Reinforcement Learning Based Framework for Real Time Fault Tolerance

Smart autonomous systems are system that take decisions independently and on run time without the need for human interaction. One of the most important components that plays a big roll in system autonomy is fault tolerance and avoidance which is a basic capability that autonomous systems should have in order to be able to survive surrounding environment state change without the need for supervision. In this paper, a framework is proposed where fault tolerance and avoidance is achieved through a reinforcement learning based framework. The framework adapts to changes and learns new processes for fault tolerance and avoidance whenever environment states change. The framework has a set of predefined actions and an observable environment. Reinforcement learning is being applied in order to learn the sequence of actions that needs to be taken to avoid or tolerate failure. The outcome of the learning process is a sequence of actions to help the system reach a desired state while avoiding fault states. These are used for later execution when the same situation occurs when the agent is in similar environment state while having the similar readings. Two Theorems and a Lemma are proposed to define the validity and correctness of the framework. The proposed framework is then simulated and tested.