Showing papers by "Munindar P. Singh published in 1996"

Book review: ARCHON: An Architecture for Multi-agent Systems. Edited by Wittig, Thies (Ellis Horwood Limited (distributed by Prentice Hall), 1992)

Abstract: The term agent has become a popular buzzword, and the non-technical media particularly relish it. Similarly, our colleagues in traditional computer science, who often frown upon AI, now like to describe their programs as agents of various kinds. Agents might be new to the media and to computer network managers, but they have been around for decades in AI. Researchers in distributed artificial intelligence (DAI) have carried the torch, sometimes in the face of remarkable resistance. Recently, some of us have noticed that new entrants into the study of agents often tend to ignore the past achievements of 1 DAI. The present book, ARCHON, should serve as an excellent reminder of those achievements. Although collected volumes on DAI topics have appeared before, the present book is perhaps the first thorough description of a large DAI system. The book is essentially a midterm report on the five-year ARCHON project, which was funded by the Commission of European Communities under the ESPRIT program. (I wish other continents had similar programs!) The project involved fourteen organizations in eleven countries. The book has contributions by thirty-eight people, so it was itself a significant exercise in cooperation. The book consists of eight chapters written by different groups. It is a credit to Wittig's organizational and editorial efforts that it still reads like a book. Not the least of this book's charms is its brevity-the main body of the book is just over 100 pages. The ARCHON project developed a domain-independent architecture of multiagent systems. While some of the techniques are new, many are borrowed from previous work in DAI. This is as one would expect in a large system, and the authors are generous in giving credit. Two of the main challenges in building large systems with real applications are the autonomy and heter-ogeneity of the legacy components that must be logically integrated. The ARCHON architecture addresses these challenges by layering agent functionality above existing systems without significantly violating their autonomy. The additional function-ality is called the Archon layer, which handles interactions and cooperation among the systems. The Archon layer and the legacy component jointly constitute an agent. Agents-or rather, 1. I use 'ARCHON' to refer to the book, 'ARCHON' to refer to the project, and 'Archon' to refer to some of its components. their Archon layers-work like humans might in a similar system. Thus a cooperative system is produced from isolated components. It is …

Scheduling workflows by enforcing intertask dependencies

Abstract: Workflows are composite activities that can be used to support and automate multisystem applications involving humans, heterogeneous databases and legacy systems. The traditional atomic transaction model, successful for centralized and homogeneous applications, is not suitable for supporting such workflows. Intertask dependencies, which are conditions involving events and dependencies among workflow tasks, are used to specify the coordination requirements among the workflow tasks and are a central component of most workflow models. They form a basis for developing a uniform formal framework for workflows, which is a key contribution of this work. In this paper, we formalize intertask dependencies using temporal logic. This involves event attributes, which are needed to determine whether a dependency is enforceable and to properly schedule events. Each dependency is represented internally as a finite state automaton that captures the computations that satisfy the given dependency. Sets of automata are combined into a scheduler that produces global computations satisfying all relevant dependencies, thus enacting the given workflow. This algorithm is rigorously proved correct; it has been implemented.

Synthesizing distributed constrained events from transactional workflow specifications

Abstract: Workflows are the semantically appropriate composite activities in heterogeneous computing environments. Such environments typically comprise a great diversity of locally autonomous databases, applications and interfaces. Much good research has focused on the semantics of workflows and how to capture them in different extended transaction models. We address the complementary issues pertaining to how workflows may be declaratively specified and how distributed constraints may be derived from those specifications to enable local control, thus obviating a centralized scheduler. Previous approaches to this problem were limited and often lacked a formal semantics.

Quality of service and scientific workflows.

Abstract: The advent of high-performance computing engines and networks is bringing serious numerical and problem-solving environments (PSEs) closer to a broad base of users with widely differing needs. From the perspective of these users, a key issue will be the quality of service (QoS) PSEs offer. In the broader sense, QoS includes parameters such as network delays and throughput, as well as end-user quality factors such as system availability, system functionality, content quality, and semantic interoperability. In order to facilitate integration of the QoS and PSE we introduce scientific workflows, to mean a series of structured activities and computations that arise in scientific problem-solving. Scientific workflows are expected to coexist and cooperate with other user workflows (e.g., business workflows, educational workflows, legislative workflows). As such they must support compatible QoS. We use data from existing systems and workflows to quantitatively bound some of the PSE QoS parameters. Use of multimedia imposes additional restrictions, while end-user risks impose bounds on the security and reliability of numerical computations and algorithms It is our belief that the next generation of PSEs must have QoS parameters designed into the system, or these PSEs will fail to live up to user needs and expectations.

A Conceptual Analysis of Commitments in Multiagent Systems

Abstract: The notion of commitment is central to understanding agents and multiagent systems. At least two kinds of commitment can be identified in the AI literature---the internal or psychological and the external or social. While these notions must not be conflated with each other, they are not entirely unrelated. We review the historical development of commitments in AI and distributed computing, and show the various roles they might play in multiagent systems. We discuss the key interrelationships among these concepts, and study their implementational aspects, both in traditional individual agents, and in agents that are recursively created as systems of other agents. We close with a discussion of the key research challenges.

Toward Interaction-Oriented Programming

Abstract: Although much progress has been made in agent theory and practice, bottlenecks remain in the construction of complex multiagent systems. We introduce *interaction-oriented programming (IOP)* as an approach to orchestrate the interactions among agents. As envisioned, IOP is more tractable and practical than general agent programming, especially in settings such as open information environments, where the internal details of autonomously developed agents are not available. IOP facilitates multiagent system design by enabling declarative specification and enactment of agent interactions, thereby channeling the intellectual energies of designers into the most amenable and effective design tasks. We have begun the work on realizing the above vision. Our preliminary approach formally specifies certain interactions among agents, and executes them in a distributed manner. We have implemented it in an actor language.

Synthesizing Distributed Constrained Events from Transactional Workflow

Abstract: Workflows are the semantically appropriate composite activities in heterogeneous computing environments. Such environments typically comprise a great diversity of locally autonomous databases, applications, and interfaces. Much good research has focused on the semantics of workflows, and how to capture them in different extended transaction models. Here we address the complementary issues pertaining to how workflows may be declaratively specified, and how distributed constraints may be derived from those specifications to enable local control, thus obviating a centralized scheduler. Previous approaches to this problem were limited and often lacked a formal semantics.

Challenges for Machine Learning in Cooperative Information Systems

Abstract: Cooperative Information Systems (CISs) are multiagent systems with organizational and database abstractions geared to the large open heterogeneous information environments of today. CIS is also the name of the associated research area, which has emerged from the synthesis of distributed databases and distributed artificial intelligence. In CIS, software agents mitigate an information environment's heterogeneity by interacting through common protocols, and manage its large size by making intelligent local decisions without centralized control. In order to cope with the dynamism presented by open environments, CIS agents must have the ability to adapt and learn. We discuss some of the most important problems involving learning and adaptivity in CISs, including requirements for reconciling semantics and improving coordination. We present a “customers' view” of learning technology as might find ready application in CISs.

Toward a Model Theory of Actions: How Agents do it in Branching Time

Abstract: A clear understanding and formalization of actions is essential to computing, and especially so to reasoning about and constructing intelligent agents. Several approaches have been proposed over the years. However, most approaches concentrate on the causes and effects of actions, but do not give general characterizations of actions themselves. A useful formalization of actions would be based on a general, possibly nondiscrete, model of time that allows branching (to capture agents'' choices). A desirable formalization would also allow actions to be of arbitrary duration and would permit multiple agents to act concurrently. We develop a branching-time framework that allows great flexibility in how time and action are modeled. We motivate and formalize several coherence constraints on our models, which capture some nice intuitions and validate some useful inferences relating actions with time.

Commitments in the Architecture of a Limited, Rational Agent

Abstract: Rationality is a useful metaphor for understanding autonomous, intelligent agents. A persuasive view of intelligent agents uses cognitive primitives such as intentions and beliefs to describe, explain, and specify their behavior. These primitives are often associated with a notion of commitment that is internal to the given agent. However, at first sight, there is a tension between commitments and rationality. We show how the two concepts can be reconciled for the important and interesting case of limited, intelligent agents. We show how our approach extends to handle more subtle issues such as precommitments, which have previously been assumed to be conceptually too complex. We close with a proposal to develop conative policies as a means to represent commitments in a generic, declarative manner.

Distributed Scheduling of Workflow Computations

Abstract: Workflows are composite activities that achieve interoperation of a variety of system and human tasks. Workflows must satisfy subtle domain-specific semantic and organizational requirements. Consequently, flexibility in execution is crucial. We address the problem of scheduling workflows from declarative specifications in terms of intertask dependencies and event attributes. Our approach goes beyond previous approaches in being generic and distributed, and having a rigorous formal semantics. It involves distributed events, which are automatically set up to exchange the necessary messages. Our approach uses symbolic reasoning to (a) determine the initial constraints on events (or guards), (b) preprocess the guards, and (c) execute the events. This approach has been implemented in an actor language.