# How to choose value of additive smoothing in naive Bayes and why a higher value gives bad accuracy?

In Naive Bayes we often do additive smoothing as a fail safe. Consider the following expression: Lets say $$P(X_i) = \frac{count(X_i) + \alpha}{\sum_i^n count(X_i) + \alpha*total\_size}$$

How to tune the parameter $\alpha$ here. I observed that my classifier accuracy increased for value from $1$ to $N$ (where N is an Integer) and then started to decrease for higher values of $\alpha$. So why does this happen and how to choose this parameter.