Showing papers in "Intelligence\/sigart Bulletin in 1991"
TL;DR: Dyna as mentioned in this paper is an AI architecture that integrates learning, planning, and reactive execution, where learning methods are used both for compiling planning results and for updating a model of the effects of the agent's actions on the world.
Abstract: Dyna is an AI architecture that integrates learning, planning, and reactive execution. Learning methods are used in Dyna both for compiling planning results and for updating a model of the effects of the agent's actions on the world. Planning is incremental and can use the probabilistic and ofttimes incorrect world models generated by learning processes. Execution is fully reactive in the sense that no planning intervenes between perception and action. Dyna relies on machine learning methods for learning from examples---these are among the basic building blocks making up the architecture---yet is not tied to any particular method. This paper briefly introduces Dyna and discusses its strengths and weaknesses with respect to other architectures.
681 citations
TL;DR: The PRODIGY architecture is described, describing its planning and problem solving capabilities and touching upon its multiple learning methods, as well as issues in architectural design, providing a context to examine the underlying tenets of the PRODigY architecture.
Abstract: Artificial intelligence has progressed to the point where multiple cognitive capabilities are being integrated into computational architectures, such as SOAR, PRODIGY, THEO, and ICARUS. This paper reports on the PRODIGY architecture, describing its planning and problem solving capabilities and touching upon its multiple learning methods. Learning in PRODIGY occurs at all decision points and integration in PRODIGY is at the knowledge level; the learning and reasoning modules produce mutually interpretable knowledge structures. Issues in architectural design are discussed, providing a context to examine the underlying tenets of the PRODIGY architecture.
259 citations
TL;DR: A symbol level account of some of the representation and reasoning structures within the LOOM knowledge representation system, which is unique in that it constructs a separate taxonomy for each of seven kinds of non-composite descriptions, and uses a marker passing algorithm to replace the quadratic time subsumption test found in most classifiers with a linear time test.
Abstract: This paper presents a symbol level account of some of the representation and reasoning structures within the LOOM knowledge representation system. Reasoning in LOOM centers around a classifier whose primary function is to construct a taxonomy of all descriptions that have been entered into the system. The LOOM classifier is unique in that it constructs a separate taxonomy for each of seven kinds of non-composite descriptions, and uses a marker passing algorithm to replace the quadratic time subsumption test found in most classifiers with a linear time test. We briefly illustrate how the selection of data structures within LOOM impacts the completeness of the classification algorithm, and we describe the LOOM option that allows concepts to be reasoned with in either a forward-chaining or a backward-chaining mode.
217 citations
TL;DR: An overview of the ANA architecture is presented, which describes the functionalities that has been implemented, the results that have been obtained with robotic and simulated ANA agents, and finally it discusses (current) limitations and future work.
Abstract: The goal of my work is to develop and implement an architecture for an autonomous agent, which I refer to as "ANA". An ANA agent consists of a distributed set of "competence modules". Competence modules are linked in a network. A spreading activation process operates on the network to decide what the "relevance" or relative strength of a competence module is in the current context. This process implements a competition among modules for activation energy. The higher the activation energy level of a module, the more likely it is that this module determines what the autonomous agent does or communicates to believe. Learning is a central, completely integrated feature of the architecture. The competence module network is continuously being developed and changed on the basis of experience: links are added and deleted depending on real world observations and new "macro modules" are created whenever a goal is achieved. This paper presents an overview of the architecture. It describes the functionalities that have been implemented, the results that have been obtained with robotic and simulated ANA agents, and finally it discusses (current) limitations and future work.
182 citations
TL;DR: It turned out that almost all implemented KL-ONE systems such as BACK, KL-TWO, LOOM, NIKL, SB-ONE use sound but incomplete algorithms.
Abstract: The knowledge representation system KL-ONE first appeared in 1977. Subsequently many systems based on the idea of KL-ONE have been built. The formal model-theoretic semantics which has been introduced for KL-ONE languages [9] provides means for investigating soundness and completeness of inference algorithms. It turned out that almost all implemented KL-ONE systems such as BACK, KL-TWO, LOOM, NIKL, SB-ONE use sound but incomplete algorithms.
133 citations
TL;DR: The BACK project was begun at the Technical University Berlin in January 1985 and supports complex representation of a domain terminology, description of domain objects using that terminology, and database access via a uniform interface language.
Abstract: The BACK project was begun at the Technical University Berlin in January 1985 as part of a larger project within the ESPRIT programme1. Our task in the project group was the specification, design and implementation of a knowledge representation system which we called BACK ("Berlin Advanced Computational Knowledge Representation System"). It is a based on a terminological logic (term description language) and supports complex representation of a domain terminology, description of domain objects using that terminology, and database access via a uniform interface language.
105 citations
TL;DR: This paper discusses what is easy, what is hard, and where the research frontiers lie in behavior based systems and an orthogonal view of integration issues is highlighted.
Abstract: Behavior based systems require an orthogonal view of integration issues. In this paper we highlight those issues, discuss what is easy, what is hard, and where the research frontiers lie.
96 citations
TL;DR: The CLASSIC system explores the expressiveness vs. tractability tradeoff, driven by concerns of usefulness and usability in several real applications and embodies the views of what a knowledge representation system should be.
Abstract: Our work on the CLASSIC knowledge representation system covers a broad range from theory to practice. While CLASSIC was implemented primarily to provide a simple, easy to learn and use, locally available tool for a relatively limited set of applications, it has a substantial theoretical foundation, based on a formal "terminological" logic. The logical foundation provides the semantics of a term description language, which is used to define structured concepts and make assertions about individuals in a knowledge base. These concepts and individuals are organized into a generalization hierarchy by classification and subsumption algorithms. The CLASSIC system explores the expressiveness vs. tractability tradeoff, driven by concerns of usefulness and usability in several real applications. Within this context, it embodies our views of what a knowledge representation system should
94 citations
[...]
TL;DR: In attempting to construct a broad agent, constraints seem to arise between components of the architecture and in this brief note, some of these constraints are discussed.
Abstract: The Oz project at Carnegie Mellon is developing technology for dramatic virtual worlds. One requirement of such worlds is the presence of broad, though perhaps shallow, agents. To support our needs, we are developing an agent architecture that provides goals and goal directed reactive behavior, emotional state and its effects on behavior, some natural language abilities (especially pragmatics based language generation), and some memory and inference abilities. We are limiting each of these capacities whenever necessary to allow us to build a broadly capable, integrated agent.In attempting to construct a broad agent, constraints seem to arise between components of the architecture. In this brief note, we discuss some of these constraints.
79 citations
TL;DR: The K-Rep system was built to explore the utility of a KL-One style knowledge representation in the development of expert systems, and abandoned the rule based approach in favor of organizing the system as a set of problem solvers around a common conceptual core.
Abstract: The K-Rep system was built to explore the utility of a KL-One style knowledge representation in the development of expert systems. Beginning in about 1985, our activity in expert systems has been centered on the FAME (FinAncial Marketing Expertise) system[4]. FAME attempts to provide support to an IBM marketing representative in the financing decisions involved in the acquisition of large mainframe computer systems. Based on our experience in building a feasibility demonstration of FAME using a rule based approach, we concluded that the rule based technology would not easily scale up. Thus we abandoned the rule based approach in favor of organizing the system as a set of problem solvers around a common conceptual core. Since diverse problem solvers would be utilized in FAME, it was thought desireable that the conceptual core have a well-defined, enforceable semantics. These considerations led us to the KL-One[3] style knowledge representation.
71 citations
TL;DR: An architecture for controlling autonomous mobile robots based on control of continuous activities (processes) rather than discrete actions is presented, and it is argued that different levels of activities require different sorts of computational mechanisms to control them.
Abstract: We present an architecture for controlling autonomous mobile robots based on control of continuous activities (processes) rather than discrete actions. We define a hierarchy of activity, and argue that different levels of activities require different sorts of computational mechanisms to control them. Many controversial issues concerning the use of persistent internal state and higher levels of abstraction can be better understood in terms of this hierarchy. Two experiments using the architecture to control mobile robots performing complex navigation tasks are described.
TL;DR: User Models in Dialog Systems as mentioned in this paper is an informative and thought-provoking book based upon a user modeling conference held in 1986, which was used to promote the idea that interactive software systems can eventually become intelligent partners in goal-directed dialogs with humans.
Abstract: User Models in Dialog Systems is informative and thought-provoking. The book is based upon a user modeling conference held in 1986. Since then, the various contributors have obviously worked at adding tutorial and review material to the chapters to supplement the research reports. This is very useful, especially since the primary effect of the book will be to promote the idea that interactive software systems can eventually become intelligent partners in goal-directed dialogs with humans. The book is by no means an engineering handbook; it is primarily theoretical, but includes sketches of requirements specifications and feasibility demonstrations.
TL;DR: The designs for ICARUS are described, an integrated architecture for controlling an intelligent agent in a complex physical environment that includes components for perceiving the environment, for generating plans to solve problems, and for executing the plans generated.
Abstract: We describe our designs for ICARUS, an integrated architecture for controlling an intelligent agent in a complex physical environment. By navigating between locations and manipulating other objects, the agent influences the world and achieves its goals. The architecture includes components for perceiving the environment, for generating plans to solve problems, and for executing the plans generated. A fourth component manages the agent's long-term memory. Our assessment of the design suggests that it is general, versatile, scales well to larger problems, and is consistent with a variety of psychological results.
TL;DR: Some of the methodological considerations and the technical decisions that are both different from conventional thinking and that have significantly influenced the design of the CycL language are reviewed.
Abstract: CycL is the language in which the Cyc Knowledge Base is being encoded. This paper reviews some of the methodological considerations and the technical decisions (resulting from these considerations) that are both different from conventional thinking and that have significantly influenced the design of the language.
TL;DR: This paper addresses Laird's issues concerning integrated intelligent architectures in the context of work done by the Situated Automata group at SRI International, CSLI, and Teleos Research.
Abstract: This paper addresses Laird's issues concerning integrated intelligent architectures in the context of work done by the Situated Automata group at SRI International, CSLI, and Teleos Research. The first section contains a brief general introduction to the theory and programming methodologies of the situated-automata approach, the second section discusses some terminological difficulties; and the third section addresses the specific issues.
TL;DR: This paper gives a linear time algorithm for the n - queens problem, an extension of one of the previous local search algorithms that is capable of solving problems with 3,000,000 queens in approximately 55 seconds.
Abstract: The n - queens problem is a classical combinatorial search problem. In this paper we give a linear time algorithm for this problem. The algorithm is an extension of one of our previous local search algorithms [3, 4, 6]. On an IBM RS 6000 computer, this algorithm is capable of solving problems with 3,000,000 queens in approximately 55 seconds.
TL;DR: Access-Limited Logic, though incomplete, still has a well defined semantics and a weakened form of completeness, Socratic Completeness, which guarantees that for any fact which is a logical consequence of the knowledge-base, there is a series of preliminary queries and assumptions after which a query of the fact will succeed.
Abstract: Access-Limited Logic (ALL) is a theory of knowledge representation which formalizes the access limitations inherent in a network structured knowledge-base. Where a deductive method such as resolution would retrieve all assertions that satisfy a given pattern, an access-limited logic retrieves only those assertions reachable by following an available access path. The time complexity of inference in ALL is a polynomial function of the size of the accessible portion of the knowledge-base, rather than an exponential function of the size of the entire knowledge-base (as in much past work). Access-Limited Logic, though incomplete, still has a well defined semantics and a weakened form of completeness, Socratic Completeness, which guarantees that for any fact which is a logical consequence of the knowledge-base, there is a series of preliminary queries and assumptions after which a query of the fact will succeed.
TL;DR: Over the years, AI has divided itself into many subfields corresponding roughly to different components of intelligence such as planning, knowledge representation, learning, vision, robotics, natural language, and interaction with other agents.
Abstract: Over the years, AI has divided itself into many subfields corresponding roughly to different components of intelligence such as planning, knowledge representation, learning, vision, robotics, natural language, and interaction with other agents. Although this division has allowed the field to make progress in understanding each of these components, it often ignores the issues involved in creating integrated systems that draw from more than one of these subfields. SHRDLU was an early example of the integration of planning and natural language, but in general we find ourselves with individual capabilities that are not integrated within a complete system.
TL;DR: The n-queens problem is often used as a benchmark problem for AI research and in combinatorial optimization and a polynomial time algorithm for finding a solution was presented in this magazine.
Abstract: The n-queens problem is often used as a benchmark problem for AI research and in combinatorial optimization. An example is the recent article [1] in this magazine that presented a polynomial time algorithm for finding a solution. Several CPU-hours were spent finding solutions for some n up to 500,000.
TL;DR: KRS demonstrates that it is now possible to develop knowledge representation servers on a par with floating-point arithmetic units and numeric libraries, as modules with well-defined functionality and fast, reliable performance.
Abstract: The design and application of KRS, a knowledge representation server in the KL-ONE, KRYPTON, CLASSIC family, are described. The server is designed as an open architecture module that can be used as a stand-alone service or embedded in other systems. It accepts the constraints necessary to make subsumption and recognition tractable, and maintains a careful distinction between definitions and assertions. It is implemented as a class library in an object-oriented language using generic, reusable objects. The approach taken to the integration of the server with external knowledge representation servers of similar or dissimilar types is analyzed. The server supports reasoning with exceptions and incomplete data through computation of a three-valued subsumption relation that is able to determine the possibility of further inferences if more assertions are made. KRS demonstrates that it is now possible to develop knowledge representation servers on a par with floating-point arithmetic units and numeric libraries, as modules with well-defined functionality and fast, reliable performance. It is also proving a useful tool for the empirical investigation of some large-scale knowledge representation server applications.
TL;DR: The SB-ONE workbench is described, the motivations and philosophies behind its components are illustrated, some feedback received from users is reported, and some of the components have been employed by a fairly large number of people in various natural-language applications.
Abstract: SB-ONE is a workbench for the representation of conceptual knowledge, with the emphasis on applications in natural-language systems Besides the SB-ONE language, the workbench comprises three different interfaces, a partition mechanism, a consistency maintenance system for the syntactic well-formedness of SB-ONE knowledge bases, a classifier, a realizer, a pattern matcher, a spreading-activation mechanism, an interpreter and classifier for SB-ONE to SB-ONE translation rules, an integration mechanism for an external frame-based representation, and a connection between SB-ONE and an extended Prolog These components have been employed by a fairly large number of people in various natural-language applications This paper describes the SB-ONE workbench, illustrates the motivations and philosophies behind its components, and reports some feedback received from users
TL;DR: A structured bibliography of the body of knowledge that has begun to accumulate on how to integrate intelligent computers and human practitioners into an effective cooperative system is presented.
Abstract: This paper presents a structured bibliography of the body of knowledge that has begun to accumulate on how to integrate intelligent computers and human practitioners into an effective cooperative system. Work on this topic is divided into four major sections. The first covers empirical work related to human-intelligent system cooperation. The second covers work in system building, i.e., prototypes that instantiate new concepts and capabilities for more effective cooperative interaction with people. The third section reviexvs concepts for human-intelligent system cooperation based on models of human performance and errors or models of the cognitive demands of domain tasks. The fourth section includes review articles, books and workshops relevant to this are&
TL;DR: This volume is a collection of previously published (save for one) papers on the subfield of artificial intelligence dealing with planning and the action, providing a general framework for study in this field.
Abstract: Readings In Planning (RIP) is another volume in a fine series of anthologies published by Morgan Kaufman. This volume is a collection of previously published (save for one) papers on the subfield of artificial intelligence (AI) dealing with planning and the action. This volume presents many of the historically important papers while providing a general framework for study in this field.
TL;DR: A set of ITS knowledge acquisition tools tailored for usability by teachers, which facilitates rapid prototyping and testing of curriculum and multiple tutoring strategies, and compares with related research in generic tutoring systems and knowledge acquisition.
Abstract: This research addresses the widening gap between research in intelligent tutoring systems and practical use of this technology by the educational community. In order to insure that intelligent tutoring systems (ITSs) are effective, teachers must be involved in their design and evaluation. We have followed a user participatory design process to build a set of ITS knowledge acquisition tools tailored for usability by teachers. The system facilitates rapid prototyping and testing of curriculum and multiple tutoring strategies. Teachers use the system to create, modify, and test the system's domain content and tutoring strategy knowledge bases. The design includes novel methodologies for strategy representation and overlay student modeling, and incorporates considerations from instructional design theory. Tools have been designed to provide the user with visual models of the concepts and structures of the underlying framework. In close collaboration with a veteran high school teacher, we have used the interface to design a tutor for statics (part of a high school physics course). In this paper we describe the system (called KAFITS), report on our experience involving educators in ITS development, discuss issues of ITS knowledge representation and acquisition, and compare the system with related research in generic tutoring systems and knowledge acquisition.
TL;DR: This work is extending the Construction-Integration Model even further so that it can be considered a general purpose Cognitive Architecture so thatIt can be compared to other architectures.
Abstract: The Construction-Integration Model is a theory of discourse comprehension that is being used as the foundation for a new Integrated Intelligent Architecture. It is a hybrid model, combining symbolic features with connectionist techniques, that emphasizes bottom-up, data-driven comprehension processes over more rigid top-down search strategies. Recent work has focussed on using the Model to simulate problem solving behavior as might be seen during routine computing tasks. As a consequence of this latter work, we are extending the Model even further so that it can be considered a general purpose Cognitive Architecture so that it can be compared to other architectures.
TL;DR: This work demonstrates a fully intergrated physical system (a mobile robot) performing navigation, landmark detection, map learning, and planning, without explicit integration, and is fully reactive.
Abstract: We address the problem of integrating a number of capabilities into a situated system without introducing a hybrid solution consisting of a reactive low-level, and a deliberative high-level. We demonstrate a fully intergrated physical system (a mobile robot) performing navigation, landmark detection, map learning, and planning, without explicit integration. Inspired by biology and the subsumption architecture, our approach proceeds bottom-up and is fully reactive. It maximizes the use of direct information from the world, and minimizes communication within the system. The control is fully distributed over all of the components, and does not use any manipulable data structures or symbolic representations. Instead, the representation of the entire system is homogeneous; it consists of simple reactive rules which encode both the control strategy and the knowledge of the system. The exchange of knowledge within the system is mostly done implicitly, by observing, and relying on, the effects of the other components. Integration is accomplished through the components' interaction with the environment. The homogeneous, behavior-based nature of the representation does not make a distinction between the reactive and deliberating parts of the system, making them all active and interactive.
TL;DR: An integrated system for planning and learning based on the Direct Memory Access Parser of Martin (1990) and the Reactive Action Packages of Firby (1989) is developed.
Abstract: We are developing an integrated system for planning and learning. The type of learning we are interested in is advice-taking, or learning by being told. Our model of planning is based on the Reactive Action Packages (RAPs) of Firby (1989). Our model of learning is based on the Direct Memory Access Parser (DMAP) of Martin (1990).
TL;DR: The MVL theorem-proving system is an implementation of theoretical work currently under way at Stanford University that is regularly included in MVL to ensure that the ideas developed theoretically can be tested experimentally as well.
Abstract: The MVL theorem-proving system is an implementation of theoretical work currently under way at Stanford University. Theoretical progress made there is regularly included in MVL in order to ensure that the ideas developed theoretically can be tested experimentally as well.
TL;DR: The Entropy Reduction Engine is described, motivated, presented, and analyzed in terms of its different components; namely, problem reduction, temporal projection, and situated control rule execution.
Abstract: This paper describes the Entropy Reduction Engine, an architecture for the integration of planning, scheduling, and control. The architecture is motivated, presented, and analyzed in terms of its different components; namely, problem reduction, temporal projection, and situated control rule execution. Experience with this architecture has motivated the recent integration of learning, and this paper also describes the learning methods and their impact on architecture performance.
TL;DR: The paper briefly describes the control constructs of TCA, and characterizes their applicability along several important dimensions.
Abstract: To operate effectively with complex tasks in complex environments, agents must intelligently manage their limited physical and computational resources. This can be accomplished by explicitly coordinating the agent's planning, perception, and action. The Task Control Architecture (TCA) provides a vocabulary of coordination constructs that combines both deliberative and reactive aspects. The paper briefly describes the control constructs of TCA, and characterizes their applicability along several important dimensions.