Active Audition for Robots using Parameter-Less Self-Organising Maps

How can a robot become aware of its surroundings? How does it create its own subjective, inner representation of the real world, so that relationships in the one are reflected in the other? It is well known that structures analogous to Self-Organising Maps (SOM) are involved with this task in animals, and this thesis undertakes to explore if and how a similar approach can be success- fully applied in robotics. In order to study the environment-to-abstraction mapping with a minimum of guidance from directed learning and built-in design assumptions, this thesis examines the active audition task in which a system must determine the direction of a sound source and orient towards it, both in horizontal and vertical direction. Previous explanations of directional hearing in animals, and the implementation of directional hearing algorithms in robots have tended to focus on the two best known directional clues; the intensity and time differences. This thesis hypothesises that it is advantageous to use a synergy of a wider range of metrics, namely the phase and relative intensity difference. A solution to the active audition problem is proposed based on the Parameter- Less Self-Organising Map (PLSOM), a new algorithm also introduced in this thesis. The PLSOM is used to extract patterns from a high-dimensional input space to a low-dimensional output space. In this application the output space is mapped to the correct motor command for turning towards the source and focusing attention on the selected source by filtering unwanted noise. The dimension-reducing capability of the PLSOM enables the use of more than just two directional clues for computation of the direction. This thesis presents the new PLSOM algorithm for SOM training and quantifies its performance relative to the ordinary SOM algorithm. The mathematical correctness of the PLSOM is demonstrated and the properties and some applications of this new algorithm are examined, notably in automatically modelling a robot's surroundings in a functional form: Inverse Kinematics (IK). The IK problem is related in principle to the active audition problem - functional rather than abstract representation of reality - but raises some new questions of how to use this internal representation in planning and execution of movements. The PLSOM is also applied to classification of high-dimensional data and model-free chaotic time series prediction. A variant of Reinforcement Learning based on Q-Learning is devised and tested. This variant solves some problems related to stochastic reward functions. A mathematical proof of correct state-action pairing is devised.