Exploration vs. Exploitation in Airborne Wind Energy Systems via Information-Directed Sampling Control

Airborne Wind Energy systems (AWEs) are an emerging wind generation technology. They differ from conventional turbines in that they are attached to the ground by a tether and can evolve from low to high altitudes (approx. 1km). Informed altitude control of AWEs is key to track favorable wind speed and maximize power output in a time-varying and partially-observable environment. Leveraging recent advances in Multi-Armed Bandit problems, we recursively estimate the wind profile distribution and use the residuals to fit the noise covariance in an online fashion. This filtering approach paves the way for the computation of (i) the distribution of the wind-output given past observations and (ii) the expected reduction in entropy in the optimum distribution with respect to the potential future altitude set-point. We implement an Information Directed Sampling controller that minimizes the ratio of squared-regret per bit of information gained about the optimum. We finally compare our controller with different baseline controllers using real-world data.