The Use of Relevance to Evaluate Learning Biases

This paper describes Probabilistic Bias Evaluation (PBE), a method for evaluating learning biases formally analyzing the tradeoff between the expected accuracy and complexity of alternative biases. Intelligent agents must filter out irrelevant aspects of the environment, in order to minimize the costs of learning. In PBE, probabilistic background knowledge about relevance is used to compute expected accuracy; and the complexity of a bias is used to estimate the costs of learning. These are combined into a single value that can be used to select the best bias: one which maximizes predictive accuracy while minimizing computation cost. 1