Tabula Rasa is a data-driven video installation. Inspired by the philosophical concept of the clean slate, it explores the technology of Deep Reinforcement Learning in machine learning.
An agent’s training is visually interpreted through abstract-organic aesthetics as it is simulated in a fluid environment. As the agent’s single goal is to move forward, all motion data has been generated in its unsupervised learning process.
According to the theory of Tabula rasa, the agent learns by constantly interacting with the environment. Like on a blank sheet of paper, it starts acting without any experience. The only motivation to progress and learn is to be rewarded as high as possible.
To visualize this learning progress, the agent’s appearance corresponds to the training’s state at a time. The rewards and the strategy performance determine the colors and look of the geometry.
The visible progress and the interval until the training’s completion are visualized distinctively. To demonstrate differences in unsupervised learning processes, multiple agents, spawn parallel, not interacting with each other.
Each layer of the accompanying music is metonymy and a metaphor for life events. It overlaps, shakes, disappears, accumulates, shifts between simple and complicated. This process is irreversible: As the agent learns to move forward, time and sound move forward.