Three research challenges at the intersection of machine learning, statistical induction, and systems
Moisés Goldszmidt, Ira L. Cohen, Armando Fox, Steve Zhang · 2005
results for performance debugging and failure diagnosis and detection in systems by using approaches based on automatically inducing models and deriving correlations from observed data. This paper explores research questions and preliminary results regarding the next steps in advancing this line of work. We specifically formulate three challenges. First, as new data is collected from a system, there is a need to continuously assess the validity of previously induced models, with the ultimate aim of achieving online adaption to system changes. The second challenge deals with interactions with human operators, including interpretation of model findings to generate explanations, enabling operator feedback to improve the models, and handling false positives and missed detections. The third challenge focuses on transforming the output of these models into structured representations or signatures of system state, so that they can serve as a machine-readable index for diagnosis and repairs. These challenges arise directly from the application of statistical and machine learning techniques to real-world systems, and we contend that through domain knowledge of this same application we may identify well-engineered solutions that allow the full potential of statistical and machine learning approaches to be realized. 1