The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9813

Title:Dynamical Interaction Between Learning, Sensing, Attention, and Behaviour
Date: 1998
Abstract:The report describes the investigation of the dynamical interaction between robot learning, visual attention, and behaviour. Jun Tani has proposed a robot control architecture built onto a vision based robot. His robot was found to be able to create an analogical model of the environment and learn from such a model and its behaviour to adapt itself to the environment. In this project, it was proposed to carry out a similar investigation by modifying Tani's control architecture and building onto a robot with a simpler sensing system. The robot chosen was the Khepera robot which has only infra-red sensors aboard. The architecture consists of multiple neural network modules. The recurrent neural network (RNN) learns the sequence of events encountered by the robot incrementally so that RNN can make a prediction of the future based on such sequences. There are two tasks for the sensing system: the wall following task and the light classification task. Attention between these two tasks is switched by the control of the topdown prediction made by the RNN. The prediction is based on the learning status of theRNN. The experiment is conducted using the Khepera robot simulator in learning both static and dynamic environments. since the nature of the light source classification task was so simple and there existed an uncertainty in the performance of the wall following model, only the interaction between the robot's sensing attention and the robot's bbehaviour was established. This interaction, without the help from the learning process, was found to be enough for the Khepera robot to adapt to the environment.

[Search These Pages] [DAI Home Page] [Comment]