Analog neural networks for real-time constrained optimization

Architectures and circuit techniques for implementing general piecewise constrained optimization problems using VLSI techniques are explored. Discrete-time analog techniques are considered due to their inherent accuracy, programmability, and reconfigurability. A general architecture for minimizing piecewise functions by using gradient schemes is introduced. Switched-capacitor (SC) building blocks featuring improved characteristics in terms of area occupation and operation speed are presented. The implementation of the architectures by using the newest switched-current techniques is discussed. The layout of a 3-μm CMOS SC prototype for a quadratic optimization problem with linear constraints is given.