The application of adaptive resonance theory and reinforcement learning to mapping and control
In this thesis, the ideas of Adaptive Resonance Theory (ART) and Reinforcement Learning (RL) are applied to the problems of mapping and control. A neural architecture, fuzzy ARTMAP is considered as an alternative to standard feedforward networks for noisy mapping tasks. It is one of a series of architectures based upon ART. Fuzzy ARTMAP has advantages over feedforward networks--such as increased autonomy- and is especially suited to classification-type problems. Here it is used to estimate a continuous mapping from noisy data. Results show that properties useful for classification problems are not necessarily advantageous for noisy mapping problems. One particular feature is found to cause specialisation to the data. A modified variant is proposed which stores probability information in a sub-unit of the architecture. The proposed fuzzy ARTMAP variant is found to outperform fuzzy ARTMAP in a mapping task. Another novel self-organising architecture, loosely based upon a particular implementation of ART, is proposed here as an alternative to the fixed state-space decoder in a seminal implementation of reinforcement learning. A well-known non-linear control problem is considered. Input / output pattern pairs, desired state-space regions and the network size / topology are not known in advance. Results show that, although learning is not smooth, the novel ART-based RL implementation is successful and develops a meaningful control mapping. The new decoder increases its information capacity as necessary and indicates that such a self-organising approach to control is viable. The self-organising properties of the new decoder allow the neurocontroller to retain previously learned information and to adapt to newly encountered states throughout its operation, on-line. A fuzzy version of the original RL implementation is implemented to investigate the possibility of distributing control information across more than one state-space region. The fuzzy version is found to outperform the original RL implementation in a control task.