Ensemble Stacking Case-Based Reasoning for Regression

This paper presents a case-based reasoning algorithm with a two-stage iterative double stacking to find approximate solutions to one and multidimensional regression problems. This approach does not require training, so it can work with dynamic data at run time. The solutions are generated using stochastic algorithms in order to allow exploration of the solution space. The evaluation is performed by transforming the regression problem into an optimization problem with an associated objective function. The algorithm has been tested in comparison with nine classical regression algorithms on ten different regression databases extracted from the UCI site. The results show that the proposed algorithm generates solutions in most cases quite close to the real solutions. According to the RMSE, the proposed algorithm globally among the four best algorithms, according to MAE, to the fourth best algorithms of the ten evaluated, suggesting that the results are reasonably good.