Suggesting and Justifying Model Updates for Improved Troubleshooting

摘要

For the troubleshooting domain, machine learning is used to suggest model updates using in-service troubleshooting and component testing records. One of the challenges of using these updates is how to justify each change to the system experts; a novel approach for the justification of updates to a Bayesian network troubleshooting model is presented. The results of experiments into the performance of the approach suggest that the changes suggested by maximum likelihood learning improve a model by fitting to the new records and a good justification for each change can be obtained from quantities already computed during the learning process.