In gradient boosting, a composition of N base algorithms has already been built. What will be the target for the new, N+1-th algorithm?