|
| EPSRC Reference: |
GR/R69679/01 |
| Title: |
Developmental Learning Algorithms for Embedded Agents |
| Principal Investigator: |
Professor MH Lee |
| Other Investigators: |
|
| Researcher Co-investigator: |
|
| Project Partner: |
|
| Department: |
Computer Science |
| Organisation: |
Aberystwyth University |
| Scheme: |
Standard Research |
| Starts: |
01 October 2002 |
Ends: |
31 March 2006 |
Value (£): |
208,307
|
| EPSRC Research Topic Classifications: |
| Learning Engineering Systems |
New and Emerging Computer Paradigms |
|
| EPSRC Industrial Sector Classifications: |
| Healthcare |
Information Technologies |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary |
The methods are inspired by concepts in developmental psychology, have some novel characteristics, and promise to be of considerable value for future generation domestic embedded agents in spatial domains. We aim to deliver results and demonstrations of a simplified working handleye system that learns to coordinate its sensory and motor systems from initial primitive exploratory acts to skilled reaching and coordinated action towards objects and events in the environment.
Novel aspects include: a new design of home robot, with several new features; the replacement of textual programming with learning by example; adaptation to user changes, permitting changing preferences, learning new routes and task correction and reconfiguration.
|
| Final Report Summary |
The aim of this project was to investigate the concept of development'' as a mechanism for learning in robots and autonomous systems. Psychologists have long observed and theorised about the origins and growth of adaptive behaviour in infants. The first year of life is one of the most remarkable periods of cognitive growth and emulating some of this growth and behaviour might provide powerful learning algorithms for intelligent systems.
An experimental hardware and software environment has been created and used for extensive exploration and experimentation on developmental algorithms. The experiments have produced hand/eye behaviour that is very similar to reported infant behaviour and displays a continuous growth of sensory-motor learning through a series of distinct stages of competence. This has produced insights into spatial learning, motor activity, and multi-modal coordination.
A novel mapping system has been designed that provides a computational substrate for understanding the key factors in developmental behaviour. Our approach is important as its open architecture and low representational overhead exposes the relations and mechanisms involved in coordination, controlling attention and action selection, etc
The benefits of this work lie in expanding our options in designing learning methods for robotics and other embedded systems. This approach adds to our understanding of psychology inspired (top-down) models rather than the much more popular neuroscience (bottom-up) models. This project has directed attention to developmental robotics and in particular to our approach towards algorithmic generators of behaviour. This has been an adventurous, exploratory project and it has exposed many new interesting aspects of the exciting interaction between psychology and computing.
|
| Further Information: |
|
| Organisation Website: |
|
|
|