An Architecture for Multilevel Learning and Robotic Control based on Concept Generation

Iravani, Pejman (2005). An Architecture for Multilevel Learning and Robotic Control based on Concept Generation. PhD thesis The Open University.



Robot and multi-robot systems are inherently complex systems, for which designing the programs to control their behaviours proves complicated. Moreover, control programs that have been successfully designed for a particular environment and task can become useless if either of these change. It is for this reason that this thesis investigates the use of machine learning within robot and multi-robot systems. It explores an architecture for machine learning, applied to autonomous mobile robots based on dividing the learning task into two individual but interleaved sub-tasks.

The first sub-task consists of finding an appropriate representation on which to base behaviour learning. The thesis explores the viability of using multidimensional classification techniques to generalise the original sensor and motor representations into abstract hierarchies of 'concepts'. To construct concepts the research used standard classification techniques, and experimented with a novel method of multidimensional data classification based on 'Q-analysis'. Results suggest that this may be a powerful new approach to concept learning.

The second sub-task consists of using the previously acquired concepts as the representation for behaviour learning. The thesis explores whether it is possible to learn robotic behaviours represented using concepts. Results show that is possible to learn low-level behaviours such as navigation and higher-level ones such as ball passing in robot football.

The thesis concludes that the proposed architecture is viable for robotic behaviour learning and control, and that incorporating Q-analysis based classification results in a promising new approach to the control of robot and multi-robot systems.

Viewing alternatives

Download history


Public Attention

Altmetrics from Altmetric

Number of Citations

Citations from Dimensions

Item Actions