To solve the problem of extreme learning machine(ELM) on-line training, an algorithm, fixed-memory extreme
learning machine(FM-ELM), is proposed. FM-ELM adopts the latest training sample and abandons the oldest training
sample iteratively to enhance its adaptive capacity. The output weights of FM-ELM are determined recursively based on
Sherman-Morrison formula. Thus, the computational cost of FM-ELM training procedure is effectively reduced. Numerical
experiments on nonlinear system on-line condition prediction show that FM-ELM has better performance in adjusting speed
and prediction accuracy in comparison with on-line sequential extreme learning machine(OS-ELM).