Partial spectral projected gradient method with active-set strategy for linearly constrained optimization (Q849145)

The authors consider problem of minimizing a continuously differentiable function \(f: \mathbb{R}^n\to\mathbb{R}\) under linear inequality constraints. New algorithm based on a relaxed form of Spectral Projected Gradients (SPG) method is proposed. The algorithm overcomes the main drawback of the SPG methods consisting in expensive computing projections. The new algorithm enlarges the set onto which the projection must be computed, and combines the SPG-like iterations with ``essentially'' unconstrained iterations, which can be easily computed. A careful management of the constraints that must be at each of the iterations preserved gives rise to the partial projected gradient algorithm presented in this paper. Convergence of the algorithm is proved, and its practical implementation is discussed. Report on results of numerical tests on examples from the CUTE library is placed in the concluding part of the paper.