From Formal Workflow Models to Intelligent Agents

Todays business workflows, e.g. in administration, have a strong need for support by information technology, because they heavily rely on the efficient and fluent interaction of human actors and artificial systems. At the same time, workflow tasks supported or automatically executed by computers get more and more complex. Intelligent agents try to cope with both of these settings by providing a Software Engineering abstraction, which incorporates complex functionality for task planning and execution with elaborate interaction capabilities. This paper presents a method for identifying and specifying such agents and their behavior by careful analysis of formal workflow models. It follows a small example out of an industrial-size case study, which is currently conducted at Europe’s biggest hospital, Charit~ Berlin. There are at least two aspects of todays business workflows, which have to be considered for support by information technology. The first aspect is the partially or fully automated execution of workflow tasks. These tasks get more and more complex and enforce the use of sophisticated methods, including planning, scheduling and constraint satisfaction. There are various successful contributions from Artificial Intelligence on these topics (rf. e.g. to (Allen et al. 1990; Zweben and Fox 1994; Yokoo et al. 1998)). The second aspect is the control of the task execution process itself. Since most realworld tasks and their execution are distributed spatially and among individuals, the coordination and interaction between the task execution entities is central to the management of business processes. Workflow Management has made successful contributions to this field since the beginning of 1990 (rf. e.g. to (Lawrence 1996; Jablonski 1995)). The notion of an Intelligent Agent (rf. e.g. to (Huhns and Singh 1998; Jennings and Wooldridge 1998)) is recent concept, which tries to incorporate the merits of classical AI methods with strong communication and interaction capabilities. Hence, it seems to be predestined for application to workflow management, especially under the assumptions proposed by HEWITT’s Open Systems (Hewitt 1988). Unfortunately, it is not simple to identify possible agents in a given business context and to specify their behavior. Should agents represent certain individuals, roles or even organizational units in the business domain or should they rather implement representations of tasks, processes and resources? In my opinion, these questions cannot be answered generally but only in the context of the given application domain. Nevertheless, I will present a method, which describes the identification and the specification of agents and their properties by a careful analysis of formal workflow models. This method can be applied to a wide range of applications, which involve distributed process control. The term "agent" can be interpreted as being a new abstraction in Software Engineering, just like objects. Even pioneers of object-oriented technology are aware of the usefulness of intelligent, goal-directed entities. A. COCKBURN states in (Cockburn 1997), "I hope that one day we shall have more formal tracking of goals and backup goals.", and I. GRAHAM Can be quoted, "Clearly the goal or contract is in the mind of some agent rather than being something possessed by the task itself." Nevertheless, there has been a lot of confusion about this term, especially outside the Distributed Artificial Intelligence community. An interesting survey on possible agent definitions is given in (Franklin and Graesser 1996). Throughout this paper, I will use the term "agent" to denote an artificial representative, which autonomously acts in behalf of a human or organizational actor in the application domain. For presentation, I will follow a small example out of an industrial-size case study. This case study is explained in the first section. In the following section a brief description of the used workflow model is given. The next section presents the method for identification and specification of agents and their skills. Concluding remarks follow in the last section.