Derivative-free reinforcement learning: a review

Reinforcement learning is about agent models that make the best sequential decisions in unknown environments. In an environment, needs to explore environment while exploiting collected information, which usually forms a sophisticated problem solve. Derivative-free optimization, meanwhile, capable of solving problems. It commonly uses sampling-and-updating framework iteratively improve solution, where exploration and exploitation are also needed be well balanced. Therefore, derivative-free optimization deals with similar core issue as reinforcement learning, has been introduced approaches, under names classifier systems neuroevolution/evolutionary learning. Although such methods have developed for decades, recently, exhibits attracting increasing attention. However, recent survey on this topic still lacking. article, we summarize date, organize aspects including parameter updating, model selection, exploration, parallel/distributed methods. Moreover, discuss some current limitations possible future directions, hoping article could bring more attentions serve catalyst developing novel efficient approaches.