A safe control of evolution consists in preventing past errors of evolution to be repeated, which could be done by keeping track of the history of evolution. But maintaining and exploiting the complete history is intractable. This paper therefore investigates the use of machine learning (ML), in order to extract a manageable information from this history. More precisely, induction from examples of past trials and errors provides rules discriminating errors from trials. Such rules allow to a priori estimate the opportunity of next trials; this knowledge can support powerful strategies of control. Several strategies of ML-based control are experimented on the Royal Road, a GA deceptive and a combinatorial optimization problem. The control of mutations unexpectedly compares to that of crossovers.