[4S1-OS-30a] OS-30

[4S1-OS-30a-01] Emergence of non-trivial behavior for homeostasis through deep homeostatic reinforcement learning

〇Naoto Yoshida Yoshida¹, Yasuo Kuniyoshi^1,2 (1. The University of Tokyo, 2. Next Generation Artificial Intelligence Research Center)

Keywords:Homeostatic Reinforcement Learning, Homeostasis, Open-ended, Deep Reinforcement Learning, Emergence of Behavior

Homeostasis is the most fundamental property of animal survival. Homeostatic reinforcement learning (Homeostatic RL) realizes the emergence of integrated behaviors for survival through internal motivation based on the coupled dynamics inside and outside the body. We have achieved behavior emergence with deep homeostatic RL. However, the behaviors we have observed so far are limited to the emergence of trivial behaviors for homeostasis. In this study, we show experimentally that seemingly non-trivial behaviors can in fact be derived from homeostasis. To this end, we performed deep homeostatic RL in an open environment and constructed an agent that can survive for a long period of time. By evaluating this agent, we observed the emergence of non-trivial behaviors such as attacking zombies and building shelters, as well as trivial behaviors such as collecting food and water. Finally, we discuss the nature of the environmental dynamics necessary for the emergence of these non-trivial behaviors.