A Survey of Curriculum Learning in Deep Reinforcement Learning

Curriculum learning (CL) is considered as a key component of the approach in solving many machine learning problems. CL used in deep reinforcement learning (DRL) improves the training speed, generality as well as robustness through organizing the learning trajectory for agent according to the importance of a sample throughout the training process, which provides new ideas for dealing with issues such as sparse rewards and multi-goal tasks. This paper firstly introduces the concept of curriculum learning and its relationship with reinforcement learning, then discusses the paradigms and key points of curriculum learning in deep reinforcement learning. Finally, the paper summarizes several future research hotspots in this field, so as to explore the path for subsequent related research.