Dynamic Instance Selection for First Order Adaptable Systems

P. Gczy, S. Usui (Japan), J. Chmrny, D. Repck, M. Lka and J. tulrajter (Slovak Republic)


Dynamic Instance Selection, Training, Learning, First Order Systems


The article presents a novel dynamic instance selec tion algorithm. Efficient training of adaptable systems such as neural networks indispensably depends on the training exemplar set. Appropriate presenta tion of exemplars brings benefits in terms of faster convergence speed, lower computational complexity and better properties of trained neural networks. Conventionally, the neural network is presented the complete set of training exemplars with little or no concern on their immediate suitability at a given iteration of learning. Only recently researchers have started to address the issue of exemplar selectivity. Dynamic instance selection techniques feature the highest flexibility in selection of training instances. The instances are selected dynamically at each iteration of adaptation procedure and presented exem plar set may vary in size and content at each iteration.

Important Links:

Go Back