yorick@nmsu.CSNET.UUCP (05/11/87)
ABSTRACTS OF MEMORANDA IN COMPUTER AND COGNITIVE SCIENCE Computing Research Laboratory New Mexico State University Box 30001 Las, Cruces, NM 88003. Kamat, S.J. (1985), Value Function Approach to Multiple Sensor Integration, MCCS-85-16. A value function approach is being tried for integrating multiple sensors in a robot environment with known objects. The state of the environment is characterized by some key parameters which affect the performance of the sensors. Initially, only a handful of discrete environmental states will be used. The value of a sensor or a group of sensors is defined as a function of the number of possible object contenders under consideration and the number of contenders that can be rejected after using the sensor information. Each possible environmental state will have its effect on the function, and the function could be redefined to indicate changes in the sampling frequency and/or resolution for the sensors. A theorem prover will be applied to the sensor information available to reject any contenders. The rules used by the theorem prover may be different for each sensors, and the integration is provided by the common decision domain. The values for the different sensor groups will be stored in a database. The order of use of the sensor groups will be according to the values, and can be stored as the best search path. The information in the database can be adaptively updated to provide a training methodology for this approach. Cohen, M. (1985), Design of a New Medium for Volume Holographic Information Processing, MCCS-85-17. An optical analog of the neural networks involved in sensory processing consists of a dispersive medium with gain in a narrow band of wavenumbers, cubic saturation, and a memory nonlinearity that may imprint multiplexed volume holographic gratings. Coupled mode equations are derived for the time evolution of a wave scattered off these gratings; eigenmodes of the coupling matrix $$kappa$$ saturate preferentially, implementing stable reconstruction of a stored memory from partial input and associative reconstruction of a set of stored memories. Multiple scattering in the volume reconstructs cycles of associations that compete for saturation. Input of a new pattern switches all the energy into the cycle containing a representative of that pattern; the system thus acts as an abstract categorizer with multiple basins of stability. The advantages that an imprintable medium with gain biased near the critical point has over either the holographic or the adaptive matrix associative paradigms are (1) images may be input as non-coherent distributions which nucleate long range critical modes within the medium, and (2) the interaction matrix $$kappa$$ of critical modes is full, thus implementing the sort of `full connectivity' needed for associative reconstruction in a physical medium that is only locally connected, such as a nonlinear crystal. Uhr, L. (1985), Massively Parallel Multi-Computer Hardware = Software Structures for Learning, MCCS-85-19. Suggestions are made concerning the building and use of appropriately structured hardware/software multi-computers for exploring ways that intelligent systems can evolve, learn and grow. Several issues are addressed such as: what computers are, the great variety of topologies that can be used to join large numbers of computers together into massively parallel multi-computer networks, and the great sizes that the micro-electronic VLSI (``very large scale integration'') technologies of today and tomorrow make feasible. Finally, several multi-computer structures that appear especially appropriate as the substrate for systems that evolve, learn and grow are described, and a sketch of a system of this sort is begun. Partridge, D. (1985), Input-Expectation Discrepancy Reduction: A Ubiquitous Mechanism, MCCS-85-24. The various manifestations of input-expectation discrepancy that occurs in a broad spectrum of research on intelligent behavior is examined. The point is made that each of the different research activities highlights different aspects of an input-expectation reduction mechanism and neglects others. A comprehensive view of this mechanism has been constructed and applied in the design of a cognitive industrial robot. The mechanism is explained as both a key for machine learning strategies, and a guide for the selection of appropriate memory structures to support intelligent behavior. Ortony, A., Clore, G. & Foss, M. A. (1985), Conditions of Mind, MCCS-85-27. A set of approximately 500 words taken from the literature on emotion was examined. The overall goal was to develop a comprehensive taxonomy of the affective lexicon, with special attention being devoted to the isolation of terms that refer to emotions. Within the taxonomy we propose, the best examples of emotion terms appear to be those that (a) refer to [i]internal, mental[xi] conditions as opposed to physical or external ones, (b) are clear cases of [i]states[xi], and (c) have [i]affect[xi] as opposed to behavior or cognition as their predominant referential focus. Relaxing one or another of these constraints yields poorer examples or nonexamples of emotions; however, this gradedness is not taken as evidence that emotions necessarily defy classical definition. Wilks, Y. (1985), Machine Translation and Artificial Intelligence: Issues and their Histories, MCCS-85-29. The paper reviews the historical relations, and future prospects for relationships, between artificial intelligence and machine translation. The argument of the paper is that machine translation is much more tightly bound into the history of artificial intelligence than many realize (the MT origin of Prolog is only the most striking example of that), and that it remains, not a peripheral, but a crucial task on the AI agenda. Coombs, M.J. (1986), Artificial Intelligence Foundations for a Cognitive Technology: Towards The Co-operative Control of Machines, MCCS-85-45. The value of knowledge-based expert systems for aiding the control of physical and mechanical processes is not firmly established. However, with experience, serious weaknesses have become evident which, for solution, require a new approach to system architecture. The approach proposed in this paper is based on the direct manipulation of models in the control domain. This contrasts with the formal syntactic reasoning methods more conventionally employed. Following from work on the simulation of qualitative human reasoning, this method has potential for implementing truly co-operative human/computer interaction. Coombs, M.J., Hartley, R. & Stell J.F. (1986), Debugging User Conceptions of Interpretation Processes, MCCS-85-46. The use of high level declarative languages has been advocated since they allow problems to be expressed in terms of their domain facts, leaving details of execution to the language interpreter. While this is a significant advantage, it is frequently difficult to learn the procedural constraints imposed by the interpreter. Thus, declarative failures may arise from misunderstanding the implicit procedural content of a program. This paper argues for a \fIconstructive\fR approach to identifying poor understanding of procedural interpretation, and presents a prototype diagnostic system for Prolog. Error modelling is based on the notion of a modular interpreter, misconceptions being seen as modifications of correct procedures. A trace language, based on conceptual analysis of a novice view of Prolog, is used by both the user to describe his conception of execution, and the system to display the actual execution process. A comparison between traces enables the the correct interpreter to be modified in a manner which progressively corresponds to the user's mental interpreter. Dorfman, S.B. & Wilks, Y. (1986), SHAGRIN: A Natural Language Graphics Package Interface, MCCS-85-48. It is a standard problem in applied AI to construct a front-end to some formal data base with the user's input as near English as possible. SHAGRIN is a natural language interface to a computer graphics package. In constructing SHAGRIN, we have chosen some non-standard goals: (1) SHAGRIN is just one of a range of front-ends that we are fitting to the same formal back-end. (2) We have chosen not a data base in the standard sense, but a graphics package language, a command language for controlling the production of graphs on a screen. Parser output is used to generate graphics world commands which then produce graphics PACKAGE commands. A four-component context mechanism incorporates pragmatics into the graphics system as well as actively aids in the maintenance of the state of the graph world. Manthey, M.J. (1986), Hierarchy in Sequential and Concurrent Systems or What's in a Reply, MCCS-85-51. The notion of hierarchy as a tool for controlling conceptual complexity is justifiably well entrenched in computing in general, but our collective experience is almost entirely in the realm of sequential programs. In this paper we focus on exactly what the hierarchy-defining relation should be to be useful in the realm of concurrent programming. We find traditional functional dependency hierarchies to be wanting in this context, and propose an alternative based on shared resources. Finally we discuss some historical and philosophical parallels which seem to have gone largely unnoticed in the computing literature. Huang, X-M (1986), A Bidirectional Chinese Grammar in A Machine Translation System, MCCS-85-52. The paper describes a Chinese grammar which can be run bidirectionally, ie., both as a parser and as a generator of Chinese sentences. When used as a parser, the input to the grammar is single Chinese sentences, and the output would be tree structures for the sentences; when used as a generator, tree structures are the input, and Chinese sentences, the output. The main body of the grammar, the way bidirectionality is achieved, and the performance of the system with some example sentences are given in the paper. Partridge, D. & Wilks, Y. (1986), Does AI have a methodology different from Software Engineering?, MCCS-85-53. The paper argues that the conventional methodology of software engineering is inappropriate to AI, but that the failure of many in AI to see this is producing a Kuhnian paradigm ``crisis''. The key point is that classic software engineering methodology (which we call SPIV: Specify-Prove-Implement-Verify) requires that the problem be circumscribable or surveyable in a way that it is not for areas of AI like natural language processing. In addition, it also requires that a program be open to formal proof of correctness. We contrast this methodology with a weaker form SAT ( complete Specification And Testability - where the last term is used in a strong sense: every execution of the program gives decidably correct/incorrect results) which captures both the essence of SPIV and the key assumptions in practical software engineering. We argue that failure to recognize the inapplicability of the SAT methodology to areas of AI has prevented development of a disciplined methodology (unique to AI, which we call RUDE: Run-Understand-Debug-Edit) that will accommodate the peculiarities of AI and also yield robust, reliable, comprehensible, and hence maintainable AI software. Slator, B.M., Conley, W. & Anderson, M.P (1986), Towards an Adaptive Front-end, MCCS-85-54. An adaptive natual language interface to a graphics package has been implemented. A mechanism for modelling user behavior operating over a script-like decision matrix capturing co-occurrence of commands is used to direct the interface, which uses a semantic parser, when ambiguous utterances are encountered. This is an adaptive mechanism that forms a model of a user's tendencies by observing the user in action. This mechanism provides a method for operating under conditions of uncertainty, and it adds power to the interface - but, being a probabilistic control scheme, it also adds a corresponding element of nondeterminism. A hidden operator experiment was conducted to collect utterance files for a user-derived interface development process. These empirical data were used to design the interface; and a second set, collected later, was used as test data. Lopez, P., Johnston, V. & Partridge, D. (1986), Automatic Calibration of the Geometric Workspace of an Intelligent Robot, MCCS-85-55. An intelligent robot consisting of an arm, a single camera, and a computer, functioning in an industrial environment, is described. A variety of software algorithms that compute and maintain, at task-execution time, the mappings between robot arm, work environment (the robot's world), and camera coordinate systems, are presented. These mappings are derived through a sequence of arm movements and subsequent image ``snapshots'', from which arm motion is detected. With the aid of world self-knowledge (i.e., knowledge of the length of the robot arm and the height of the arm to the base pivot), the robot then uses its ``eye'' to calculate a pixel-to-millimeter ratio in two known planes. By ``looking'' at its arm at two different heights, it geometrically computes the distance of the camera from the arm, hence deriving the mapping from the camera to the work environment. Similarly, the calculation of the intersection of two arm positions (where wrist location and hypothetical base location form a line) gives a base pivot position. With the aid of a perspective projection, now possible since the camera position is known, the position of the base and its planar angle of rotation in the work environment (hence the world to arm mapping) is determined. Once the mappings are known, the robot may begin its task, updating the approximate camera and base pivot positions with appropriate data obtained from task-object manipulations. These world model parameters are likely to remain static throughout the execution of a task, and as time passes, the old information receives more weight than new information when updating is performed. In this manner, the robot first calibrates the geometry of its workspace with sufficient accuracy to allow operation using perspective projection, with performance ``fine-tuned'' to the nuances of a particular work environment through adaptive control algorithms. Fass, D. (1986), Collative Semantics: An Approach to Coherence, MCCS-85-56. Collative Semantics (CS) is a domain-independent semantics for natural language processing that focusses on the problem of coherence. Coherence is the synergism of knowledge (synergism is the interaction of two or more discrete agencies to achieve an effect of which none is individually capable) and plays a substantial role in cognition. The representation of coherence is distinguished from the representation of knowledge and some theoretical connections are established between them. A type of coherence representation has been developed in CS called the semantic vector. Semantic vectors represent the synergistic interaction of knowledge from diverse sources (including the context) that comprise semantic relations. Six types of semantic relation are discriminated and represented: literal, metaphorical, anomalous, novel, inconsistent and redundant. The knowledge description scheme in CS is the senseframe, which represents lexical ambiguity. The semantic primitives in senseframes are word-senses which are a subset of the word-senses in natural language. Because these primitives are from natural language, the semantic markerese problem is avoided and large numbers of primitives are provided for the differentiated description of concepts required by semantic vectors. A natural language program called meta5 uses CS; detailed examples of its operation are given. McDonald, D.R. & Bourne, L.E. Jr. (1986), Conditional Rule Testing in the Wason Card Selection Task, MCCS-85-57. We used the Wason card selection task, with variations, to study conditional reasoning. Disagreement exists in the literature, as to whether performance on this task improves when the problem is expressed concretely and when instructions are properly phrased. In order to resolve some inconsistencies in previous studies, we examined the following variables, (1) task intructions, (2) problem format, and (3) the thematic compatibility of solution choices with formal logic and with pre-existing schemas. In Experiment 1, performance was best in an 8-card, rather than a 4-card or a hierarchical decision-tree format. It was found in Experiment 2 that instructions directing subjects to make selections based on ``violation'' of the rule, rather than assessing its truth or falsity, resulted in more correct responses. Response patterns were predictable in part from formal logical considerations, but primarily from mental models, or schemas, based on (assumed) common prior experience and knowledge. Several explanations for the findings were considered. Partridge, D, McDonald, J., Johnston, V. & Paap, K. (1986) AI Programs and Cognitive Models: Models of Perceptual Processes, MCCS-85-60. We examine and compare two independently developed computer models of human perceptual processes: the recognition of objects in a scene and of words. The first model was developed to support intelligent reasoning in a cognitive industrial robot - an AI system. The second model was developed to account for a collection of empirical data and known problems with earlier models - a cognitive science model. We use these two models, together with the results of empirical studies of human behaviour, to generate a generalised model of human visual processing, and to further our claim that AI modelers should be more cognizant of empirical data. A study of the associated human phenomena provides an essential basis for understanding complex models as well as valuable constraints in complex and otherwise largely unconstrained domains.
yorick@nmsu.CSNET.UUCP (05/11/87)
Computing Research Laboratory New Mexico State University Box 30001 Las, Cruces, NM 88003. Krueger, W. (1986) Transverse Criticality and its Application to Image Processing, MCCS-85-61. The basis for investigation into visual recognition of objects is their representation. One appealing approach begins by replacing the objects themselves by their bounding surfaces. These then are represented by surfaces which have been smoothed according to various prescriptions. The resulting smoothed surfaces are subjected to geometric analysis in an attempt to find critical events which correspond to ``landmarks'' that serve to define the original object. Many vision researchers have used this outline, often incorporating it into a larger one that uses the critical events as constraints in surface generation programs. To deal with complex objects these investigators have proposed a number of candidates for the notion of critical event, most of which take the form of zero-crossings of some differentially defined quantity associated to surfaces (e.g. Gaussian curvature, etc.). Many of these require some a posteriori geometric conditioning (e.g. planarity) in order to be visually significant. In this report, we introduce the notion of a transverse critical line of a smooth function defined on a smooth surface. Transverse criticality attempts to capture the trough/crest behavior manifested by quantities which are globally defined on surfaces (e.g. curvature troughs and crests, irradiance troughs and crests). This notion can be used to study both topographic and photometric surface behavior and includes, as special cases, definitions proposed by other authors, among which notions are the regular edges of Phillips and Machuca [PM] and the interesting flutings of Marr [BPYA]. Applications are made to two classes of surfaces which are important in computer vision height surfaces and generalized cones. Graham, N. & Harary, F. (1986) Packing and Mispacking Subcubes into Hypercubes, MCCS-85-65. A node-disjoint packing of a graph G into a larger graph H is a largest collection of disjoint copies of G contained in H; an edge disjoint packing is defined similarly, but no two copies of G have a common edge. Two packing numbers of G into H are defined accordingly. It is easy to determine both of these numbers when G is a subcube of a hypercube H. A mispacking of G into H is a maximal collection of disjoint copies of G whose removal from H leaves no subgraph G, such that the cardinality of this collection is minimum. Two mispacking numbers of G into H are defined analogously. Their exact determination is quite difficult but we obtain upper bounds. Dietrich, E. & Fields, C. (1986), Creative Problem Solving Using Wanton Inference: It takes at least two to tango, MCCS-85-70. This paper introduces \fBwanton inference\fR, a problem solving strategy for creative problem solving. The central idea underlying wanton inference is that creative solutions to problems are often generated by ignoring boundaries between domains of knowledge and making new connections between previously unassociated elements of one's knowledge base. The major consequence of using the wanton inference strategy is that the size of search spaces is greatly increased. Hence, the wanton inference strategy is fundamentally at odds with the received view in AI that the essence of intelligent problem solving is limiting the search for solutions. Our view is that the problem of limiting search spaces is an artificial problem in AI, resulting from ignoring both the nature of creative problem solving and the social aspect of problem solving. We argue that this latter aspect of problem solving provides the key to dealing with the large search spaces generated by wanton inference. Ballim, A. (1986), The Subjective Ascription of Belief to Agents, MCCS-85-74. A computational model for determining an agent's beliefs from the viewpoint of an agent known as the system is described. The model is based on the earlier work of Wilks and Bien(1983) which argues for a method of dynamically constructing nested points of view from the beliefs that the system holds. This paper extends their work by examining problems involved in ascribing beliefs called meta-beliefs to agents, and by developing a representation to handle these problems. The representation is used in ViewGen, a computer program which generates viewpoints. Partridge, D. (1986), The Scope and Limitations of First Generation Expert Systems, MCCS-85-43. It is clear that expert system's technology is one of AI's greatest successes so far. Currently we see an ever increasing application of expert systems, with no obvious limits to their applicability. Yet there are also a number of well-recognized problems associated with this new technology. I shall argue that these problems are not the puzzles of normal science that will yield to advances within the current technology; on the contrary, they are symptoms of severe inherent limitations of this first generation technology. By reference to these problems I shall outline some important aspects of the scope and limitations of current expert system's technology. The recognition of these limitations is a prerequisite of overcoming them as well as of developing an awareness of the scope of applicability of this new technology. Gerber, M., Dearholt, D.W., Schvaneveldt, R.W., Sachania, V. & Esposito, C. (1987), Documentation for PATHFINDER: A Program to Generate PFNETs, MCCS-87-47. This documentation provides both user and programmer documentation for PATHFINDER, a program which generates PFNETs from symmetric distance matrices representing various aspects of human knowledge. User documentation includes instructions for input and output file formats, instructions for compiling and running the program, adjustments to incomplete or incompatable data sets, a general description of the algorithm, and a glossary of terms. Programmer documentation includes a detailed description of the algorithm with an explanation of each function and procedure, and hand execution examples of some of the more difficult to read code. Examples of input and output files are included. Ballim, A. (1986) Generating Points of View, MCCS-85-68. Modelling the beliefs of agents is normally done in a static manner. This paper describes a more flexible dynamic approach to generating nestings which represent what the system believes other agents believe. Such nestings have been described in Wilks and Bien (1983) as has their usefulness. The methods presented here are based upon those described in Wilks and Bien (ibid) but have been augmented to handle various problems. A system based on this paper is currently being written in Prolog. The Topological Cubical Dimension of a Graph Frank Harary MCCS-86-80 A cubical graph G is a subgraph of some hypercube $Q sub n$. The cubical dimension cd(G) is the smallest such n. We verify that the complete graph $K sub p$ is homeomorphic to a cubical graph H \(sb $Q sub p-1$. Hence every graph G has a subdivision which is a cubical graph. This enables us to define the topological cubical dimension tcd(G) as the minimum such n. When G is a full binary tree, the value of tcd is already known. Computer scientists, motivated by the use of the architecture of a hypercube for massively parallel supercomputers, defined the dilation of an edge e of G within a subdivision H of G as the lenth of the image of e in H, and the dilation of G as the maximum dilation of an edge of G. The two new invariants, tcd(G) and the minimum dilation of G among all cubical subdivisions H of G, are studied. CP: A Programming Environment for Conceptual Interpreters M.J. Coombs and R.T. Hartley MCCS-87-82 A conceptual approach to problem-solving is explored which we claim is much less brittle than logic-based methods. It also promises to support effective user/system interaction when applied to expert system design. Our approach is ``abductive'' gaining its power from the generation of good hypotheses rather than deductive inference, and seeks to emulate the robust cooperative problem-solving of multiple experts. Major characteristics include: (1) use of conceptual rather than syntactic representation of knowledge; (2) an empirical approach to reasoning by model generation and evaluation called Model Generative Reasoning; (3) dynamic composition of reasoning strategies from actors embedded in the conceptual structures; and (4) characterization of the reasoning cycle in terms of cooperating agents. Semantics and the Computational Paradigm in Cognitive Psychology Eric Dietrich MCCS-87-83 There is a prevalent notion among cognitive scientists and philosophers of mind that computers are merely formal symbol manipulators, performing the actions they do solely on the basis of the syntactic properties of the symbols they manipulate. This view of computers has allowed some philosophers to divorce semantics from computational explanations. Semantic content, then, becomes something one adds to computational explanations to get psychological explanations. Other philosophers, such as Stephen Stich have taken a stronger view, advocating doing away with semantics entirely. This paper argues that a correct account of computation requires us to attribute content to computational processes in order to explain which functions are being computed. This entails that computational psychology must countenance mental representations. Since anti-semantic positions are incompatible with computational psychology thus construed, they ought to be rejected. Lastly, I argue that in an important sense, computers are not formal symbol manipulators. Problem Solving in Multiple Task Environments Eric Dietrich and Chris Fields MCCS-87-84 We summarize a formal theory of multi-domain problem solving that provides a precise representation of the inferential dynamics of problem solving in multiple task environments. We describe a realization of the theory as an abstract virtual machine that can be implemented on standard architectures. We show that the behavior of such a machine can be described in terms of formally-specified analogs of mental models, and present a necessary condition for the use of analogical connections between such models in problem solving. An Automated Particulate Counting System for Cleanliness Verification of Aerospace Test Hardware \fIJeff Harris and Edward S. Plumer\fR MCCS-87-86 An automated, computerized particle counting system has been developed to verify the cleanliness of aerospace test hardware. This work was performed by the Computing Research Laboratory at New Mexico State University (CRL) under a contract with Lockheed Engineering and Management Services Company at the NASA Johnson Space Center, White Sands Test Facility. Aerospace components are thoroughly cleaned and residual particulate matter remaining on the components is rinsed onto 47 mm diameter test filters. The particulates on these filters are an indication of the contamination remaining on the components. These filters are examined under a microscope, and particles are sized and counted. Previously, the examination was performed manually; this operation has now been automated. Rather than purchasing a dedicated particle analysis system, a flexible system utilizing an IBM PC-AT was developed. The computer, combined with a digitizing board for image acquisition, controls a video-camera-equipped microscope and an X-Y stage to allow automated filter positioning and scanning. The system provides for complete analysis of each filter paper, generation of statistical data on particle size and quantity, and archival storage of this information for further evaluation. The system is able to identify particles down to 5 micrometers in diameter and discriminate between particles and fibers. A typical filter scan takes approximately 5 minutes to complete. Immediate operator feedback as to pass-fail for a particular cleanliness standard is also a feature. The system was designed to be operated by personnel working inside a class 100 clean room. Should it be required, a mechanism for more sophisticated recognition of particles based on shape and color may be implemented. Solving Problems by Expanding Search Graphs: Mathematical Foundations for a Theory of Open-world Reasoning Eric Dietrich and Chris Fields MCCS-87-88 We summarize a mathematical theory describing a virtual machine capable of expanding search graphs. This machine can, at least sometimes, solve problems where it is not possible to precisely and in detail specify the space it must search. The mechanism for expansion is called wanton inference. The theory specifies which wanton inferences have the greatest chance of producing solutions to given problems. The machine, using wanton inference, satisfies an intuitive definition of open-world reasoning. Software Engineering Constraints Imposed by Unstructured Task Environments Eric Dietrich and Chris Fields MCCS-87-91 We describe a software engineering methodology for building multi-domain (open-world) problem solvers which inhabit unstructured task environments. This methodology is based on a mathematical theory of such problem solving. When applied, the methodology results in a specification of program behavior that is independent of any architectural concerns. Thus the methodology produces a specification prior to implementation (unlike current AI software engineering methodology). The data for the specification are derived from experiments run on human experts. Multiple Agents and the Heuristic Ascription of Belief. Yorick Wilks and Afzal Ballim MCCS-86-75 A method for heuristically generating nested beliefs (what some agent believes that another agent believes ... about a topic) is described. Such nested beliefs (points of view) are esential to many processes such as discourse processing and reasoning about other agents' reasoning processes. Particular interest is paid to the class of beliefs known as \fIatypical beliefs\fR and to intensional descriptions. The heuristic methods described are emboddied in a program called \fIViewGen\fR which generates nested viewpoints from a set of beliefs held by the system. An Algorithm for Open-world Reasoning using Model Generation M.J. Coombs, E. Dietrich & R.T. Hartley MCCS-87-87 The closed-world assumption places an unacceptable constraint on a problem-solver by imposing an \fIa priori\fR notion of relevance on propositions in the knowledge-base. This accounts for much of the brittleness of expert systems, and their inability to model natural human reasoning in detail. This paper presents an algorithm for an open-world problem-solver. Termed Model Generative Reasoning, we replace deductive inference with a procedure based on the generation of alternative, intensional domain descriptions (models) to cover problem input, which are then evaluated against domain facts as alternative explanations. We also give an illustration of the workings of the algorithm using concepts from process control. Pronouns in mind: quasi-indexicals and the ``language of thought'' Yorick Wilks, Afzal Ballim, & Eric Dietrich MCCS-87-92 The paper examines the role of the natural-formal language distinction in connection with the "language of thought" (LOT) issue. In particular, it distinguishes a realist-uniform/attributist-uniform approach to LOT and seeks to link that distinction to the issue of whether artificial intelligence is fundamentally a science or engineering. In a second section, we examine a particular aspect of natural language in relation to LOT: pronouns/indexicals. The focus there is Rapaport's claims about indexicals in belief representations. We dispute these claims and argue that he confuses claims about English sentences and truth conditions, on the one hand, with claims about beliefs, on the other. In a final section we defend the representational capacity of the belief manipulation system of Wilks, Bien and Ballim against Rapaport's published criticisms.