Exploring Alternative Biases Prior to Learning in Scientific Domains

Before machine learning can be applied to a new scientific domain, considerable attention must be given to finding appropriate ways to characterize the learning problem. A central idea guiding our work is that we must make explicit more of the elements of a program’s bias and understand the criteria by which we prefer one bias over another. We illustrate this exploration with a problem to which we have applied the RL induction program, the problem of predicting whether or not a chemical is a likely carcinogen.