Each concept description language and search strategy has an inherent inductive bias, a preference for some hypotheses over others. No single inductive bias performs optimally on all problems. This paper describes a system that couples induction with optimization to carry out an efficient search of large regions of inductive bias space. Experimental results are reported demonstrating the system's capacity to choose optimal biases even for complex and noisy problems.
David K. Tcheng, Bruce L. Lambert, Stephen C. Y. L