DNA for Optimal Control

We introduce Diffusion Network Adaptation (DNA), a framework for finding approximate solutions to continuous time, continuous state, continuous action optimal control problems. We present two reinforcement learning algorithms developed under this framework, one model based and the other model free. We test the algorithms in computer simulations and in a complex pneumatic humanoid robot that had to learn how to kick a ball. The algorithms are mathematically elegant, easy to use, and achieve state of the art performance. The DNA framework provides interesting links to recent reinforcement learning algorithms and helps explain why these algorithms work well in conditions that violate the assumptions under which they were originally developed.