Enhanced sampling methods have the capability to facilitate crossing large (free) energy barriers. Despite many efforts that have been made in this area, two major difficulties still remain: How to sample high-dimensional free energy landscapes? And how to select the collective variables (CVs) if there is limited prior knowledge of a system? Here we introduce adaptive reinforced dynamics (ARiD), a method to efficiently explore the high-dimensional free energy landscape based on the free energy profile and error indicator estimated by deep neural network models. ARiD is flexible in determining the number of CVs and highly parallel, and it requires little prior knowledge and human intervention. We illustrate this method by studying various representative and challenging examples.