Convergence

A good way to monitor the algorithm to detect convergence is to maintain a history list of the last few values of $Q$. If there is no improvement in $Q$ for the duration of the history list, terminate the training. Note that because of pruning, etc, it is possible for $Q$ to go down.