Optimizing Control by Robustly Feasible Model Predictive Control and Application to Drinking Water Distribution Systems

The well-known Nonnegative Matrix Factorization (NMF) method can be provided with more flexibility by generalizing the non-normalized Kullback-Leibler divergence to αdivergences. However, the resulting α-NMF method can only achieve mediocre sparsity for the factorizing matrices. We have earlier proposed a variant of NMF, called Projective NMF (PNMF) that has been shown to have superior sparsity over standard NMF. Here we propose to incorporate both merits of α-NMF and PNMF. Our α-PNMF method can produce a much sparser factorizing matrix, which is desired in many scenarios. Theoretically, we provide a rigorous convergence proof that the iterative updates of α-PNMF monotonically decrease the α-divergence between the input matrix and its approximate. Empirically, the advantages of α-PNMF are verified in two application scenarios: (1) it is able to learn highly sparse and localized part-based representations of facial images; (2) it outperforms α-NMF and PNMF for clustering in terms of higher purity and smaller entropy. Supported by the Academy of Finland in the project Finnish Center of Excellence in Adaptive Informatics Research.