In this paper we further develop our research line of chaining several
preprocessing methods with classi ers [
We systematically generate trees representing work ow DAGs using typed genetic programing initialization designed for polymorphic and parametric types. Terminal nodes of a tree correspond to the nodes of the DAG, where each node contains a computational intelligence method. Function nodes of a tree represent higher-order functions combining several DAGs in a serial or parallel manner. We use types to distinguish between input data (D) and predictions (P) so the generated trees represent meaningful work ows. In order to make the method general enough to handle methods like k-means (where k a ects the topology of the DAG) correctly, we had to use the polymorphic type \list of s of size n" ([ ]n) with a natural number parameter n and an element type parameter . The generating method systematically produces work ow DAGs from simple ones to more complex and larger ones, working e ciently with symmetries.
To demonstrate our rst results we have chosen the winequality-white [
To provide a baseline, we tested each of the four classi ers separately on the two selected datasets. The parameters of the classi ers were set using an extensive grid search with 5-fold cross-validation; the classi ers were compared using the quadratic weighted kappa metric. Next, we generated more than 65,000 di erent work ows using the proposed approach, and evaluated all of them. All computational intelligence methods used the default settings, or the tuned settings of the individual methods (denoted as `default' or `tuned' in Fig. 1c).
Tomas Kren et al. The best work ows for the two datasets are presented in Figs. 1a and 1b, and their numerical results are presented in Fig.1c.
We have demonstrated how the valid work ow DAGs can be easily generated
by a typed genetic programming initialization method. The generated work ows
beat the baseline obtained by the hyper-parameter tuning of single classi er
by a grid search, which is not surprising as the single method is also among
the generated DAGs. On the other hand the work ows do not use any
hyperparameter tuning. In our future work, we will extend this approach to a full
genetic programming solution, which will also optimize the hyper-parameters of
the work ows and we intend to include the method in our multi-agent system
for meta-learning { Pikater [
Acknowledgment This work was supported by SVV project no. 260 224, GAUK project no. 187 115, and Czech Science Foundation project no. P103-15-19877S.