Gramian-based combination policies for diffusion learning over networks,