Barry Haynes

Pruning artificial neural networks using neural complexity measures.

This paper describes a new method for pruning artificial neural networks, using a measure of the neural complexity of the neural network. This measure is used to determine the connections that should be pruned. The measure computes the information-theoretic complexity of a neural network, which is similar to, yet different from previous research on pruning. The method proposed here shows how overly large and complex networks can be reduced in size, whilst retaining learnt behaviour and fitness. The technique proposed here helps to discover a network topology that matches the complexity of the problem it is meant to solve. This novel pruning technique is tested in a robot control domain, simulating a racecar. It is shown, that the proposed pruning method is a significant improvement over the most commonly used pruning method Magnitude Based Pruning. Furthermore, some of the pruned networks prove to be faster learners than the benchmark network that they originate from. This means that this pruning method can also help to unleash hidden potential in a network, because the learning time decreases substantially for a pruned a network, due to the reduction of dimensionality of the network.

The addition of neural networks to the inner feedback path in order to improve on the use of pre-trained feed forward estimators

A learning control architecture which uses a multi-layer feed forward neural network with error back propagation is described. The architecture includes a feed forward estimator which is pre-trained and a feedback controller which continues to learn. The properties of the architectures are investigated through a series of experiments, and the application of the prototype adaptive controllers is described. To examine the performance of this controller, square wave demand signals are applied to the controller and the results are presented.

A self‐healing mobile wireless sensor network using predictive reasoning

Purpose – The paper aims to investigate performance benefits associated with adopting a mobile wireless sensor network (WSN). Sensor nodes are generally energy constrained due to the latter being acquired from onboard battery cells. If one or more sensor nodes fail, possible coverage holes may be created which could invariantly lead to a reduced network lifetime. The paper proposes that instead of rendering the entire WSN inoperative, sensor nodes should physically change position within the region of interest thus adaptively altering the WSN topology with a view of recovering from failures. This type of motion will be referred to as “self healing”.Design/methodology/approach – This paper presents a mobility scheme based on Bayesian networks for predictive reasoning (BayesMob) which is essentially a distributed self healing algorithm for coordinating physical relocation of sensor nodes. Using the algorithm, sensor nodes can predict the performance of the WSN in terms of coverage given that the node moves ...

A Bayesian network approach to a biologically inspired motion strategy for mobile wireless sensor networks

Mobility strategies for wireless sensor networks (WSNs) are presented. We introduce a grazing mobility strategy for mobile WSNs, inspired by the foraging behaviour of herbivores grazing pastures. We present Bayesian network GRAZing (BNGRAZ) that implements the proposed WSN grazing strategy. BNGRAZ uses local neighbourhood information to predict coverage and connectivity performance changes related to sensor node motion characteristics. This enables a sensor node to predict the performance implications related to its direction of movement. We implement the BNGRAZ approach to grazing in a custom built mobile WSN simulator. The WSN performance criteria considered during the validation process include coverage, redundancy, connectivity, and network lifetime.

Application of soft-computing techniques in modelling of buildings

The paper presents recent results on the application of soft computing techniques for predictive modelling in the built sector. More specifically, an air-conditioned zone (Anglesea Building, University of Portsmouth), a naturally ventilated room (Portland Building, University of Portsmouth), and an endothermic building (St Catherine’s Lighthouse, Isle of Wight) are considered. The zones are subjected to occupancy effects and external disturbances which are difficult to predict in a quantitative way and hence the soft computing approach seems to be a better alternative. In fact, the overall complexity of the problem domain makes the modelling of the internal climate in buildings a difficult task which is not always carried out in a satisfactory way by traditional deterministic and stochastic methods. The approach adopted uses fuzzy logic for modelling, as well as neural networks for adaptation and genetic algorithms for optimisation of the fuzzy model. The latter is of the Takagi-Sugeno type and it is built by subtractive clustering as a result of which the initial values of the antecedent non-linear membership functions and the consequent linear algebraic equations parameters are determined. A method of a combinatorial search over all possible fuzzy model structures for a specified plant order is presented. The model parameters are further adjusted by a back-propagation neural network and a real-valued genetic algorithm in order to obtain a better fit to the measured data. Modelling results with actual data from the three buildings are presented where the initial (fuzzy) and the final (fuzzy-neuro and fuzzy-genetic) models are shown.