This thesis examines a balance between designer effort required in biasing a robot's learning of a task, and the effort required from an experienced agent in influencing the learning using social interactions, and the effect of this balance on learning performance. In order to characterise this balance, a two dimensional design space is identified, where the dimensions represent the effort from the designer, who abstracts the robot's raw sensorimotor data according to the salient parts of the task to increasing degrees, and the effort from the experienced agent, who interacts with the learner robot using increasing degrees of complexities to actively accentuate the salient parts of the task and explicitly communicate about them. While the influence from the designer must be imposed at design time, the influence from the experienced agent can be tailored during the social interactions because this agent is situated in the environment while the robot is learning. The design space is proposed as a general characterisation of robotic systems that learn from social interactions. The usefulness of the design space is shown firstly by organising the related work into the space, secondly by providing empirical investigations of the effect of the various influences on the robot's experiences and how learning performance varies as a function of these influences, and finally by identifying how the conclusions from these investigations apply to the related work for improving learning performance. The empirical investigations implement different learning approaches, and are conducted with simulated and physical mobile robots learning wall-following and phototaxis tasks from an experienced simulated robot or an experienced human, and with a simulated humanoid robot learning an object-interaction task from an experienced simulated robot. The design space is used not only to characterise these investigations and related work, but also to characterise a typical performance surface that can be used to guide the design of new and existing systems. The characterisation shows that a particular level of performance can be maintained by compensating one source of influence for the other, and that performance can generally be improved by increasing the influence from any of these sources. It also shows that the best performance depends on various factors that affect the robot's overall learning potential, such as the available learning resources. The thesis argues that characterising the balance between designer effort and social interactions and how learning performance is affected is crucial for addressing a difficult trade-off: increasing designer effort for biasing the learning of a particular task in a particular environment and thus providing more reliability, versus increasing the influence from the social interactions thus providing more generality.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:726386 |
| Date | January 2003 |
| Creators | Marom, Yuval |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/24903 |
Page generated in 0.0024 seconds