Awesome

Deriv - A set of Matlab tools for Automatic Differentiation

Overview

This directory contains a set of tools for performing automatic differentiation in Matlab. Automatic differentiation is performed using Matlab's operator overloading. The intention is to replace double objects with Deriv objects (which contain both values of the variable (acting as a double) as well as the derivative of the variable with respect to the parameter(s) of interest. To perform the derivative calculations in Matlab, Deriv must overload any of the double class operators that Matlab encounters when executing your function. The value of the derivative of any intermediate calculation can be extracted using the Get___deriv function.

A differentiating feature of Deriv is that it goes back to the continuous form of a problem to compute derivatives of non-differentiable objects. For example, we avoid differentiation of the time-step selection process in ode23 by simultaneously computing the solution of the original and differentiated ode in our own algorithm. This includes new algorithms for implementation of Matlab functions such as interp1, QR, etc., which would be approximated by finite differences in other automatic differentiation algorithms.

Repository Contents

• Deriv A Matlab class (this can stand alone and must be in your path) that implements automatic differentiation by operator overloading. All of the overloaded double functions are contained in this file.

• Dzeros A Matlab function that takes care of the preallocation problem. Any preallocated variables in your differentiated function that depend on the independent variable must be replaced with this Dzeros function for now. A simple global change and replace zeros->Dzeros would work, though more efficiency can be introduced by selectively replacing only those functions that are affected by the independent variable.

• Get___gradient A Matlab function that uses the forward mode of automatic differentiation to compute the gradient of a function.

• LICENSE.md The LGPL license.

• README.md This file.

• Set___variable A Matlab function that defines the independent variable for automatic differentiation.

• test___Deriv A series of unit tests for Deriv.

Basic Introduction

To illustrate how this works, we use the function Set_variable to define the variable used as the independent variable. Any number of intermediate calculations (with or without using alpha) can be performed. The derivative of those calculations with respect to alpha can be extracted with the embedded Get_deriv method (hidden inside Deriv).

    >> alpha = Set_variable(7);
    >>
    >> % Perform intermediate calculations to arrive at the desired output.
    >>
    >> output = exp((alpha-5)^3)*sin(alpha);
    >> f = Get_value(output);  % extracts the value of the output
    >> d = Get_deriv(output);  % extracts the derivative of the output wrt alpha
    >>
    >> disp(f)
       1.9584e+03
   
    >> disp(d)
       2.5749e+04

Development Tasks

• overload basic double operators

• overload matrix functions

• overload ode/dae solvers

• overload interp options

• scalar variable, forward mode

• vector variable, forward mode

• vector variable, reverse mode

• Jacobian-vector product, reverse mode

• user documentation (this README.md)

• algorithm documentation (paper, wiki)

Author

Jeff Borggaard, Interdisciplinary Center for Applied Mathematics, Virginia Tech jborggaard@vt.edu

License

These files are provided under the Gnu LGPL License.

Acknowledgments

This software was developed as auxiliary software to support the computation of sensitivity derivatives and derivatives for optimization for a number of research projects including funded projects by the Air Force Office of Scientific Research under contracts FA9550-10-1-0201 and FA9550-12-0-0173, the Department of Energy under contract DE-EE0004261 and National Science Foundation under contracts DMS-1016450 and EAR-0943415. This support is greatly appreciated and facilitates additional software projects such as this one.