ChemGrid: An Open-Ended Benchmark Domain for an Open-Ended Learner

This paper introduces ChemGrid, which is a novel video-game-based domain for exploring open-ended artificial intelligence capable of setting arbitrary goals for itself. In this domain, agents must satisfy a minimal goal of crafting a preset “survival molecule” but otherwise can pursue their own agendas related to crafting in an artificial chemistry based on the concept of pathway complexity from artificial life. In particular, agents can discover new molecular recipes by joining existing recipes together and breaking them apart. This paper explores the design space afforded by the artificial chemistry. The main contribution is the introduction of ChemGrid itself, which is intended as a future benchmark domain for general game-playing algorithms.