An extended Bayesian belief network model of multi-agent systems for supply chain management

This dissertation develops a theoretical model, called an extended Bayesian Belief Network (eBBN), of a Multi-agent System for Supply Chain Management (MASCM), which formalizes agent interactions in uncertain environments.; MASCM is an electronic marketplace as well as a supply chain management system where agents sell and buy products on behalf of their owners to gain profits. A virtual chain consists of agents connected by commitments triggered by an end order. The system performance is measured by whether the management goal, e.g. end customer satisfaction, shared by all virtual chains can be reached.; Due to the uncertain nature of internal and external decision factors, a commitment made by an agent may eventually not be fulfilled. Uncertainty concerning one agent's commitments may propagate over the chain via its supplier-customer connections. Uncertainty and its propagation may have negative impacts on agents' operations, cause inventories to be increased, the chain to be disturbed or destroyed, and eventually end orders to be delayed.; To reduce potential damage from uncertainty, agents may choose to cooperate with each other by sharing information. This type of agent interaction in uncertain environments is formalized as eBBN, in which the effects of uncertainty are modeled as agents' beliefs about the failure of commitments, relationships between these beliefs as direct causal links, and information sharing as belief update and propagation. By properly incorporating actions and their consequences into the network, eBBN further extends the representation and inference capability of traditional Bayesian Belief Networks (BBNs). The model can not only reason about the effects of agents' strategic behaviors in updating beliefs but can also describe dynamic causal structures as virtual chains evolve over time.; As a formal model, eBBN provides a sound basis for developing effective algorithms of uncertainty management. It can serve as an analytic platform to quantitatively study the relationship between agents' local behaviors and overall system performance in an uncertain environment. Several algorithms for both local decisions and global optimization have been developed and tested. The simulation results present that the system with agents using these algorithms can achieve stable performance even when uncertain events occur with high frequency.