Neural networks for nonlinear and mixed complementarity problems and their applications.

Area of Science:

• Computational Mathematics
• Artificial Intelligence
• Optimization Theory

Background:

• Nonlinear and mixed complementarity problems (MCPs) are fundamental in various scientific and engineering fields.
• Existing methods for solving MCPs can be computationally intensive or lack guaranteed convergence.
• Feedback neural networks offer a promising alternative for solving complex optimization and equilibrium problems.

Purpose of the Study:

• To propose two novel feedback neural network models for solving nonlinear and mixed complementarity problems.
• To demonstrate the efficiency and stability of the proposed networks.
• To extend the applicability of these networks to related problems like nonlinear convex programming and variational inequalities.

Main Methods:

• Design of a parameter-free feedback neural network for strictly monotone MCPs, emphasizing hardware implementability.
• Construction of a second feedback neural network utilizing the first as a subnetwork for general monotone MCPs, minimizing state variables.
• Mathematical proofs for the stability and convergence properties of the proposed neural networks.

Main Results:

• The first network ensures uniform and asymptotic stability for strictly monotone MCPs.
• The second network guarantees convergence to an exact solution from any initial point for MCPs with multiple solutions.
• Simulation experiments validate the feasibility and efficiency of both proposed neural networks.

Conclusions:

• The developed feedback neural networks provide robust and efficient solutions for nonlinear and mixed complementarity problems.
• The networks exhibit strong stability and convergence properties, even for problems with multiple solutions.
• The proposed approach demonstrates versatility, applicable to nonlinear convex programming and monotone variational inequalities.