Describe the issue linked to the documentation

There is almost nothing in the doc about it. And you don't find informative ressources about it online easily. So it's probably really obscure for 99% of users.

Suggest a potential alternative/fix

In the params doc: say that it's usually used for additive learning (e.g.when using decision trees as base learner for gradient boosting method).

In the Mathematical Formulation section
Add the friedman MSE criterion. This will be awkard as friedman MSE criterion doesn't minimize the same thing than other criterion. Instead of minizing a $G(Q_m, \theta)$ that can be expressed as:

$$G(Q_m, \theta) = \frac{n_m^{left}}{n_m} H(Q_m^{left}(\theta)) + \frac{n_m^{right}}{n_m} H(Q_m^{right}(\theta))$$

like other criteria, it maximizes:

$$G(Q_m, \theta) = \frac{n_m^{left} n_m^{right}}{n_m} \left( \bar{y}_m^{left} - \bar{y}_m^{right} \right)^2$$

But impurity values of the nodes is the MSE...

In the random forest params doc: maybe add someting here too.