An active inference implementation of phototaxis

Active inference is emerging as a possible unifying theory ofperception and action in cognitive and computational neuro-science. On this theory, perception is a process of inferringthe causes of sensory data by minimising the error betweenactual sensations and those predicted by an innergenerative(probabilistic) model. Action on the other hand is drawn as aprocess that modifies the world such that the consequent sen-sory input meets expectations encoded in the same internalmodel. These two processes, inferring properties of the worldand inferring actions needed to meet expectations, close thesensory/motor loop and suggest a deep symmetry betweenaction and perception. In this work we present a simpleagent-based model inspired by this new theory that offers in-sights on some of its central ideas. Previous implementationsof active inference have typically examined a “perception-oriented” view of this theory, assuming that agents are en-dowed with a detailed generative model of their surround-ing environment. In contrast, we present an “action-oriented”solution showing how adaptive behaviour can emerge evenwhen agents operate with a simple model which bears littleresemblance to their environment. We examine how variousparameters of this formulation allow phototaxis and presentan example of a different, “pathological” behaviour