Reinforcement Learning for Reliability Optimisation

Software Reliability Optimization problem is aimed at bridging the reliability gap in an optimal way. In an industrial setting, focussed testing at the component level is the most favored solution exercised to fill the reliability gap. However, with the increased complexity in the software systems coupled with limited testing timing constraints finding an optimal set of test suite to bridge the reliability gap has become an area of intense research. Furthermore, the stochastic nature of the reliability improvement estimates associated with each test suite manifolds the complexity. Here, we propose Reinforcement Learning (RL), as a mechanism to find an optimal solution. We have shown how an interactive learning is used to estimate the true outcome of the selection and the action selection policy so as to maximise the long term reward. The estimation methodology and the selection policy is inspired by Multi-armed bandit solution strategies. Firstly, we employ a sample average estimation technique for deriving the true outcomes. Secondly, a variant of simple greedy algorithm coined as epsilon-greedy algorithm is considered for action selection policy. These two steps are iteratively exercised until the selection criteria converges. The efficacy of the proposed approach is illustrated on a real time case study.