Two observations underpin this thesis; 1. There is a need for automated pattem-recognition techniques that allow processes requiring skills normally associated with the human brain to be carried out rapidly, reliably and cheaply, and; 2. The current methods applied to solving artificial intelligence (AI) problems are insufficient to the task of creating generalised systems capable of pattem-recognition and environmental interaction. Neural networks (NNs) are a good method of solving AI problems that are difficult or impossible to solve using knowledge-based or symbolic techniques. NNs provide the flexibility to analyse poorly-defined systems or systems that are general in nature, and also provide the ability to learn from noisy, complex data sets. The main problem with the use of NNs to date has been that one NN's structure and dynamics may work for a specific problem, but if this problem is changed slightly then it is difficult to determine the optimal settings for the network to enable it to adapt to the new situation. The use of evolutionary methods is emphasised throughout this thesis as a way of optimising NN system performance. Several methods have been developed through the course of this thesis that improve the performance of NN models. One of the most important is the use of a biologically plausible node and connection modification algorithm. In this method, local effects such as the activation levels of nodes at either end of a connection or a node's past activation history are the only input parameters which network components use for their adjustment. Included in the biological plausibility argument are NN structuring methods that mimic specific areas of the brain. One example is the visual system, in which a pyramidal structure is applied that permits a hierarchical pattern recognition process to develop. This process builds the image recognition up from small 'substructures' in successive layers, allowing the system to recognise objects that are not specifically defined by the user. Arguments are made that an AI systems's utility is limited if it does not have the capability of interacting with its environment. A system that merely observes without attempting to alter or exist within an environment is only half of the story. From a biological standpoint, intelligence is the result of successive generations of organisms interacting with and altering their environment. Limiting an AI system's ability to interact with the environment can only place restrictions on the capabilities of that system, not improve them. Following development of a suite of applicable pattem-recognition techniques, work is carried out in order to implement these methods within a simple environment. For the moment, a virtual 'block world' is used that is relatively easy and cheap to manipulate. The importance of both modularity and sensory feedback to the ability to develop complex behaviours is investigated, with these two concepts included in the overall evolutionary strategy of system development. The results obtained show that the techniques developed provide a pattem- recognition and learning system that is capable of being applied to general problems and that learns without human intervention. In comparison to classical NN techniques the systems developed show superior learning abilities and can be applied in less specific situations. The use of modularity and sensory feedback in the animat simulations has allowed the development of behavioural patterns that are difficult to achieve using homogeneous, input-output systems. Evolutionary methods have allowed system optimisation in a way that is impossible to achieve through trial and error, and which also permit the system to be easily fine-tuned towards specific problems and situations. With current advances in computer speed and memory capacity, it is now possible to implement NNs comparable in size to the nervous systems of small animals. The methods used here provide the potential to provide these NNs with the sophistication displayed by their organic counterparts.