David M. Steier

Varieties of learning in Soar: 1987

Abstract : Soar is an architecture for intelligence that integrates learning into all of its problem-solving behavior. The learning mechanism, chunking, has been studied experimentally in a broad range of tasks and situations. This paper summarizes the research on chucking in Soar, covering the effects of chunking in different tasks, task-independent applications of chunking and our theoretical analyses of effects and limits of chunking. We discuss what and when Soar has been able to learn so far. The results demonstrate that the variety of learning in Soar arises from variety in problem solving, rather than from variety in architectural mechanisms. Keywords: Artificial intelligence, Machine learning, cognitive architecture.

The Roles of Execution and Analysis in Algorthm Design

The analysis and execution of partial algorithm descriptions is an important part of the algorithm design process (as is borne out by studying the behavior of human algorithm designers). In this paper, we describe a language for representing partially designed algorithms and a process, developmental evaluation, that can discover useful knowledge to guide design. Using these and other results from our research in artificial intelligence, we are building a system, DESIGNER, that automatically designs algorithms. This paper also compares developmental evaluation to execution and analysis techniques used for testing complete programs and for validation of abstract specifications; concepts similar to those found in developmental evaluation are thus shown to apply to all stages of the software life cycle.

A knowledge-based mathematical model formulation system

The value of model-based approaches to support software development is well recognized. Modeling methodologies such as process modeling [4], behavioral analysis [19], and object-oriented analysis [11] among others, have been proposed to support conventional software development. While these approaches have a role to play in the development of knowledge systems, they do not directly address the knowledge-engineering bottleneck--the difficulty of incorporating large amounts of task-specific and search control knowledge. Furthermore, the development of knowledge systems is an ew)lutionary process. Being able to conceptualize and program using higherlevel models can significantly improve productivity. To accommodate these requirements, knowledge system developq n the sequel, we use the term mathematical models to refer to mathematical programming models. 2MFS currently represents only a subset of the kinds of knowledge listed. Its knowledge base is predominantly made up of detailed model construction knowledge related to the formulation of linear and integer programming models. ers have adopted models that describe a knowledge system at varying levels of abstraction. These range from the knowledge-level model (KLM), which provides an implementation-independent description of the system in terms of its goals, actions, and the knowledge it uses to select among its actions to the symbol-level model, in which implementation-specific commitments have been made [13]. The Soar architecture that we have employed to implement MFS provides three models that span this spectrum: the KLM, the problemspace-level model (PSM), and the symbol-level model. We will describe the KLM and the PSM of MFS. Using Soar to Develop Knowledge Systems Soar is a theory of how intelligent behavior can arise from a small set of cognitively plausible mechanisms. We do not cover the arguments for Soars cognitive plausibility here (the reader is referred to [14]) so that we can concentrate on behavior and the mechanisms. If one describes an agent with the ability to affect its environment as having goals, knowledge, and a capability for taking action, then one is describing it at the knowledge level. This level abstracts from the details of the internal processing (i.e., the inference mechanism or how the knowledge is represented), to yield a specification of how the agent would act if it could utilize all its available knowledge to perform tasks. Beause they have limited resources, physically realizable agents such as humans and computers, do not function with such perfect intelligence in solving complex tasks. Thus knowledgelevel descriptions are usually approximations. Nevertheless, this type of approximation serves a useful purpose for designing systems, by specifying the ideal behavior to be implemented by structure described at lower levels. At the knowledge level, one specifies the environment in which the agent operates, the goals of the agent, the knowledge the agent possesses, the perceptions (input) to the agent, and the possible actions (output) of the agent. Two key principles of the knowledge level are then added to such a specification. The principle of rationality dictates that the agent will ! 3 8 September 1992/Vo1.35, No.9/COMMUNICATIONS OF THE ACM

On Heterogeneous Database Retrieval: A Cognitively Guided Approach

Retrieving information from heterogeneous database systems involves a complex process and remains a challenging research area. We propose a cognitively guided approach for developing an information-retrieval agent that takes the users information request, identifies relevant information sources, and generates a multidatabase access plan. Our work is distinctive in that the agent design is based on an empirical study of how human experts retrieve information from multiple, heterogeneous database systems. To improve on empirically observed information-retrieval capabilities, the design incorporates mathematical models and algorithmic components. These components optimize the set of information sources that need to be considered to respond to a user query and are used to develop efficient multidatabase-access plans. This agent design, which integrates cognitive and mathematical models, has been implemented using Soar, a knowledge-based architecture.

Combining multiple knowledge sources in an integrated intelligent system

Using a stratified approach to system design embodied in Soar, multiple knowledge sources are integrated to implement systems performing different tasks: natural-language comprehension, production scheduling, and algorithm design. These three systems demonstrate that architectural mechanisms can play a key role in constructing systems to perform difficult knowledge-intensive tasks. Basic Soar principles are reviewed, and it is noted that Soar mechanisms reduce both design-time and runtime overhead associated with knowledge integration.<<ETX>>

Intelligent control of external software systems

Abstract This paper focuses on the relatively unexplored set of issues that arises when an intelligent agent attempts to use external software systems (EESs). The issues are illustrated initially in the context of the complex agent-ESS interactions in an engineering design example. Approaching the area from the perspective of artificial intelligence (AI) research, we find that in general, agent-ESS interactions vary widely. We characterize the possible variations in terms of performance capabilities required, skill levels at which performance is exhibited, and knowledge sources from which capabilities can be acquired. We are exploring these variations using Soar as our candidate AI agent; the document briefly describes seven Soar-based projects in early stages of development, in which agent-ESS issues are addressed. We conclude by placing agent-ESS research in the context of other work on software technology, and discuss the research agenda we have set for ourselves in this area.

Applying an architecture for general intelligence to reduce scheduling effort

A system called Merle-Soar is described which demonstrates how a specific architecture for general intelligence and learning (Soar) can be used to reduce scheduling effort when solving simple scheduling problems. In particular, we describe how Merle-Soar schedules sequences of jobs on a single bottleneck machine in a job shop. The knowledge of dispatching, acquired from examining how a human expert performs the task, is cast as search rules. A study was conducted which examined the extent to which learning could contribute to decreases in scheduling effort; specifically, the contribution of learning within-tasks was explored—the change in reasoning effort while solving a particular scheduling problem as knowledge is accumulated from successive trials. The results indicated that dramatic reductions in scheduling effort (in terms of the Soar architecture) were obtained. Knowledge gained early in the scheduling task was subsequently applied later in the task to reduce deliberation, and knowledge gained from one trial successfully reduced deliberation effort in subsequent trials. Additionally, the reduction exhibited the general power law of learning documented in psychological studies of skill acquisition. INFORMS Journal on Computing , ISSN 1091-9856, was published as ORSA Journal on Computing from 1989 to 1995 under ISSN 0899-1499.

A STRATEGY FOR RESEARCH ON MODEL FORMULATION

The starting point for this commentary is Sacks’ and Doyles conclusion that a central problem for qualitative physics is automating mathematical model formulation. We believe that model formulation is also a central problem for operations research, and although we have focused on models for production planning rather than for engineering systems analysis, our experience confirms that of Sacks and Doyle, that at least parts of model formulation are amenable to automation. In terms of their recommendations for future research, their strategy seems to emphasize the formalization of mathematical knowledge. We wish to stress that understanding the design or analysis context, the problem domain, and resource constraints on the modeling process is equally important. Methods used in cognitive psychology for understanding human problem solving, such as protocol analysis, can complement mathematical study by helping us understanding the processing that human modelers use to bring mathematical knowledge to bear. We have been using the results of such analyses to guide the creation of a model formulation system (MFS) within the Soar architecture. The use of cognitive studies and computer models in tandem seems to represent a viable strategy for making progress in this area.

Panel On Knowledge-based Design

The knowledge-based design panel explored new ideas for support of complex design processes and also explored the relationship between knowledge-based software engineering and other areas of knowledge-based design. This sununary consists of digests written by the panel members. Dr. Michael Lowry describes the knowledge life cycle, which is the maturation of design knowledge for an application domain from the initial research stage to the cookbook engineering stage, and its interaction with knowledge-based tools. Professor Gail Kaiser describes the current state of knowledge-based process support, and the need to incorporate models of cooperative design to scale up to real world problems. Professor Dorothy Setliff contrasts VLSI design tools with current software design tools. Dr. David Steier describes research toward providing intelligent support for using CAD tools in various engineering domains.