New Fuzzing Biases for Action Policy Testing

• DOI: 10.1609/icaps.v34i1.31472
• Publication Date: 2024-05-30T13:15:42Z
• Authors: Jan Eisenhut; Xandra Schuler; Daniel Fišer; Daniel Höller; Maria Christakis; Jörg Hoffmann
• Keywords: Fuzz testing; TRACE (psycholinguistics)

Testing was recently proposed as a method to gain trust in learned action policies classical planning. Test cases this setting are states generated by fuzzing process that performs random walks from the initial state. A bias attempts these towards policy bugs, is, where sub-optimally. Prior work explored simple based on policy-trace cost. Here, we investigate topic more deeply. We introduce three new biases analyses of shape, estimating whether trace is close looping back itself, it contains detours, and its goal-distance surface does not smoothly decline. Our experiments with two kinds neural show improve bug-finding capabilities many cases.