Alternating minimization algorithms are typically used to find interference alignment (IA) solutions for multiple-input multiple-output (MIMO) interference channels with more than K=3 users. For these scenarios many IA solutions exit, and the initial point determines which one is obtained upon convergence. In this paper, we propose a new iterative algorithm that aims at finding the IA solution that maximizes the average sum-rate. At each step of the alternating minimization algorithm, either the precoders or the decoders are moved along the direction given by the gradient of the sum-rate. Since IA solutions are defined by a set of subspaces, the gradient optimization is performed on the Grassmann manifold. The step size of the gradient ascent algorithm is annealed to zero over the iterations in such a way that during the last iterations only the interference leakage is being minimized and a perfect alignment solution is finally reached. Simulation examples are provided showing that the proposed algorithm obtains IA solutions with significant higher throughputs than the conventional IA algorithms.