Demystify Painting with RL

Given an image of an arbitrary scene, experienced artists are skillful at accurately perceiving the visual contents within the scene, such as objects, lighting, and tint, and presenting them in different painting styles. Essentially, this artistic creation procedure starts from a blank canvas and proceeds in a stroke-by-stroke manner, which could be modeled as a sequence of carefully-chosen stroke actions. Given the fact that reinforcement learning (RL) offers a principled approach to tackling sequential decision-making tasks, it is natural to consider applying appropriate RL techniques to training machines to mimic the painting procedure accomplished by human artists. In this project, we aim to learn an agent capable of automatically planning a sequence of strokes that result in a painting with desired visual contents and artistic styles, just as what human painters would do during artistic creation. We perform extensive experiments under different algorithmic designs as an attempt to demystify the learning mechanism and capability of current RL-based painting agents.