Information transfer and causality in the sensorimotor loop

This thesis investigates information-theoretic tools for detecting and describing causal influences in embodied agents. It presents an analysis of philosophical and statistical approaches to causation, and in particular focuses on causal Bayes nets and transfer entropy. It argues for a novel perspective that explicitly incorporates the epistemological role of information as a tool for inference. This approach clarifies and resolves some of the known problems associated with such methods. Here it is argued, through a series of experiments, mathematical results and some philosophical accounts, that universally applicable measures of causal influence strength are unlikely to exist. Instead, the focus should be on the role that information-theoretic tools can play in inferential tests for causal relationships in embodied agents particularly, and dynamical systems in general. This thesis details how these two approaches differ. Following directly from these arguments, the thesis proposes a concept of “hidden” information transfer to describe situations where causal influences passing through a chain of variables may be more easily detected at the end-points than at intermediate nodes. This is described using theoretical examples, and also appears in the information dynamics of computer-simulated and real robots developed herein. Practical examples include some minimal models of agent-environment systems, but also a novel complete system for generating locomotion gait patterns using a biologically-inspired decentralized architecture on a walking robotic hexapod.