A recurrent neural network for solving a class of generalized convex optimization problems.

Area of Science:

• Computational Mathematics
• Artificial Intelligence
• Optimization Theory

Background:

• Constrained optimization problems are fundamental in various scientific and engineering disciplines.
• Solving nonsmooth optimization problems with generalized convex objectives presents significant challenges.
• Existing methods may struggle with convergence guarantees or applicability to diverse constraint types.

Purpose of the Study:

• To propose a novel penalty-based recurrent neural network (RNN) for solving constrained optimization problems.
• To demonstrate the model's applicability to nonsmooth optimization with affine equality and convex inequality constraints.
• To provide theoretical guarantees for the network's convergence properties.

Main Methods:

• Development of a penalty-based recurrent neural network model.
• Utilizing differential inclusion to describe the network's simple structure.
• Analysis of the network's behavior for regular and pseudoconvex objective functions.

Main Results:

• The proposed recurrent neural network globally converges to the feasible region in finite time.
• The network's state vector remains within the feasible region after convergence.
• The model is proven to converge to the optimal solution set of the constrained optimization problem.

Conclusions:

• The penalty-based recurrent neural network offers an effective approach for solving a class of constrained optimization problems.
• The model's theoretical convergence guarantees make it a robust tool for nonsmooth and generalized convex optimization.
• This work contributes a new computational framework for tackling challenging optimization tasks.