ICRS-Filter: A randomized direct search algorithm for constrained nonconvex optimization problems

• A novel algorithm and its implementation for the stochastic optimization of generally constrained Nonlinear Programming Problems.
• Use of filter methods in random sampling stochastic optimization method.
• Penalty-free algorithm for global optimization.
• Ability to handle very large scale applications.
• Global optimization success rate 92% of the 104 case studies presented in this work.

This work presents a novel algorithm and its implementation for the stochastic optimization of generally constrained Nonlinear Programming Problems (NLP). The basic algorithm adopted is the Iterated Control Random Search (ICRS) method of Casares and Banga (1987) with modifications such that random points are generated strictly within a bounding box defined by bounds on all variables. The ICRS algorithm serves as an initial point determination method for launching gradient-based methods that converge to the nearest local minimum. The issue of constraint handling is addressed in our work via the use of a filter based methodology, thus obviating the need for use of the penalty functions as in the basic ICRS method presented in Banga and Seider (1996), which handles only bound constrained problems. The proposed algorithm, termed ICRS-Filter, is shown to be very robust and reliable in producing very good or global solutions for most of the several case studies examined in this contribution.