Behavioral Optimization for Homeostasis and Its Implications

Homeostasis of internal bodily states is essential for animal survival. Numerous neuroscientific studies suggest that processes such as water regulation, energy balance, and even social interactions can be interpreted as forms of homeostatic control over the body’s physical and chemical states. In computational neuroscience, reinforcement learning (RL) theory is widely recognized as a framework for modeling decision-making. However, the concept of reward in RL typically remains a free parameter. A central challenge in recent research is to establish a general framework that grounds not only the “process” of cognition but also the dynamically changing “motivation” of animals. Addressing this issue, homeostatically regulated reinforcement learning (HRRL) theory has been proposed (Keramati & Gutkin, 2011), which grounds motivation in the homeostatic regulation of internal bodily states. In this talk, I will present our recent findings from computational and robotic experiments demonstrating that implementing HRRL produces a range of intriguing properties in motor control. I will also introduce deep HRRL, the core technology that enables these results.