Motivational engine and long-term memory coupling within a cognitive architecture for lifelong open-ended learning

This paper considers a cognitive architecture that revolves around a network memory based Long-Term Memory and how it can lead to a working lifelong learning system that can deal with open-ended learning. It focuses on the mutual interaction between the Motivational Engine and the Long-Term Memory and, in particular, on autonomously producing high-level utility representations in order to allow for development. Thus, the main point is to study how this architecture allows to start from primitive policies and models operating over continuous and large state/action spaces and progressively move towards higher level structures defined over smaller and discrete state/action spaces. This progression is demonstrated in a series of experiments carried out on a real robotic setup that involves different contexts, both in terms of domains (worlds) and tasks (goals).