UC Santa Barbara

Reinforcement learning (RL) is an efficient class of sequential decision-making algorithms that have achieved remarkable success in a broad range of applications, such as robotic manipulations, strategic games, or autonomous driving. The most well-known example of reinforcement learning is AlphaGo, a computer program that plays the board game Go and outperforms top human Go players. Unlike other two major machine learning categories, supervised learning and unsupervised learning, in which media artists are actively engaged, reinforcement learning has yet to result in many creative applications.

Generative art is usually driven, in whole or in part, by autonomous systems that are derived from a set of rules. Interestingly, an RL policy can be seen as an autonomous system where the rules are learned by interacting with its environment. Regardless of its initial purpose, reinforcement learning has the potential to expand the boundary of generative art. However, a formal process of applying reinforcement learning to generative art does not yet exist and the current RL tools require an in-depth understanding of RL concepts.

To bridge the gap, the first part of the dissertation introduces a conceptual framework to adapt reinforcement learning for generative art. The framework proposes a term RL-based generative art to denote a novel form of generative art of which the use of RL agents is the key element. The creative process of RL-based generative art and possible emergent behaviors are discussed in the framework. This leads to a discussion of several author's related practices on generative art, deep-learning art, and reinforcement learning. Those practices are critical for understanding the conceptual and technical details of each component in order to construct the framework.

The second part introduces RL5, a JavaScript library for rapidly prototyping RL environments and training RL policies in web browsers. The library combines RL algorithms and RL environments into one framework and is fully compatible with p5.js. RL5 is developed with a particular focus on simplicity to favor (re)usability of RL algorithms and development of RL environments. Specifically, the library implemented three RL algorithms, Tabular Q-learning, REINFORCE, and DDPG, to cover all the three families of model-free RL, and nine RL environments that six of them address autonomous agents in steering behaviors, which can be used as building blocks for complex systems. Finally, the author demonstrates four different use cases of how to apply RL5 for pedagogical and creative applications.