Probability from similarity∗

The probability calculus provides an attractive canonical form for reasoning but its use requires numerous estimates of chance. Some of the estimates needed in artificial systems can be recorded individually or via Bayesian networks. Others can be tabulated as relative frequencies from stored data. For the shifting contexts of commonsense reasoning, however, the latter sources are likely to prove insufficient. To help fill the gap, we show how sensible conditional probabilities can be derived from absolute probabilities plus information about the similarity of objects and categories. Experimental evidence from studies of human reasoning documents the naturalness of the numbers we derive.