OpenDiscreteDynamicProgrammingTemplate uses second order optimization and neural networks to find optimal constrained parameters for discrete controls, replacing Hessian with directional derivatives and backpropagation for digital filters.
The OpenDDPT project is a software designed to assist individuals in automatically creating control devices for various engineering problems. Using differential equations and systems of differential equations, the software is able to formulate mathematical models of the real world. This is accomplished through an optimization template that finds optimal control parameters based on the fact that differential equations are parameter dependent.
OpenDDPT employs a innovative non-convex second-order optimization tool-box that replaces the hessian structure, which has heavy memory storage and an expensive computational-cost, with directional-derivatives. This reduces the computational cost of the second order information by one order without approximation. Furthermore, the software utilizes an extended form of the back-propagation rule used for neural-networks to find control's local information such as gradient and its directional derivative.
This new extended form is capable of performing not only weighted-sum and sigmoidal saturations but also multiplication, addition, time-delay, and any non-linearity. Additionally, it can be used as a generic expression for realizing any non-linear discrete time control device. With the ability to describe any real device, such as electronic circuits, automatic controls, and electro-magnetic wave emissions, OpenDDPT offers a powerful tool to assist individuals in creating control devices with ease.
OpenDDPT employs a innovative non-convex second-order optimization tool-box that replaces the hessian structure, which has heavy memory storage and an expensive computational-cost, with directional-derivatives. This reduces the computational cost of the second order information by one order without approximation. Furthermore, the software utilizes an extended form of the back-propagation rule used for neural-networks to find control's local information such as gradient and its directional derivative.
This new extended form is capable of performing not only weighted-sum and sigmoidal saturations but also multiplication, addition, time-delay, and any non-linearity. Additionally, it can be used as a generic expression for realizing any non-linear discrete time control device. With the ability to describe any real device, such as electronic circuits, automatic controls, and electro-magnetic wave emissions, OpenDDPT offers a powerful tool to assist individuals in creating control devices with ease.
What's New
Version 0.3.2: N/A