Combining policies: the best of human expertise and neurocontrol

Abstract : We consider sequential decision making in the case where a generative model and a parametric policy are available. Such a framework is naturally tackled with Direct Policy Search, i.e. parametric op-timisation over simulations. We propose a simple method that combines this parametric policy with a more generic neural network, where all parameters are trained simultaneously. As such, our approach doesn't require any computational overhead. We show that the resulting policy significantly outperforms both the domain specific policies and the neural network on a unit commitment test problem.