GENSYNTH: Synthesizing Datalog Programs without Language Bias

Techniques for learning logic programs from data typically rely on language bias mechanisms to restrict the hypothesis space. These methods are therefore limited by user's ability tune them such that space is simultaneously large enough include target program but small admit a tractable search. We propose technique learn Datalog input-output examples without requiring user specify any bias. It employs an evolutionary search strategy mutates candidate and evaluates their fitness using off-the-shelf interpreter. have implemented our approach in tool called GenSynth evaluate it diverse tasks knowledge discovery, analysis, relational queries. Our experiments show can correct few examples, including require recursion invented predicates, robust noise.