Optimising path tracing with genetic algorithms

Abstract

While being capable of rendering light phenomena such as global illumination and caus- tics, the Path Tracing algorithm has an empirical time complexity that increases linearly with respect to the quality of the frame by an intractable coe cient (due to Monte Carlo point-sampling). By proposing a Genetic Programming (GP) framework using domain- speci c function sets, this work aims to explore whether it is possible to optimise the performance of this physically-based rendering technique over multiple levels of abstrac- tion, and discover alternative implementations of the algorithm that may help guide future research. The results obtained showed that by using the proposed GP framework, it was possible to develop alternative approaches to solving problems that may not be immediately evident to human programmers. Additionally, by observing the limitations of the proposed system, a number of techniques were developed such as domain-partitioning (a form of importance sampling) and the use of variance in calculating error to improve convergence rates for general problem solving. Lastly, this project highlights the shortcomings of the proposed system, presents so- lutions to these problems, and subsequently details related topics that deserve further research, such as meta-heuristics.