A comparison of genetic programming representations for binary data classification

The choice of which representation to use when applying genetic programming (GP) to a problem is vital. Certain representations perform better than others and thus they should be selected wisely. This paper compares the three most commonly used GP representations for binary data classification problems, namely arithmetic trees, logical trees, and decision trees. Several different function sets were tested to determine which functions are more useful. The different representations were tested on eight data sets with different characteristics and the findings show that all three representations perform similarly in terms of classification accuracy. Decision trees obtained the highest training accuracy and logical trees obtained the highest test accuracy. In the context of GP and binary data classification the findings of this study show that any of the three representations can be used and a similar performance will be achieved. For certain data sets the arithmetic trees performed the best whereas the logical trees did not, and for the remaining data sets the logical tree performed best whereas the arithmetic tree did not.