Online ship rolling prediction using an improved OS-ELM

Abstract

In this paper, an improved online sequential extreme learning machine (OS-ELM) is applied on ship roll motion prediction. The OS-ELM is improved by temporal difference (TD) learning which is one of the mostly conventionally used prediction methods in reinforcement learning problem; the model dimension is also optimized by Akaike information criterion (AIC). Online sequential extreme learning machine is an efficient algorithm for on-line construction of single-hidden-layer feedforward networks (SLFNs). Ship's roll motion is hard to be predicted because it is a complex process influenced by various time-varying navigational status and environmental factors. The improved OS-ELM was applied to the simulation of online ship roll motion prediction. Results demonstrate that the proposed method can online give predictions for ship roll motion with extreme fast speed and considerable high accuracy.