Abstract
We incorporate future information in the form of the estimated value offuture gradients in online convex optimization. This is motivated by demandresponse in power systems, where forecasts about the current round, e.g., theweather or the loads' behavior, can be used to improve on predictions made withonly past observations. Specifically, we introduce an additional predictivestep that follows the standard online convex optimization step when certainconditions on the estimated gradient and descent direction are met. We showthat under these conditions and without any assumptions on the predictabilityof the environment, the predictive update strictly improves on the performanceof the standard update. We give two types of predictive update for variousfamily of loss functions. We provide a regret bound for each of our predictiveonline convex optimization algorithms. Finally, we apply our framework to anexample based on demand response which demonstrates its superior performance toa standard online convex optimization algorithm.