Although it sounds strange but in reinforcement learning there are approaches where actors learn on real world sensory input and then do a couple of rounds of i

Although it sounds strange but in reinforcement learning there are approaches where actors learn on real world sensory input and then do a couple of rounds of internal simulations to refine the parameters. Then, this is tested against real worlds. I remember that this outperformed algorithms without such simulations. Also dreaming found in humans and other animals seems like this.