Path Tracing with Q-Learning
Bachelor Thesis, National Tsing Hua University, 2019

overview

Implementation

framework

Results

Regular Grid

regular grid
Visualization of regular grid in each scene.

Light Sampling

light sampling
Comparison under 8-spp.

BRDF Sampling

brdf sampling
Comparison under 8-spp and ray depth = 16.

Final

final
Final comparison of each scene. Red regions in Q(s, a) indicate high reward.

Limitation

  • Insufficient resolution of regular grid.
    • insufficient resolution
  • Grid artifact.
    • hyperparameter tuning

Future Works

  • Atomic operation hurts performance.
  • Other state representation such as surfel.
  • Maybe some ideas of deep reinforcement learning.

Citation

Acknowledgements

The website template was borrowed from Michaël Gharbi , Jonathan T. Barron and Thomas Müller .