Next: Federated Peer Learning Up: Negotiated Retrieval Previous: Negotiated Retrieval Algorithm

CBR-TEAM: A multi-agent design system

In this section, we present a brief summary of the CBR-TEAM system[Nagendra Prasad, Lesser, &Lander1996] that uses negotiated retrieval to compose coherent design cases. Note however, that our experience with negotiated retrieval is still rudimentary and we will be able to give further insights into its effectiveness in future.

CBR-TEAM, whose core is derived from TEAM[Lander1994], is a parametric design system that uses a set of heterogeneous cooperative agents for designing steam condenser components. It consists of three agents - motor-agent, pump-agent and vbelt-agent - that are responsible for the design of the motor, pump and vbelt components of a steam condenser. The user gives a problem specification that consists of minimum head size for the pump in the required design. The agents in CBR-TEAM retrieve and use suitable members from libraries of manufacturer-specified models for designing their components and use the negotiated retrieval strategy to arrive at mutually acceptable designs. When the components of the individual agents are being assembled, violation of constraints due to mismatches on shared parameters lead to information exchange followed by redesign. Interface parameters are those features of a component that are shared by more than one agent. All the relevant agents have to reach an agreement on the values of the shared parameters.

During the initial phase of retrieval, the agents may have only partial information on the requirements of other interacting components. So, each of the agents chooses the lowest cost component based on the information available to it. Trying to assemble these components into an overall design may lead to conflicts due to mismatches in the parameters that are shared by two or more components. For example, motor and pump components have required-pump-power as a shared parameter and both motor-agent and pump-agent impose their own set of constraints on this parameter. A mismatch on this parameter involves one agent assigning a value to the parameter that violates the constraints in another agent. When a conflict is detected, the agent detecting it sends feedback to the other agents involved. In CBR-TEAM, feedback involves communicating the relevant violated explicit constraints, all of which are single-parameter numerical-valued constraints. When an agent receives feedback from others, it assimilates the feedback. Assimilation involves adding the feedback constraints to the set of local constraints to effect further searches from there on. The agents iteratively perform further rounds of retrieval using the previous information and the new requirements from other agents to get better cases to be assembled into a design that does not produce the same conflict.

Next: Federated Peer Learning Up: Negotiated Retrieval Previous: Negotiated Retrieval Algorithm
Mon Sep 16 17:23:45 EDT 1996