Light transport as an optimal control problem

Abstract

Recently, reinforcement learning (RL) has attracted some attention in the research field of physically-based rendering. This thesis explores a deeper connection between both fields. In particular, we identify the light transport equations (LTE) used in physically-based rendering with the Bellman expectation equations used in optimal control. With this identification, we formulate an optimal control problem that can be used for scene rendering based on the framework of linearly-solvable Markov decision processes. This formulation provides a novel perspective on the use of RL techniques for rendering.