Iterative Software Engineering for Multiagent Systems

Although agents and agent-based computing have been an active research area for many years, it is only until now that these topics begin to gain industrial relevance: agent technology is beeing recognized as a powerful tool for the development of large and complex systems. These days, typical software architectures contain many dynamically interacting components, each with their own thread of control and engage in complex coordination protocols [Ciancarini and Wooldridge, 2000]. Therefore, a new programming metaphor that captures these systems is needed. Although the basic structural elements of the agent-based approach as well as their connections are not yet fully understood, it nonetheless seems to be a promising means for dealing with these highly complex systems.

In this chapter, I will outline the basic characteristics of intelligent agents and multiagent systems (MAS) and their relation to Software Engineering issues. I will start with an introduction to intelligent agents and then extend the single-agent case to systems with several intelligent agents. The basic concepts are still presented from a very general point-of-view and will be refined in subsequent chapters; the reader already familiar with MAS may safely skip this chapter and continue with chapter 3; a more thorough introduction to the field for the unfamiliar reader can be found in [Weiss, 1999]. After the general introduction, I will then briefly discuss some related research fields and finally provide a personal view on various topics of the relatively new research field of “agent-oriented software engineering” [Jennings et al., 1998].

The software development method that is presented in this book is built upon a number of standard Software Engineering concepts and combines them effectively into a single, coherent model. In this chapter, I will introduce these basic concepts and their relationships by starting with an investigation of a cognitive model of design in general. From this general model, I will derive some basic skills that are essential for a successful software engineer and I will outline some requirements for software engineering tools and methods that follow from these considerations.

The ideas and concepts of the previous chapter constitute the basis of general Software Engineering methods which will now be refined for the particular case of multiagent system development. To this end, I will now explain the basic concepts and ideas of the Massive method how its building blocks are assembled into a coherent method that can be used to develop multiagent applications.

The product model of Massive is the core of the entire method. It allows the system designer to break the target system down into several views that concentrate on particular aspects of the system and abstract away from others. In each of the following sections, we will at first discuss the general nature and the intended scope of a view as well as a number of features and design patterns that belong to a view. However, the features and patterns that are presented are by no means an exhaustive collection, they rather represent the current state of my personal experience in designing multiagent applications. According to the basic idea of the Experience Factory, the potential user of the Massive method is encouraged to add new aspects that are necessary or to remove aspects that are not important in a particular context. In each of the following sections, the general considerations are applied to a case study in order to demonstrate how the theoretical concepts are used in a practical situation.

A development method will never be accepted in an industrial context if it cannot prove its validity in practice. The Massive method is not a method that was developed in the laboratory and then transfered to actual projects. Rather it is derived from projects that were successfully carried out at the DFKI and elsewhere and that were analyzed after completion in order to find similarities in the product and process models. The advantage of this approach is that it provides further case studies that show how the method works and that demonstrate that the method can be used for a broad range of multiagent applications.

The basic properties of an ideal software development method for multiagent systems were defined in the introduction: the method should be flexible, open, simple, scalable, support learning and reuse, have little institutional overhead and relate to established software engineering concepts.

In this chapter, I will review three tool-kits for the development of multiagent applications. These packages usually offer predefined structures or even tools that can be used by the application developer to build his or her own system using the provided structures that are common to most multiagent applications. All tool-kits discussed in this section are freely available and result from research work in the area, commercial frameworks have not been considered for this section.

The planning process for a single task completion consists of three steps: 1. Finding the shortest path between the source and the destination node. 2. Forward propagation of arrival times. 3. Backward propagation of departure times.