When considering how to reduce the learning effort required for Reinforcement Learning (RL) agents on complex tasks, designers can apply several common approaches. Reward shaping boosts the immediate reward provided by the environment, effectively encouraging (or discouraging) specific actions. Curriculum learning (Bengio et al. 2009) aims to help an agent learn a complex task by learning a sequence of simpler tasks. Hints may also be provided (e.g., a yellow brick road), which fall outside the notion of shaping or a curricula. Despite the prevalence of these approaches, few studies examine how they compare to (or complement) each other or when an approach is better.
展开▼