Deep Free Energy Principle for Adaptive Behavior
Ryoya Ogishima, Shogo Yonekura, Yasuo Kuniyoshi · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2018
Free Energy Principle enables agents to understand the generative models of the environment, to have beliefs about their current states by perceptual inference, and to behave adaptively to environmental changes by minimizing their prediction errors. This work combines Free Energy Principle from computational cognitive neuroscience and Deep Learning from computer science, suggesting its potentials to be applied to the understanding of agents' adaptive behaviors in complex environments. As an example, this paper shows that an agent can behave adaptively when it is given an expert's goal-directed belief.