Towards a Biologically Plausible Computational Model of Developmental Learning with Robotic Applications

Makukhin, Kirill (2014). Towards a Biologically Plausible Computational Model of Developmental Learning with Robotic Applications PhD Thesis, School of Information Technology and Electrical Engineering, The University of Queensland.

       
Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
s4216724_phd_35b.pdf Thesis full text application/pdf 13.47MB 3
s4216724_phd_abstract.pdf Thesis abstract application/pdf 32.48KB 1
Author Makukhin, Kirill
Thesis Title Towards a Biologically Plausible Computational Model of Developmental Learning with Robotic Applications
School, Centre or Institute School of Information Technology and Electrical Engineering
Institution The University of Queensland
Publication date 2014
Thesis type PhD Thesis
Supervisor Marcus Gallagher
Scott Bolland
Total pages 127
Total colour pages 18
Total black and white pages 109
Language eng
Subjects 080101 Adaptive Agents and Intelligent Robotics
080199 Artificial Intelligence and Image Processing not elsewhere classified
090602 Control Systems, Robotics and Automation
Abstract/Summary Building robots that are able to efficiently operate in the real world is a formidable challenge. Functioning in the real world is extremely difficult, because it requires a vast amount of knowledge about the world, specific tasks, and the robot’s own body, as well as the ability to handle unexpected situations. Despite the undoubtable power of modern machine learning approaches and great success in the solving of task-specific problems (especially in controlled environments), such algorithms as supervised and reinforcement learning have significant limitations in real world environments, because they require from the designer either a prepared set of labelled examples of desired behaviours or an adequate reward function. Even having these pre-requisites, the algorithm faces the problem of the greater size of the environment comparing to the agent’s learning abilities. In contrast, humans or animals efficiently gather knowledge of the world during the developmental stage. Children and young animals are intrinsically motivated to use exploration and play to select and learn information about the world that is useful in the future, but also fits into their existing mental schemas and satisfies the physical constraints. According to White, intrinsically motivated learning (also referred to as self-motivated exploratory activity) is responsible for the formation of a wide range of salient competences, from grasping and walking to language. Moreover, this ability to explore and learn for the sake of knowledge typically persists over a lifetime, progressing in complexity and allowing the adaption of the organism’s behaviour to the dynamically changing environment. The current research project aims to apply the developmental learning concept to robotics, taking an approach of building a biologically plausible model of intrinsic motivation. Page ii In particular, we have proposed a theory that explains neurological mechanisms underlying an inverted U-shaped dependence of the activity of deep cortical areas with respect to familiarity of perceived information. We have shown that our model is able to account for recent neuroimaging studies that have not been explained yet. Furthermore, we have linked the neural activity to behavioural response through the feeling of pleasure, showing that the neural activity effectively might be considered as a reinforcement signal that shapes the animal behaviour. We have conducted several computational experiments showing that such behaviour could structure exploration and is beneficial for building a world model in close-to-real world conditions. Finally, we have demonstrated in a simulation that exploration on the periphery of knowledge could lead to an emergence of competences.
Keyword developmental learning
robotics
intrinsic motivation
emergence of competences

 
Citation counts: Google Scholar Search Google Scholar
Created: Sat, 20 Sep 2014, 13:42:19 EST by Kirill Makukhin on behalf of Scholarly Communication and Digitisation Service