**4. Conclusion**

In this work, the authors deal with an investigation of the possible improvement of the river flow predictions. A new methodology was investigated in which ensemble modeling by datadriven models was applied and in which the harmony search was used to optimize the ensemble's structure. Because various data-driven models with strong prediction capability already exist, the authors were trying to evaluate in the case study presented in this paper (2-day ahead prediction of river flows), whether an ensemble paradigm would also bring some gain in cases when strong algorithms are used as ensemble members. Although the improvement in precision was not relatively as high as in the case when the ensemble consists of weak learners, it was proved that the ensemble model worked better than any of its constituents. These results mean, of course, that the proposed ensemble also works better than the ensembles with weak learners which are usually applied, because these were actually among the members of the proposed ensemble.

The authors' intention was to emphasize one important detail: how the input data for a harmony search optimization of weights should be properly computed. In the authors' investigation, it was verified that using the results of testing folds from cross-validation is the best option. This procedure is described in Section 3.

The authors like to emphasize the following practical aspect about ensemble modeling at the end of this paper. It is well known that for different datasets various algorithms may suit as best choice for prediction and it is never certain in advance, which one of these algorithms will perform with best results. This is known as "no free lunch" theorem. Because of this uncertainty, more algorithms must be usually trained, tested and evaluated during data mining process. These three activities (training, testing and evaluation) together with data preparation are quite laborious and computationally intensive. When this work is already done, instead of choosing only one of these algorithms for obtaining final results, it is wiser to use all already tuned algorithms for ensemble prediction of unknown variable (or subset of these algorithms). Updating prediction using ensemble paradigm almost always brings an improvement in precision as was also confirmed in the case study presented (the results are in **Table 2**). It does not mean a lot of extra work because tuned algorithms for a given task are already available. Gain will be different for different datasets, but as was confirmed also in this study it is surely worth to try this for such a little effort.
