Modern manufacturing systems have to deliver products to increasingly competitive and volatile markets. These markets force companies to shorten product-life cycles, reduce time-to-market, increase product variety, and instantly satisfy demand, while maintaining high quality and reducing investment costs. This, in turn, has two major consequences for the manufacturing systems. First of all, manufacturing systems must become more flexible in order to cope with constant product changes and an increasingly volatile demand. Secondly, manufacturing systems must exhibit more robustness with respect to disturbances in order to maximize the total use of the manufacturing equipment, and thus to further reduce the overall investment costs. Flexible and robust production systems, however, require an intelligent control system that makes efficient use of the (hardware) flexibility. Classical control technology is no longer sufficient because it has been developed to optimize a fixed process. Modern manufacturing systems thus will have to continuously adapt their process to changing production conditions.
Software agents are just the right information technology to meet the above challenges of modern manufacturing systems. They model the manufacturing process as it is with no artificial central control unit. Resources are allocated dynamically by a continuous coordination process among the relevant agents. Unlike in Computer Integrated Manufacturing, there is no need to handle all the contingencies of a complex manufacturing process at design time; rather, agents negotiate proper allocations among themselves during execution. Although some of their joint decisions may not be optimal, all decisions are, nevertheless, made on the basis of the current situation – in the long run leading to a better system performance.
Many agent-based approaches have been proposed to achieve the vision of agent-based manufacturing; the earliest dating back to the middle of the eighties. Probably the very first application of agent-oriented concepts to manufacturing control was the prototype factory control system YAMS of Parunak et al. [Parunak 1985] [Parunak 1986]. In YAMS, the manufacturing enterprise is modelled as a hierarchy of production units. The hierarchy, however, records only composition, not control. Task distribution down the hierarchy is done through a negotiation process that is based on the contract-net protocol [Smith 1980]. A node announces a task, units capable of performing the task reply with a bid, and the node in turn assigns the task to the unit with the best bid. Because of this negotiation process between superior and subordinate units, the assignment process is able to take into account any changes or disturbances that may occur during the planning process. In parallel to the work of Parunak et al., Shaw and Whinston [Shaw 1985] also developed a distributed scheduling method based on the contract-net protocol. As in YAMS, manufacturing tasks are contracted out to workcells through a bidding process in which the workcells bid for the tasks. Other early work on agent-based manufacturing systems was proposed in [Duffie 1987], [Duffie 1988], [Tilley 1992] and [Maturana 1996]. In particular Duffie investigated the limits of centralized and hierarchical control architectures and proposed heterarchical control systems, i.e., control systems without centralized of hierarchical control [Dilts 1991].
To date only few agent-based manufacturing system have really been implemented on the shop-floor. One is the system Production 2000+ from Daimler which is described here.