Effective dynamic sample selection algorithm

Data/information overload is becoming increasingly important issue. Training adaptable systems, or machine learning agents, on large data sets is computationally costly and often ineffective. Proper management of data utilized for adaptation can speed up learning and decrease computational costs. The article presents a sample selection algorithm easily implementable into first order adaptable systems. It effectively selects an appropriate set of training exemplars at each iteration of adaptation. The selected exemplar set may vary in size and chosen data. Dynamic sample selection algorithm is computationally inexpensive and positively contributes to the increased convergence speed of the first order learning methods. The presented dynamic sample selection is theoretically justified and practically demonstrated on the tasks of neural networks training. The simulation results indicate satisfactory performance.