A Constructive Neural Network Incorporating Competitive Learning of Locally Tuned Hidden Neurons

Performance metrics are a driving force in many fields of work today. The field of constructive neural networks is no different. In this field, the popular measurement metrics (resultant network size, test set accuracy) are difficult to maximise, given their dependence on several varied factors, of which the mostimportant is the dataset to be applied. This project set out with the intention to minimise the number of hidden units installed into a resource allocating network (RAN) (Platt 1991), whilst increasing the accuracy by means of application of competitive learning techniques. Three datasets were used for evaluation of the hypothesis, one being a time-series set, and the other two being more general regression sets. Many trials were conducted during the period of this work, in order to be able to prove conclusively the discovered results. Each trial was different in only one respect from another in an effort to maximise the comparability of the results found. Four metrics were recorded for each trial- network size (per training epoch, and final), test and training set accuracy (again, per training epoch and final), and overall trial runtime. The results indicate that the application of competitive learning algorithms to the RAN results in a considerable reduction in network size (and therefore the associated reduction in processing time) across the vast majority of the trials run. Inspection of the accuracy related metrics indicated that using this method offered no real difference to that of the originalimplementation of the RAN. As such, the positive network-size results found are only half of the bigger picture, meaning there is scope for future work to be done to increase the test set accuracy.