Physical computation and embodied artificial intelligence

Physical computation and embodied artificial intelligence In this thesis, we show that the Turing machine is insufficient to model the richness of physical behaviour. Instead, we propose a model of physical computation that allows us to understand how embodied intelligent agents can simultaneously be considered to be objects operating under the laws of nature (or physics) and information-processing devices. This analysis allows us to make concrete those issues relating to the relative merits of using analogue values or symbolic representations such as numbers. Indeed, our approach allows this longstanding point of contention in artificial intelligence to be transformed from a question of philosophy to one of physics. For well-defined tasks, the results of this analysis are equivalent to Shannon’s information theory. This is as expected. However, in certain circumstances, both the computational power and energy efficiency of an analogue system may be greater than would be possible using a classical Turing machine. This supports results from both neuromorophic engineering and theoretical computer science. Finally, for large, multi-functional systems with ill-defined roles, the new model provides a novel way of thinking about the complementarity of hardware and software. Physical computation and embodied artificial intelligence v I dedicate this thesis to two men who inspired me to make a contribution to science and technology: my father, Hardial Singh Bains, who died as I began my Ph.D.; and my mentor, Professor Stephen A. Benton, who died as I was finishing it. Thanks for challenging me to do and be more. Physical computation and embodied artificial intelligence vii