Upon receiving the next example, flip a $(6-$sided$)$ fair dice. If the outcome is $1$ or $2$ return the outcome by the majority of functions in the version space $($as in the regular Halving algorithm$),$ and if the outcome is $3,4,5,$ or $6,$ return $`-\text{'}.$

After the true label is revealed, update the version space as before.

What is the tightest upper bound for the expected number of mistakes by this algorithm for adversarial inputs?