AMaze: a benchmark generator for sighted maze-navigating agents

Abstract

The need to provide fair comparisons between agents, especially in the field of Reinforcement Learning, has led to a plethora of benchmarks. While these are devoted to tailor-made problems, they offer with very little degrees of freedom for the experimenter. AMaze is instead a benchmark generator capable of producing human-intelligible environments of arbitrarily high complexity. By using visual cues in a maze-navigation task, the library empowers researchers across a large range of fields.