AGENT SYSTEMS THAT NEGOTIATE AND LEARN

Agents for office automation purposes can be seen as an extension of today's office software (word processors, spreadsheets). They should not only support single tasks, but assist their users throughout complex workflow procedures with many persons involved, if possible in much the same way as human secretaries do. For example, such programs might assist a worker in scheduling meetings, flagging important electronic mail, processing purchase orders, etc. This concept seems clearly attractive: it could make secretarial assistance available to everyone within an organization and make such assistance mobile (on notebooks and PDAs). This paper focuses on two major features which determine the success of secretarial software. First, the assistance program must be able to negotiate (with other agents as well as with humans), because most office tasks contain interaction among several people. Second, it must be able to learn. It has to learn how to adapt to its users' idiosyncrasies (and not vice versa), since people tend to develop individual work techniques and styles. It also has to learn how to adapt to specific workflows that can differ substantially from organization to organization. To ensure this adaptability we propose-similar to the way human secretaries are trained-a learning-by-being-told approach. Mechanisms for negotiation and learning have been included in secretarial agents for calendar and room management. In particular, the architecture, functionality, and capabilities of the calendar apprentice CAP II and the room reservation apprentice RAP will be described.