Learning and Retrieval of Memory Elements in a Navigation Task

Abstract

Desert ants when foraging for food, navigate by performing path integration and exploiting landmarks. In an earlier paper, we proposed a decentralized neurocontroller that describes this navigation behavior. As by real ants, landmarks are recognized depending on the context, i.e. only when landmarks belong to the path towards the current goal (food source, home). In this earlier version, neither position nor quality of the food sources can be learnt, the memory is preset. In this article, we present a new version, whose memory elements allow for learning food place vectors and quality. When the agent meets a food source, it updates the quality value, if this source is already known, or stores position and quality, if the source is new. Quality values are used to select food sources to be visited. When one source has a too low quality, the agent also finds a shortcut to another known food source.