Dissertation, Otto-Friedrich-Universität Bamberg, 2019

This thesis deals with the automatic and semantically autonomous construction of the mental representations of an agent — so-called "symbol grounding".

How can a system perform an independent semantic interpretation of its sensorimotor data, that is not just an imitation of the semantics in the head of its designer? The ability to do so is a prerequisite for general learning in unknown environments. Previous approaches try to achieve this in three different ways: by simulating a sufficiently complex biological brain (anatomically motivated), by simulating and combining functional modules of the human psyche (psychologically motivated), and by identifying a fundamental algorithm that enables different types of learning in the same way (holistically motivated).

This work follows the third approach and draws its inspiration from modern phenomenology, theories of embodied cognition, semiotics, and methods of machine learning. Previous approaches to the dynamic generation of representations are presented. After that a new approach in the field of reinforcement learning is worked out. Physically present aspects of the environment are captured as sensorimotor activations within a system so that their occurrence can be predicted probabilistically. This is implemented within the theoretical framework of conditional probabilities according to Bayes with an extension for the identification of hierarchical structures in the environment.

It can be shown, on the one hand, that a hierarchical approach exceeds previous methods for sequence prediction. On the other hand, it allows a differentiation of different subsequences and it allows their representation and the modification of these representations at runtime. The possibilities and limitations of the developed algorithm are illustrated and evaluated on the basis of various experiments.