Projected gradient descent denotes a class of iterative methods for solving optimization programs. In convex optimization, its computational complexity is relatively low whenever the projection onto the feasible set is relatively easy to compute. On the other hand, when the problem is nonconvex, e.g., because of nonlinear equality constraints, the projection becomes hard and thus impractical. In this paper, we propose a projected gradient method for nonlinear programs that only requires projections onto the linearization of the nonlinear constraints around the current iterate, similar to sequential quadratic programming (SQP). The proposed method falls neither into the class of projected gradient descent approaches, because the projection is not performed onto the original nonlinear manifold, nor into that of SQP, since second-order information is not used. For nonlinear smooth optimization problems, we assess local and global convergence to a Karush–Kuhn–Tucker point of the original problem. Further, we show that nonlinear model predictive control is a promising application of the proposed method, due to the sparsity of the resulting optimization problem.

In this paper, we analyze first-order methods to find a KKT point of the nonlinear optimization problems arising in Model Predictive Control (MPC). The methods are based on a projected gradient and constraint linearization approach, that is, every iteration is a gradient step, projected onto a linearization of the constraints around the current iterate. We introduce an approach that uses a simple ℓ_p merit function, which has the computational advantage of not requiring any estimate of the dual variables and keeping the penalty parameter bounded. We then prove global convergence of the proposed method to a KKT point of the nonlinear problem. The first-order methods can be readily implemented in practice via the novel tool FalcOpt. The performance is then illustrated on numerical examples and compared with conventional methods.