Approximating the objective gradient using perceptrons for constrained minimization with application in drag reduction

This thesis is concerned with the constrained minimization problems whose objective function does not have a closed form analytical expression. However, it can be evaluated numerically which enables the determination of an estimator using random samples. We assume that the set of constraints are well-defined and explicitly known and use perceptrons to estimate the objective function. The new method, which we call Neural Reduced Gradient Algorithm, approximates the reduced gradient by combining the gradient value of the estimated objective function with the exact gradients of the constraints. It is tested with some randomly generated and also, some available convex and nonconvex optimization problems. We also provide an interesting real application of the new method in shear stress minimization for drag reduction and compare its performance with the ones obtained using linear estimators, which are proposed as a part of this work.