Learning with curricula for sparse-reward tasks in deep reinforcement learning

Abstract

From an early age, humans spend a great deal of time playing in and exploring in their environments. We don't go from zero to AlphaZero without stopping to learn many other things along the way, and we don't learn these things alone. In many human societies, schooling and culture guide learning by providing a curricula for what is considered "intelligent" behavior. In this work I demonstrate how drawing from a curriculum developed to coax apes into successfully learning tasks can also improve performance of artificial agents, particularly in sparse-reward scenarios. I also demonstrate where curriculum learning falls short, and what these experimental results suggest for efforts in developing human-like artificial intelligence.