CAPRI: A Common Architecture for Autonomous, Distributed Diagnosis of Internet Faults using Probabilistic Relational Models

Internet fault diagnosis today is slow, costly, and error-prone because it requires humans to run diagnostic tests and interpret their results. A fully autonomous self-diagnosing network could greatly improve diagnostic accuracy and efficiency, but such a network requires a common language for expressing diagnostic knowledge and data, and a protocol for distributed probabilistic diagnostic reasoning. In this paper I show how the Common Architecture for Probabilistic Reasoning in the Internet (CAPRI) can satisfy these requirements using probabilistic relational models (PRMs). Preliminary results indicate that CAPRI agents can diagnose HTTP proxy connection failures with over 80% accuracy using TCP failure data collected using an updated version of Planetseer[11].