[4Yin2] Interactive session 2

[4Yin2-22] Model of the interaction between drawing and evaluating aesthetics by a reinforcement learning agent with embodiment

〇Yoshia Abe¹, Shogo Yonekura¹, Yoshiyuki Ohmura¹, Yasuo Kuniyoshi¹ (1.Univ. of Tokyo)

Keywords:reinforcement learning, drawing, aesthetic evaluation, embodiment

This research aims to create an autonomous agent model that acts based on its aesthetic sense and creativity. For this purpose, it is necessary that a sense of aesthetics and creativity is implemented directly as an internal mechanism of the agent, instead of making the agent acquire this sense through supervised learning. We try to address this issue by proposing a reinforcement learning model based on neuroscience and cognitive science knowledge. A drawing agent is placed in a physical environment that consists of a simulated pen and canvas in the model. Under the constraints of physical dynamics and embodiment of the agent, it repeatedly draws strokes on the canvas and makes aesthetic evaluation of the drawing in turns. In the experiment, it was possible to produce drawing behaviors by using an aesthetic evaluation score as a reward. We considered the relationship between the reward setting and drawing behaviors. In future works, we aim to understand human aesthetic sense and creativity in a constructive approach by using the proposed model as a platform for research.