Gradpy and Therapy Planner Documentation

• Overview
• Installation
  • Install gradpy Through PyPI
  • Testing gradpy
• Usage
  • Introductory demo of autodiff module with the gradpy package
  • Special functions
  • Newton’s method with automatic differentiation
• Background
• Software Organization
  • Directory Structure
  • Modules
  • Examples
  • Tests
• Implementation
  • Core Classes
  • External Dependencies
• Featured Application: Intensity Modulated Radiation Therapy (IMRT)
  • Overview
  • Basics of IMRT
  • Optimization with gradpy
  • Sample radiation therapy optimization
  • Install therapy_planner
• Summary and Future

Introduction

The gradpy package, through its primary module autodiff, implements automatic differentiation, a technique for computing derivatives of functions that is distinct from both symbolic and numerical differentiation.

The primary advantage of automatic differentiation is its ability to handle complicated functions without sacrificing the accuracy of the computed derivative. Whereas symbolic differentiation guarantees accuracy, it can be intractable for complex functions. By contrast, numerical differentiation affords ease in implementation and is amenable to any function, but it comes at the expense of accuracy. Automatic, or algorithmic, differentation is an alternative approach which allows derivatives to be computed by a computer program up to machine precision.

Automatic differentiation in science and engineering has been applied to address problems in a variety of areas, including optimization, root-finding, and implicit time-integration.

The therapy_planner package is an application of automatic differentiation. See more about therapy_planner in Featured Application: Intensity Modulated Radiation Therapy (IMRT) page.

This project also hosts the therapy_planner package, an application of automatic differentiation for dose delivery optimization relevant to Intensity Modulated Radiation Therapy (IMRT), that automatically ships with gradpy.

This project is hosted on GitHub.