Decision tree learning algorithms and their application represent one the major successes of AI. Early research on these algorithms aimed to produce classification trees that were accurate. More recently, there has been recognition that in many applications, aiming to maximize accuracy alone is not adequate since the cost of misclassification may not be symmetric and that obtaining the data for classification may have an associated cost. This has led to significant research on the development of cost-sensitive decision tree induction algorithms. One of the seminal studies in this field has been the use of genetic algorithms to develop an algorithm known as ICET. Empirical trials have shown that ICET produces some of the best results for cost-sensitive decision tree induction. A key feature of ICET is that it uses a pool that consists of genes that represent biases and parameters. These biases and parameters are then passed to a decision tree learner known as EG2 to generate the trees. That is, it does not use a direct encoding of trees. This paper develops a new algorithm called ECCO (Evolutionary Classifier with Cost Optimization) that is based on the hypothesis that a direct representation of trees in a genetic pool leads to improvements over ICET. The paper includes an empirical evaluation of this hypothesis on four data sets and the results show that, in general, ECCO is more cost-sensitive and effective than ICET when test costs and misclassifications costs are considered.