences between the rules as presented in the theoretical development and as implemented in the program are noted. The paper concludes with a discussion of the program's successful performance on a large collection of problems taken from published papers.

A clustering and data-reorganizing algorithm based on the concept of the shortest spanning path of a graph is given. This algorithm can be used to reorganize and/or cluster a large file of data.

To prove really difficult theorems, resolution principle programs need to make better inferences and to make them faster. An approach is presented for taking advantage of the structure of some special theories. These are theories with simplifiers, commutativity, and associativity, which are valuable concepts to build in, since they so frequently occur in important theories, for example, number theory (plus and times) and set theory (union and intersection). The object of the approach is to build in such concepts in a (refutation) complete, valid, efficient (in time) manner by means of a “natural” notation and/or new inference rules. Some of the many simplifiers that can be built in are axioms for (left and right) identities, inverses, and multiplication by zero. As for results, commutativity is built in by a straightforward modification to the unification (matching) algorithm. The results for simplifiers and associativity are more complicated. These theoretical results can be combined with one another and/or extended to either C -linear refutation completeness or theories with partial ordering, total ordering, or sets. How these results can serve as the basis of practical computer programs is discussed.

TO make further progress, resolution principle programs need to make better inferences and to make them faster.This paper presents a fairly general approach for taking some advantages of the structure of special theories, for example, the theories of equality, partial ordering, and sets.The object of the approach is to replace some of the axioms of a given theory by (refutation) complete, valid, efficient (in time) inference rules.The author believes that the new rules are efficient because: (1) inference for inference, a computer program embodying the rules should be faster than a resolution program computing with the corresponding axioms; (2) more importantly, certain troublesome inferences made by resolution are avoided by the new rules.In this paper, the three main applications of the approach concern "building-in" the theories of equality, partial ordering, and sets and may be stated roughly as follows.(1) If only {x =x} is retained from the equality axioms, and if the others are replaced by the functionally reflexive axioms and the rule of renamable paramodulation, refutation completeness is preserved.(2) If only {xC_x} is retained from the eight (not all independent) partial ordering axioms for { =, C:, C:} and if the other seven are replaced by the rules r(C::, C) and r(C:), refutation completeness is preserved.(3) If a certain seven of the twenty-four set axioms are retained and if the remaining seventeen are replaced by the rules r(C::, ~), r(C), r(C ), r(e) (complement), r( U ), r(['l ), and r(u) (unit sets), refutation completeness is preserved.

An inference system is C-linear complete if it is linear (ancestry filter) complete with top clause C, where C is in the original set of clauses and has suitable satisfiability properties.C-linear completeness is important for two reasons: (1) set-of-support refutation completeness is a corollary of C-linear refutation completeness, and previous computer experiments have indicated that the set-of-support strategy is efficient; (2) the search for a C-linear deduction can be naturally represented by a goal tree, and good techniques are known for searching such trees.A theorem is proved which provides a fairly general approach which, when given only a ground complete inference system, often yields a (nonground) C-linear complete system.This approach can be combined with a previously presented approach whose object is to replace some of the axioms of a given theory by a refutation complete system.The object of the combined approach is to replace some of the axioms by a C-linear refutation complete system.The approach of this paper is applied to six combinations of four inference rules.The rules are ordinary resolution, paramodulation, and two rules which respectively replace the transitivity axiom for C and the set membership axiom.C-linear refutation complete systems are found from the six combinations.For the case of resolution alone, the stronger C-linear deduction (consequence-finding) completeness is obtained.

Paramodulation can be used in conjunction with resolution for proving theoreIm in first-order logic with equality.Unit, input, and linear refutations using paramodulation and resolution are defined for theories with equality.It is proved that (a) if a set S of clauses has an input refutation, then S together with its unit factors and functionally reflexive axioms has a unit refutation; and (b) if C is a clause in an E-unsatisfiable set S of clauses including {x = x} and the functionally reflexive axioms and if (S -{C} ) is E-satisfiable, then S has a linear refutation with top clause C. ("E-unsatisfiable" is called "R-unsatisfiable" by some authors.)The refutation completeness theorem proved by Wos and Robinson for paramodulation with set of support is a corollary of our result (b).Our result (b) provides a link between heuristk programs and theorem-proving programs for first-order theories with equality. KEY WORDS AND PHRASES: theorem-proving,

The resolution principle is an inference rule for quantifier-free first-order predicate calculus. In the past, the completeness theorems for resolution and its refinements have been stated and proved for finite sets of clauses. It is easy (by Gödel's Compactness Theorem) and of practical interest to extend them to countable sets, thus allowing schemata representing denumerably many axioms. In addition, some theorems similar to Craig's Interpolation Theorem are proved for deduction by resolution. In propositional calculus, the theorem proved is stronger, whereas in predicate calculus the theorems proved are in some ways stronger and in some ways weaker than Craig's theorem. These interpolation theorems suggest procedures which could be embodied in computer programs for automatic proof finding and consequence finding.

This paper describes an algorithm which will generate all the prime implicants of a Boolean function. The algorithm is different from those previously given in the literature, and in many cases it is more efficient. It is proved that the algorithm will find all the prime implicants. The algorithm may possibly generate some nonprime implicants. However, using frequency orderings on literals, the experiments with the algorithm show that it usually generates very few ( possibly none) nonprime implicants. Furthermore, the algorithm may be used to find the minimal sums of a Boolean function. The algorithm is implemented by a computer program in the LISP language.

Many problems in artificial intelligence involve the searching of large trees of alternative possibilities—for example, game-playing and theorem-proving. The problem of efficiently searching large trees is discussed. A new method called “dynamic ordering” is described, and the older minimax and Alpha-Beta procedures are described for comparison purposes. Performance figures are given for six variations of the game of kalah. A quantity called “depth ratio” is derived which is a measure of the efficiency of a search procedure. A theoretical limit of efficiency is calculated and it is shown experimentally that the dynamic ordering procedure approaches that limit.

The heuristic program discussed searches for a constructive proof or disproof of a given proposition. It uses a search procedure which efficiently selects the seemingly best proposition to work on next. This program is multipurpose in that the domains it can handle are varied. As an initial experiment, the program was given the task of searching for proofs and disproofs of propositions about kalah end games. Kalah is a two-person game. In another experiment the program, after some modifications, played the game of kalah. This program was compared with another tree-searching procedure, the Alpha-Beta minimax procedure; the results have been encouraging since the program is fast and efficient. Its greatest usefulness is in solving large problems. It is hoped that this program has added one more step toward the goal of eventually obtaining computer programs which can solve intellectually difficult problems.

The theory of J. A. Robinson's resolution principle, an inference rule for first-order predicate calculus, is unified and extended. A theorem-proving computer program based on the new theory is proposed and the proposed semantic resolution program is compared with hyper-resolution and set-of-support resolution programs. Renamable and semantic resolution are defined and shown to be identical. Given a model M , semantic resolution is the resolution of a latent clash in which each “electron” is at least sometimes false under M ; the nucleus is at least sometimes true under M . The completeness theorem for semantic resolution and all previous completeness theorems for resolution (including ordinary, hyper-, and set-of-support resolution) can be derived from a slightly more general form of the following theorem. If U is a finite, truth-functionally unsatisfiable set of nonempty clauses and if M is a ground model, then there exists an unresolved maximal semantic clash [ E 1 , E 2 , · · ·, E q , C ] with nucleus C such that any set containing C and one or more of the electrons E 1 , E 2 , · · ·, E q is an unresolved semantic clash in U .

This paper is concerned with a class of procedures for making true-false decisions which depend on the outcome of eo sequence of elementary, binary tests.Certain of these procedures, calJed s-procedures, are conveniently represented by Boolean expressions (whose variables represent the elementary, bil~ary tests) and are easily carried out.If we are given the cost of applying each elementary, bina.rytest and the (a priori) prob'~bility of its outcome, the (average) cost of a procedure may be computed.Th.e genera!problem treated is to find efficiently a minimum-cost procedure equiv,~dent to a given s-procedure with known costs and probabilities for its elementt~ry, binary tests.Among the facts proved are two theorems showing that, under certain1 coi~ditions, the s-procedure found by a specified, efficient algorithm has mhfimum cost.The theorems and the etticient algorithm are used to solve illustrative problems in ti, e areas of computer programming, solving mat}mmatical problems, personnel selection, and medical diagnosis.

A large high-speed general-purpose digital computer (IBM 7090) was programmed to solve elementary symbolic integration problems at approximately the level of a good college freshman.The program is called SAINT, an acronym for "Symbolic Automatic INTegrator."This paper discusses the SAINT program and its performance.SAINT performs indefinite integration.It also performs definite and multiple integration when these are trivial extensions of indefinite integration.It uses many of the methods and heuristics of students attacking the same problems.SAINT took an average of two minutes each to solve 52 of the 54 attempted problems taken from the Massachusetts Institute of Technology freshman calculus final examinations.Based on this and other experiments with SAINT, some conclusions concerning computer solution of such problems are: (1) Pattern recognition is of fundamental importance.(2) Great benefit would have been derived from a larger memory and more convenient symbol manipulating facilities.(3) The solution of a symbolic integration problem by a commercially available computer is far cheaper and faster than by man.