Algorithms that learn by means of induction typically employ some sort of bias to make the task of learning computationally feasible. We examine in depth one inductive learning algorithms, the candidate elimination algorithm, and examine the effects of the choice of bias has upon performance. We consider the difficulties that occur when a poor bias is chosen, and present a framework from which we can study biases. We present bias shifting as one method of overcoming these difficulties, and present several bias shifting methods.