Efficient learning and analysis of system behavior

Abstract

In this thesis, we present techniques for more efficient learning and analysis of system behavior. The first part covers novel algorithms and tooling for testing systems based on active automata learning and Linear-time Temporal Logic (LTL) model checking, also called Learning-Based Testing (LBT). Next, we provide an improved learning algorithm that is able to deal with huge alphabets. These are commonly seen in large-scale industrial systems where input symbols contain data parameters. In the second part we discuss improvements for analyzing formal system specifications. We start out by looking at separated read, write and copy dependencies for symbolic model checking to speed up the verification of these specifications. Then, we show that bandwidth reduction techniques, originally designed for sparse matrix solvers, are very capable at reducing the memory footprint of the specifications' symbolic state space. Implementations of the presented algorithms are subjected to case studies and rigorous experimentation with scientific software competitions.