Temporary fix for convex combination method negative log-likelihood loss Relative to existing implementation, we reduce the tolerance at which we declare predictions from a given algorithm the same as another. Note that because of the way SuperLearner is structure, one needs to install the optimization software separately.