Adaptive Neural Network Committees

• Main Author: Lipnickas, Arunas
• Format: Conference Proceeding
• Language: English
• Subjects: Adaptive committees; Adaptive systems; aggregation; Artificial neural networks; Buildings; half&half sampling; Measurement uncertainty; Neural networks; Training; Training data
• DOI: 10.1109/IDAACS.2007.4488407

Combining several classifiers is an effective way for improving overall classification performance. In many cases it is possible to construct several classifiers with different characteristics. Selecting the "best" classifiers with the best individual performance can be shown as suboptimal solution in several cases, and hence here exists a need to find a member selection method to improve classification performance without increasing computational burden. In this paper on the contrary to the ordinary approach of utilising all neural networks available to make the committee decision, we propose to create adaptive committees, which are specific for each input data point. A prediction neural network is used to identify classifiers to be fused for making a committee decision about a given input data. The proposed technique is tested in three aggregation schemes and the effectiveness of the approach is demonstrated on the three real data sets.