We elaborate on such related work in that we translate and extend the basic SFL concept to accommodate also dynamic, live settings. Our concept thus allows us to continuously evaluate the success-rates of system actions via considering the success/failing of previously executed action sequences a.k.a. plans. We continuously update a corresponding reliability measure for each action, and use these data to (a) re-evaluate currently implemented action plans, and (b), when reasoning about new plans for achieving future goals. We require only very limited data for our concept, i.e., which sequences failed or succeeded in the past and which components were involved. From a technical perspective, we describe a system's actions via specific rules, pre- and postconditions and use these data for our reasoning.
While we do not isolate an issue's exact source(s) (like with MBD), we will show in our evaluation that our compromise between preciseness and computational complexity allows us to dynamically, effectively and efficiently cope with faults and other events that result in unreliable actions. As we will discuss, we rely solely on SFL for our reasoning and thus passively observed executions without deploying any exploratory component like in reinforcement learning.