Monitoring Complex Systems with Domain Knowledge: Adapting Contextual Bandits to Tracing Data
|Termin||Fr 12. Februar 2021|
|Kurzfassung||Monitoring in complex computing systems is crucial to detect malicious states or errors in program execution. Due to the computational complexity, it is not feasible to monitor all data streams in practice. We are interested in monitoring pairs of highly correlated data streams. However we can not compute the measure of correlation for every pair of data streams at each timestep.
Picking highly correlated pairs, while exploring potentially higher correlated ones is an instance of the exploration / exploitation problem. Bandit algorithms are a family of online learning algorithms that aim to optimize sequential decision making and balance exploration and exploitation. A contextual bandit additional uses contextual information to decide better.
In our work we want to use a contextual bandit algorithm to keep an overview over highly correlated pairs of data streams. The context in our work contains information about the state of the system, given as execution traces. A key part of our work is to explore and evaluate different representations of the knowledge encapsulated in traces. Also we adapt state-of-the-art contextual bandit algorithms to the use case of correlation monitoring.