Automating Web-based Infrastructure Management via Contextual Imitation Learning

Abstract

Web-based management dashboards provide intuitive interfaces to support various infrastructure management tasks. However, the lack of programmability in these dashboards makes it difficult to automate management tasks. To address this limitation, we propose a system call MAIL (Management Automation through Imitation Learning) that learns from an operator's demonstrations using imitation learning to automate management tasks on web-based dashboards. A novel Contextual Imitation Learning algorithm is proposed in MAIL to overcome the learning challenges brought by the requirements of infrastructure management automation. The effectiveness of the MAIL system has been demonstrated using both synthetic environments and production environments. It outperforms existing methods and is able to achieve an 86%-100% success rate for most tasks.