An empirical evaluation of maximum likelihood voting in failure correlation conditions

The maximum likelihood voting (MLV) strategy was recently proposed as one of the most reliable voting methods. The strategy determines the most likely correct result based on the reliability history of each software version. However, the theoretical results were obtained under the assumption that inter-version failures are not correlated by common cause faults. We first discuss the issues that arise in practical implementation of MLV, and present an extended version of the MLV algorithm that uses component reliability estimates to break voting ties. We then empirically evaluate the implemented MLV strategy in a situation where the inter-version failures are highly correlated. Our results show that, although in real situations MLV carries no reliability guarantees, it tends to be statistically more reliable, even under high inter-version correlation conditions, than other voting strategies that we have examined. We also compare implemented MLV performance with that of Recovery Block and hybrid Consensus Recovery Block approaches. Our results show that MLV often outperforms Recovery Block and that it can successfully compete with more elaborate Consensus Recovery Block. To the best of our knowledge, this is the first empirical evaluation of the MLV strategy.