Deep Active Inference Agent with Continuous Action Space

Active inference is a promising and unifying framework for the intelligent agent behavior formal description, with ongoing intensive research of the means for sophisticated model implementations which would be able to challenge complex realistic environments. In this paper we introduce a previously used deep learning-based active inference model with a few adaptations into the action-perception architecture and propose our new generative adversarial network (GAN) -based approach to optimally sample from the continuous action space unlike from discrete one as it was done in most of related works. We show that the proposed model is able to learn and solve a simple but realistic environment, optimize its effect on the world in accordance with its own preferences, exhibiting the exploration behavior at the same time. We hope our work will help to make active inference agent design principles more flexible and generalizable.