Daniel Oblinger, Gerald DeJong
A tradeoff exists between the range of learning tasks solved by an induction system, and its performance on those tasks. We propose dynamic-bias induction, an approach in which bias is dynamically constructed as a function of the learning task. This admits the possibility of a high performance inductive learner that applies to a wide range of learning tasks. We assess the benefits and limitations of dynamic-bias induction by comparing an implementation of the approach to two existing inductive learning systems.