The algorithms that work well as a metalearning in stacking will be algorithms that do well with highly correlated features. The metalearner is trained on (usually) cross-validated predicted values generated by the base learners (one column for each base learner). This means that the columns in the data frame that the metalearner is trained on are highly correlated.
I'd suggest starting with a GLM metalearner, probably with some L1 or L2 regularization or a non-negativity constraint. I've seen decent performance using deep neural networks as well, but since they are harder to tune, so you have to put more effort in to get good results.