ZiF Conference "Solving Complex Problems with Agent Systems"

Bielefeld, February 17 - 18, 1994

List of Abstracts

program - participants - abstracts - documentation - photos - about ZiF

William Bricken (HITL Seattle)

Entity-based Modeling in Virtual Environments

Immersive virtual environments provide a natural testbed for complex interaction between humans and computational agents. In VR, every object should be able to exhibit both reactive and autonomous behavior; every participant should be free to interact arbitrarily with any object. The demands of immersive interaction have lead us to a particular type of agent architecture: entity-based modeling. Entity-based modeling extends object-oriented programming to systems-oriented programming by creating agents that act as independent operating systems, controlling their own process resources, memory resources, and interprocess communication. An entity can be conceptualized as an organizationally closed quasi-biological system with control functions that define perception, action, and motivation. An essential component of this model is that every entity serves both as an object interacting within an external context and as an environment providing global context for its internal content. For VR applications, entities also serve as virtual bodies which are controlled by the dynamic activity of human participants.
Hans-Dieter Burkhard (Humboldt-Universität Berlin)

Agent-oriented Programming and Open Systems

It is common understanding that DAI does not reduce to distributed problem solving. But most tools for agent programming are related to cooperative solutions of (single) global goals. They are usually not designed to meet the needs of open systems, as e.g. continuous availability, extensibility, decentral control, asynchrony, inconsistent information, arm's-length relationships (Hewitt). Some of those defects are due to implementation only, while others concern the programming style as well. Thus, agent oriented programming for open systems needs a discussion of independent and distributed implementations at one hand, and of related programming primitives at the other hand.
Yves Demazeau (LIFIA Grenoble)

From Feature Extraction to Integration of Visual Modules Using Agent Systems

We are interested in studying how coherence between knowledge representations as well as coherence between behaviours can be programmed or can emerge in Multi-Agent Systems. Multi-Agent research at LIFIA includes looking for new computational paradigms, the implementation of tools for programming multi-agent environments, and the development of applications in several fields. We present here a detailed overview of these paradigms, the agent models which are developed, and their use to solve complex problems in the field of Image Analysis and Computer Vision. From the behavioral viewpoint, Multi-Agent Systems have to solve the coherence problem between agent behaviours in order to infer coherent behavioral information, known as collective intelligence. The PACO project considers multiple fine-grained agents - mainly reactive - that interact in order to simulate and to solve a global problem. We are developing a new paradigm called the "Coordination Patterns", in which the concepts of Autonomous Agents, Multi-Agent Systems, Complex Dynamics, and Emergent Functionality meet. On the basis of multiple reactive fine-grained agents, scope rules and possible a priori knowledge guide interaction between agents themselves and between agents and the environment. We first present the "Coordination Patterns" paradigm and then present its application for classical Image Analysis problems such as Segmentation into regions and Intelligent Boundary Detection (ELASTIC PATTERNS, PACOVISION): From the representational viewpoint, Multi-Agent Systems have to solve the coherence problem between subjective local representations of the environment in order to infer a global objective representation. The COHIA project deals with several heterogeneous agents - mainly intentional. We are developing a new paradigm called "Agent-Oriented Integration" which includes a decomposition of the knowledge representation and of the knowledge processing of complex systems, the basis of several coarse-grained agents and the use of interaction protocols to integrate the overall system. We first present the "Agent- Oriented Integration" Paradigm and then present its application for the traditional Computer Vision problem of Integrating General Purpose Vision Systems (VAP, SATURNE, MAGIC). We finally discuss our interests in collecting the results and insights from the two previous projects in order to develop a model of an autonomous agent in a multi-agent world, including both programmed and emergent functionalities and behaviours. We briefly discuss its future applications in the field of Computer Vision.
Franco di Primio/Bernd S. Müller (GMD Bonn-Sankt Augustin)

Minimal Scenarios for Studying the Transition from Individual to Competitive and to Cooperative Behavior

The main idea is to study the interaction and the behavioral organisation of agents which are based on many different but very simple sensorimotor units. This is much in the same understanding as Braitenberg's experimental and synthetic psychology. Different basic scenarios are developed in order to show the constraints which are necessary for the "emergence" of joint behavior, like the competitive or cooperative pursuit of targets, and communication. The basic research questions are: concept formation, cognitive mapping, anticipation of events. Work is done in form of simulations and real world experiments with minirobots.
Ed Durfee (University of Michigan)

Distributed Problem Solving and Multiagent Systems: Commonalities, Differences, and Examples

Distributed AI is commonly divided into two major branches: distributed problem solving and multiagent systems. Yet, typically, distributed problem solving involves multiple agents, and multiagent systems are developed to solve problems. So what are the differences between these? In this talk, I trace the history of these different trends and offer some insights as to how these branches differ, using examples of these systems to highlight these points.
Jürgen Emhardt (Freie Universität Berlin)

Task-oriented Agent for Exploring Virtual Worlds

Task-oriented agents support users in navigating through large virtual worlds and prevent them from getting lost in the virtual environment. In addition, they are able to map hypertext bases onto parts of the virtual world and solve orientation problems of hypertext systems as well. Besides of the equal treatment of virtual worlds and hypertexts, the agents let users explore the behavior of objects by serving as mediators between the user and the virtual world.
Klaus Fischer (DFKI Saarbrücken)

The Multi-Agent System Development Tool MAGSY: Methods and Applications

The rule-based multi-agent system MAGSY is presented. The kernel of an agent in MAGSY is a forward-chaining rule interpreter. Therefore, each agent has the problem solving capacity of an expert system. Each agent in MAGSY is a self-contained entity with a unique identification. If an agent knows the identification of another agent, it is able to send this agent messages. In doing so the agents perform services and call on services of other agents. Agents are able to create new agents dynamically. If an agent creates a new agent, it automatically knows the identification of this new agent. The knowledge of the agents is structured in an object-oriented knowledge representation scheme. There is a global knowledge base which contains the knowledge that may be accessed by all of the agents. Agents may store their identification in this global knowledge base and thus become known to all agents in the system. To show how the applicability of the concepts built into the MAGSY system, the design of the planning and controlling components of a flexible manufacturing system is described. The problem of role assignment in dynamically established groups and modelling cooperative task planning is described in detail.
Nick Jennings (University of London)

ARCHON: A Distributed Artificial Intelligence System for Industrial Applications

ARCHON (ARchitecture for Cooperative Heterogeneous ON-line systems) is Europe's largest project in the area of Distributed Artificial Intelligence (DAI). It has devised a general-purpose architecture, software framework and methodology which has been used to support the development of DAI systems in industrial domains. Some examples of the applications to which it has been successfully applied include: electricity distribution and supply, electricity transmission and distribution, control of a cement kiln complex, control of a particle accelerator, and control of a robotics application. The type of cooperating community that it supports has a decentralised control regime and individual problem solving agents which are large grain, loosely coupled, and semi-autonomous. This talk will tackle a broad range of issues related to the application of ARCHON to industrial applications. The rationale for a DAI approach to industrial applications will be given, the ARCHON framework will be detailed, and a description of the electricity distribution application will be undertaken.
Gerhard Kraetzschmar (FORWISS Erlangen)

Communicated Beliefs and Consistency in Multi-Agent Systems

The revision of beliefs is a well-studied problem for single-agent reasoning systems. Problem solvers based on assumption-based reasoning, as used in planning and diagnosis, for instance, must deal with this belief revision problem in order to exhibit useful behavior. Most such AI systems employ reason maintenance techniques to solve the belief revision problem. The belief revision problem is much more difficult in distributed AI systems, where multiple assumption-based reasoners communicate beliefs. How can we account for communicated beliefs? How do we represent them? How do we update them? Achieving useful, coherent behavior of the multi-agent system can crucially depend on the level of consistency we are able to maintain in such a system. In the talk the problem will be presented in detail. Various approaches to deal with communicated beliefs and to maintain different levels of consistency will be discussed and systems that embody these ideas will be presented.
Henning Lobin (Universität Bielefeld)

Artificial Agents for Grounding Natural Language Semantics

In the history of Computational Linguistics there has been an emphasis on simulating language processes in independent, autonomous modules with only a small number of interfaces to other modules. The paradigmatic form of communication in this approach to NLP is the transmission of information between a sender and a receiver. As a result, phenomena of strongly situation-embedded language usage have been paid only little attention. The emergence of new concepts for the development of situated agents, which are concentrating on the close connection of percepting, reasoning, and communicating components, leads to a different view on NLP. Rather than information-giving, the instruction of an agent by a collaborating agent must be considered as the central speech act to be investigated. In this light, the semantics of an utterance has to be formulated in terms of states and internal processes of an agent. In my paper, this paradigmatic shift will be looked at in greater detail, and the basic notions of agent-oriented NLP will be discussed.
Tim Lüth (IPR, Universität Karlsruhe)

Processing Complex Tasks with Cooperating Robot Systems

The talk deals with a multi agent control architecture of a mobile two- arm robot for autonomous assembly. At first, different control methods are presented and compared, that can be used to control complex manufacturing systems. Next a functional hierarchical and a functional distributed control architecture for the autonomous robot KAMRO are discussed. The robot consists of several subsystems like manipulators, active vision and mobile platform. The advantages and disadvantages of both approaches are explained. The talk comes to a close with some problems, which have to be solved, to operate the robot in the same manner with the distributed architecture as it is possible with the centralized controller.
Ulrich Meyer (TU Berlin)

TUB-MAGIC: An Agent Architecture for Business Applications

The TUB-MAGIC architecture has been developed to design agents that can support consulting processes in business. The economical background of the two applications - financial consulting and environmental management - made it necessary to integrate concepts of intentionality that could deal with different sorts of goals, cooperation and competition as well as performing deeds and omissions. After an introduction of the applications the requirements for the architecture will be analysed and the so far developed concepts presented.
Jürgen Müller (DFKI Saarbrücken)

Using Multi-Agent Systems for Distributed Problem Solving in the Transportation Domain

The transportation domain is presented as a multiagent scenario and the use of techniques of Distributed Artificial Intelligence (DAI) for solving cooperatively the hard problems that occur in this domain are demonstrated. After a motivation and a description of the domain we address questions of cooperation between the agents, techniques for task decomposition and task allocation, and multi-agent planning and scheduling. Examples are presented that show the utility of the approach. Finally some aspects of the implementation and preliminary results are provided.
Uwe Schnepf (GMD Bonn-Sankt Augustin)

Learning Methods for Behaviour-based Navigation in Unknown Environments

In this paper (joint work with Alexander Asteroth and Mark Sebastian Fischer), we present different approaches to the problem of building behaviour-based autonomous agents which are able to adapt themselves to the characteristics of their environment. These agents are able to structure incoming sensor data on the basis of internal learning functions and statistical features of the data. They choose appropriate actions on the basis of learned relationships between sensor data, performed actions and reward given. The first approach deals with the formation of ultrasonic sensor data categories to identify novel situations using predefined or prelearned environmental sensor patterns. The second approach tackles the problem of learning real- valued functions using Q-learning. These functions are multi- dimensional mapping functions from m-dimensional continuous sensor data to n-dimensional continuous motor functions. Finally, the third approach serves to statistically structure the m-dimensional sensor space into conjunctive but separated sensory-motor fields, where each field features one optimal behaviour.
Donald Steiner (Siemens AG/DFKI Kaiserslautern)

A Flexible Agent Model Incorporating Rationality and Reactivity

An important issue in the development of multi-agent systems is that of agent behaviour: How does that which happens outside an agent influence the agent's actions? On the one hand, changes in the environment may necessitate the immediate reaction of an agent. On the other hand, an agent may not have the a-priori knowledge of how to react, and may have to resort to rational mechanisms to formulate an appropriate reaction. Rationality especially complements reactivity in dynamic environments where a particular action is not guaranteed to achieve a desired goal. Reactivity especially complements rationality when efficiency is required. We present a flexible model supporting the use of a variety of reactive and rational mechanisms to determine an agent's behaviour. In particular, this model views both reactivity and rationality in the planning paradigm: reactivity is supported by pre-determined, easily obtainable plans, rationality is supported by plan-generation mechanisms. This model encompasses the *InteRRap* agent architecture, providing a multi-layered description of an agent's behaviour. The *Multi-Agent Interaction and Implementation Language* (MAIL) supports this model by providing an efficient representation of multi- agent plans, which are used for specifying an agent's behaviour as well as for communication among agents. The implementation of MAIL forms the foundation for the *Multi-Agent Environment for Constructing Cooperative Applications* (MECCA). Further issues to be addressed are how reactive behaviour can be interrupted (due to an unforeseen sudden change in the environment), thereby forcing rational behaviour, and how plans generated via rational behaviour can be integrated into the knowledge required for reactive behaviour.
Kurt Sundermeyer (DBresearch Berlin)

The COSY Project: Methods, Tools and Applications of Agent-Oriented Techniques

The COSY project aims at providing concepts and tools for agent-oriented analysis, design, and programming. The concepts are rooted in a belief- desire-intention agent architecture. An agent's behavior is encoded in terms of behavior scripts and cooperation protocols. The tool is a development and simulation environment which provides instruments for implementing, inspecting, and observing interacting agents. The concepts have been tested for a set of more or less complicated scenarios from quite different domains. Among these is a system from the manufacturing domain by which the robustness and flexibility properties of agent-oriented approaches shall be demonstrated.
Christoph Thomas (GMD Bonn-Sankt Augustin)

Situated Agents in Domain-oriented Support Environments

My special interest is to consider agents from the interface and the user's point of view: the user is in the center, the agent supports tasks of the user. Whether the agent appears more as a user, an interface, or a system agent depends on its task. These kinds of agents are called situated agents. They can be considered as event-driven entities or processes, which can act or react on specific situations within the normal working environment of the user: they observe and analyse user's actions. These agents are able to fulfil a specific predefined (i.e. pre-programmed or learned) task for the user and they are able to support the user by suggesting modifications to the interface or to the applications. Moreover, these agents are able to support the user within a design and/or problem-solving process. The interface concept behind that idea is what Alan Kay called the indirect management: the user is engaged in a cooperative process in which human and computer agents both initiate communication, monitor events and perform tasks, instead of unidirectional interaction via commands and/or direct manipulation. The agents should do tasks that can be done while the user is doing something else and the agent should do tasks that require considerable strategy and expertise. A situated agent does not act as a mediator between the user and the (interface of an) application. The user can at all times bypass the agent: he can initiate and observe results in the application directly without being affected by anyone of the implemented agents. Questions which arise in this field ask for the general improvement of the human-computer interaction with agent-oriented or agent-based systems, ask for architectures, for presentation techniques and for the benefits users get from agents, etc.
Frank v. Martial (DeTeMobil Bonn)

Conflict Resolution among Nonhierarchical Distributed Agents

"Conflict is endemic among agents that exist in dynamic worlds, and no realistic collection of agents can be expected to have precisely complete, consistent and compatible viewpoints at all times." [Bond & Gasser 88]. In this paper, we will deal with conflicts between the plans of distributed agents. Focussing on conflicts, we address two important issues in multiagent domains: (i) how to enable individual agents to represent and reason about the actions, plans, and knowledge of other agents in order to coordinate with them; (ii) how to enable agents to communicate and interact for conflict resolution. We have developed a model which integrates both the agent's knowledge about each other, their planning process, and their communication. We will show how agents can transfer their intended actions (plans), how they can use this knowledge to detect conflicts and coordinate themselves, and how this is connected with communication. Our coordination framework will be illustrated in the traffic domain, where it is the task of autonomous vehicles to coordinate their intended routes to avoid collisions.
Ipke Wachsmuth & Yong Cao (Universität Bielefeld)

Agent-mediated Verbal Interaction with a Virtual Environment

The overall goal in the VIENA project is to enable an intelligent communication with a technical system for the interactive manipulation of a virtual environment. In an example domain of interior design, we use simple verbal communication to manipulate scene objects, e.g., to change their positions or colors, or to modify the overall scene illumination. Our aim is to keep the user (designer) free from technical considerations such as planning of geometric details, etc. To make this possible, we are developing a set of agents which altogether form an intelligent mediator. This mediating 'agency' takes qualitative verbal instructions and translates them to quantitative commands that are used to update the visualization scene model. The VIENA parser translates an instruction to an internal deep-level representation which outputs to the mediating agents. A bookkeeping agent is authorized to access and modify the augmented graphics data base and to supply current situation information to agents on request. A space agent translates qualitative relations such as 'left of' to appropriate scene coordinates. Other agents, in similar ways, take special responsibilities in mediating an instruction. Agents cooperate to offer a goal scene corresponding to a user's inquiry. The offer can be changed in further interaction, that is, the user can negotiate the computed semantics of qualitative verbal instructions.


Ipke Wachsmuth, last updated 2000-09-27