Preserving Simulation Insight While Removing Data: Verification of Compressed Simulation Traces via Machine Learning

My Nguyen, Duc Vu, Anh Vo, Luke Liang, Philippe J. Giabbanelli

Abstract: In simple cases, a stochastic network simulation is verified by comparing the distribution of its outputs with an expected ground truth distribution. However, there are cases where the two distributions are indistinguishable, hence practitioners would believe that their code is correct whereas it has implementation errors. We recently showed that simulation traces could be converted into images so that problematic patterns could be detected by machine learning (ML) algorithms. However, using imagification to represent the outputs of a simulation required a massive amount of storage, which prevented the practical deployment of this ML-based verification system. In this paper, we show that off-the-shelf image compressions algorithms can decrease space consumption while preserving accuracy for three network simulations (SIR, rumor spread, cascading failure). Depending on the compression level, data can consume less than 25% of the original space and the algorithms that reorder the compressed images preserve the ability to detect errors.