Differentiable Design With Dynamic Programming

Kelly O. Marshall, Minsu Cho, Chinmay Hegde

Keywords: Deep Learning, Differentiable Programming, Dynamic Programming, AI-Design

TL;DR: Methods for designing input x so that f(x)=c where c is a target value and f is a function computed using dynamic-programming

Abstract: We consider problems in learning-based design subject to constraints specified in the form of Dynamic Programming (DP). Recent work from Mensch and Blondel (2018) proposes the use of a differentiable DP operator, therefore enabling DP constraints to be used in conjunction with gradient-based learning. In this paper, we introduce a differentiable technique called soft-DP that can be used to solve target- matching problems using gradient-based methods. Our technique also enables backpropagating “through” DP solutions that obey a piecewise-linear structure. To validate our approach, we report results from three showcase applications – game design, histogram approximation, and materials design – where our approach improves over data-heavy alternatives.