A non-hierarchical method for distributed hypothesis management

We describe an approach to the distributed management of multiple hypotheses within a society of agents (or peer groups) without need of a centralized controller. The distributed architecture is based on two key concepts: (1) efficient assignment, ranking and association of local hypotheses at each peer, and (2) a market-based information sharing model for exchanging hypotheses within and across peer groups. We will describe algorithms and discuss performance issues for the first issue. We will present an economic model, informative advertising, which is applied to the second topic. We will discuss necessary conditions for achieving consistency (consensus) of hypotheses within the peer group. We will describe an efficient implementation of the architecture, based on the peer-to-peer (P2P) messaging protocols from project JXTA. We will describe how directed service advertisements are used to dynamically create relationships between peers based on shared goals and contextual information. This scheme allows patterns of interaction to emerge based on the data sensed rather than on pre-staged processing pipelines. We also show how the intelligent management of hypotheses leads to flexibility in resource usage allowing peers to dynamically adapt to shortages.