Benchmarking Łukasiewicz Logic Solvers with Properties of Neural Networks

We propose new benchmarks for problems related to Łukasiewicz Infinitely-valued Logic and discuss methods for generating them. Such benchmarks comprehend instances that state properties about neural networks. In particular, reachability properties yield satisfiability instances and robustness properties yield logical consequence instances. We also present the results of empirical experiments where the proposed benchmarks were run in solvers based on SMT and MILP technologies. In this way, we are able, on the one hand, to compare the performance of different solvers and, on the other hand, to delve deeper into the investigation of neural network formal verification.