We have applied a technique recently proposed basing on learning nonlinear dynamics locally to describe the annual sunspot relative numbers. It is proved that the number of past points for prediction should be greater than 4 but less than 10. This rather simple approach yields in average relatively good results for short-term forecasts (< 11 yr). Particularly, it predicts that the current cycle no. 22 will reach a very high maximum. However, this approach must be modified in the vicinity of a grand minimum.