The incremental systems directed by hypotheses learn concepts by selecting the candidate hypothesis in advance in the search space and then testing it with examples. The selection of the "interesting" hypotheses is not based on the incoming example as in example-driven systems but on heuristics. In this paper we will focus on the monotonic algorithms that use background knowledge to define a generality relation between hypotheses. In a common framework we study and compare the heuristics or biases used in the hypothesis selection phase and show how they are closely related to the learning goals. Finally we will show that the problem of the evolution of the learning biases may be solved by the appropriate shift of the hypothesis selection bias.