A new gradient-based neural network for solving linear and quadratic programming problems.

Area of Science:

• Computational Mathematics
• Artificial Intelligence
• Operations Research

Background:

• Linear and quadratic programming (LP/QP) are fundamental optimization problems.
• Existing neural network approaches often struggle with efficiency and handling inequality constraints.

Purpose of the Study:

• To develop a novel gradient-based neural network for solving LP and QP problems.
• To enhance network efficiency and robustness by introducing a new function into the energy function.

Main Methods:

• Construction of a gradient-based neural network integrating duality theory, optimization theory, convex analysis, Lyapunov stability, and LaSalle invariance principle.
• Introduction of a new function F(x, y) into the energy function E(x, y) to ensure convexity and differentiability.
• Handling of inequality constraints without resorting to penalty or Lagrange methods.

Main Results:

• The proposed network incorporates all necessary and sufficient optimality conditions for convex QP problems.
• Strict proofs demonstrate convergence of network trajectories to optimal solutions for both primal and dual problems, regardless of the initial point.
• Simulation results validate the feasibility and efficiency of the proposed neural network.

Conclusions:

• The novel neural network provides an effective and efficient method for solving linear and quadratic programming problems.
• The approach offers advantages over existing methods, particularly in handling inequality constraints.
• The theoretical guarantees and simulation results support the practical applicability of this new network architecture.