Recovery Process Visualization based on Automaton Construction

Since information and communications technology systems become large and complex, artificial intelligence (AI) technologies for automatic failure recovery have recently been developed. While AI-based failure recovery is promising, it raises an issue of interpretability, i.e., the problem in which operators cannot understand the target system behavior during the recovery process. To overcome this issue, we propose a method for constructing human-interpretable automata that describe the recovery process for each type of failure in accordance with the concept of proactive operation including chaos engineering. We define a new class of automata called observation-labeled finite automata (OLDFAs) with an observation function, which enable us to express recovery processes of the target system that can output high-dimensional data in a mathematical form. Our method constructs an OLDFA consistent with the target system on the basis of the L
^*
algorithm and clustering techniques. We also propose a method of creating a recovery process workflow (RPW) by merging OLDFAs for distinct failures. The RPW makes visible to operators what types of failures the target system is likely to currently have after specific actions have been taken and observations obtained. This would be beneficial in that operators can understand what AI-based operation takes the actions for and what states the target system is likely to be in during the recovery process. To evaluate the effectiveness of our methods, we carried out numerical simulations and experiments with container-based systems, showing our methods successfully construct automata and that the created RPW is interpretable.