Batch Repair with Heuristic Search

Troubleshooting is the task of fixing an abnormally behaving system. Many troubleshooting algorithms employ a diagnosis engine that returns a set of diagnoses Ω and their likelihoods (p(·)), where a diagnosis ω ∈ Ω is a set of components that may have been faulty and its likelihood p(ω) is the probability that ω is all the faulty components in the system. The Batch Repair Problem (BRP) is a recently introduced troubleshooting problem in which Ω and p(·) are given as input and the task is to minimize the total repair costs incurred until the system is fixed (Stern, Kalech, and Shinitzky 2016). Unlike traditional troubleshooting problems, the total repair costs considered in BRP includes (1) the cost of repairing each faulty component (referred to as the component repair cost) and (2) additional repair overhead costs (denoted costrepair). When the repair overhead is more expensive than the component repair cost then it may be more efficient to repair a batch of components in a single repair action. BRP can be modeled as a Markov Decision Process (MDP), where the actions are the possible repair actions and the uncertainty is due to not knowing which components are faulty (and thus, repairing a set of component may or may not fix the system). However, the number of actions in each state is exponential in the number of components, as any batch of components may be considered. As an alternative, prior work modeled BRP as a combinatorial optimization problem, searching in the combinatorial space of possible repair actions for the best repair action to perform next (Stern, Kalech, and Shinitzky 2016), where best is measured according to a utility function that estimates current and future wasted repair costs. There are two challenges in implementing this approach: (1) how to define such a utility function, and (2) how to efficiently search for the repair action that maximizes this function.