Abstract. The AgentScape middleware is designed to support deployment of agent-based applications on Internet-scale distributed systems. With the design of AgentScape, three dimensions of scalability are considered: size of distributed system, geographical distance between resources, and number of administrative domains. This paper reports on the AgentScape design requirements and decisions, its architecture, and its components. 1

We present a new failure detector implementation. This implementation, a variant of the heartbeat failure detector, is both adaptable and designed for scalability. Its first specificity lies in the fact that it is designed as a shared service among several applications by way of an adaptation layer. This layer adapts the quality of service according to application needs. The second specificity is the hierarchic organization of the detection service: it allows to decrease the number of messages and the processor load. Through an experimentation evaluation, we show that our implementation is adaptable to the environment characteristics and usable with large scale applications.

This paper describes a security architecture for the Mansion mobile agent system. Mansion is a logical framework designed to support large-scale heterogenous mobile agent applications. Mansion is implemented as a multilayered middleware system in which the lowest layer provides functionality that is common to most mobile agent systems and higher layers become increasingly application aware. The security architecture presented in this paper provides secure agent communication, secure mobile agent transport and startup, and secure auditing of all changes made to agents. The system uses self-certifying names for authenticating principals in the system and provides mechanisms to control information flow.

This paper describes and empirically evaluates the middleware platform of a new network architecture called the Bio-Networking Architecture. The Bio-Networking Architecture is inspired by the observation that the biological systems (e.g., bee colonies) have already developed mechanisms necessary to achieve future network requirements such as autonomy, scalability, adaptability, and simplicity. In the Bio-Networking Architecture, a network application is implemented as a group of distributed, autonomous and diverse objects called cyber-entities (CEs) (analogous to a bee colony consisting of multiple bees). Each CE implements a functional service related to the application and follows simple behaviors similar to biological entities (e.g., reproduction and migration). In the Bio-Networking Architecture, beneficial application characteristics (e.g., autonomy, scalability, adaptability, and simplicity) arise from the autonomous interaction of CEs. The middleware platform in the Bio-Networking Architecture, the bionet platform, provides reusable software components for developing, deploying, and executing CEs. The components abstract low-level operating and networking details, and implement high-level runtime services that CEs use to perform their services and behaviors. The components in the bionet platform are designed based on several biological concepts (e.g., energy exchange and pheromone emission). This paper describes key designs of the bionet platform and empirically demonstrates that the bionet platform is efficient, scalable, reusable, and significantly simplifies development of network applications.

Abstract. Sensor Web is a revolutionary concept towards achieving a collaborative, coherent, consistent, and consolidated sensor data collection, fusion and distribution system. Sensor Webs can perform as an extensive monitoring and sensing system that provides timely, comprehensive, continuous and multi-mode observations. This new earth-observation system opens up a new avenue to fast assimilation of data from various sensors (both in situ and remote) and to accurate analysis and informed decision makings. So far, most of the sensor networks work in isolated environments rather than being integrated into a higher system of environmental systems. Our vision of the Sensor Web is an open infrastructure that allows end users to automatically access and extract and use appropriate information from multiple sensor sources over the Internet. This paper proposes an architectural framework, the Sensor Web Agent Platform (SWAP) that makes use of two of the most promising distributed architectural paradigms i.e. Web Services and Multi Agent Systems. SWAP allows for integrating arbitrary sensors or sensor networks into a loosely coupled higher level environment that facilitates developing and deploying end user applications across multiple application domains. An ontology framework and an abstract hierarchy are proposed for fusing and integrating data and storing and re-using generic information extraction techniques within the system. 1

Agents, and in particular mobile agents, offer a means for application developers to build distributed applications. In current agent systems, mobility of agents is constrained by the environment of the agents: the agent platform (which supports agents) and the agent's code base (e.g., DESIRE, Java). Generative migration is needed to adapt an agent to conform to its destination agent platform and code base. In this paper generative migration is described as a process of "transparently adapting" an agent. An agent can continue to function at its new location on a completely different agent platform.

Performance measurement of large distributed multiagent systems (MAS) offers challenges that must be addressed explicitly in the agent infrastructure. Performance data is widely distributed and voluminous, and poor data collection can impact the operation of the system itself. However, performance metrics are essential to internal system function, e.g., autonomous adaptation to dynamic environments, as well as to external assessment. In this paper we describe the tools, techniques, and results of performance characterization of the Cougaar distributed agent architecture. These techniques include infrastructure instrumentation, pluginbased instrumentation of agents, and dynamic control of metric collection. We introduce multiple redundant “channels ” for metric delivery, each serving separate quality of service requirements. We present our techniques for instrumenting the agent society, justify the metrics chosen, and describe the tools developed for collecting these metrics. We also present results from distributed agent societies comprising hundreds of agents. Categories and Subject Descriptors

Large scale open, heterogeneous, distributed environments such as the Internet, are the environments in which (intelligent) agents need to be able to function and survive. These environments need to provide distributed support, including management services, for such agent systems. In this paper a local management architecture, implemented in AgentScape, is provided together with a management-oriented life cycle model. A major feature of this model is the central role of one of the states of the life cycle model, namely the "suspended" state: the state in which an agent is manageable. A prototype implementation of the management system based on the life cycle model is described.

Abstract. The Agent Operating System (AOS) provides the basic functionality needed for secure and reliable mobile agent platforms: support for secure communication, secure agent storage and migration, and minimal primitives for agent life-cycle management. Designed as a layer between local operating systems and higher level agent platform middleware, it supports interoperability between agent platforms and between different implementations of AOS itself. AOS has been tested on interoperability, both with regard to different higher-layer middleware platforms and interoperability between two implementations of AOS in C++ and Java. 1

Open multi-agent systems need to cope with the characteristics of the Internet, e.g., dynamic availability of computational resources, latency, and diversity of services. Large-scale multi-agent systems employed on wide-area distributed systems are susceptible to both hardware and software failures. This paper describes AgentScape, a multi-agent system support environment, DARX, a framework for providing fault tolerance in large scale agent systems, and a design for the integration of the two.