Introduction

Why did the oxygen theory of combustion supersede the phlogiston theory? Why is Darwin's theory of evolution by natural selection superior to creationism? How can a jury in a murder trial decide between conflicting views of what happened? This target article develops a theory of explanatory coherence that applies to the evaluation of competing hypotheses in cases such as these. The theory is implemented in a connectionist computer program with many interesting properties.

The problem of inference to explanatory hypotheses has a long history in philosophy and a much shorter one in psychology and artificial intelligence (AI). Scientists and philosophers have long considered the evaluation of theories on the basis of their explanatory power. In the late nineteenth century, Peirce discussed two forms of inference to explanatory hypotheses: hypothesis, which involved the acceptance of hypotheses, and abduction, which involved merely the initial formation of hypotheses (Peirce 1931–1958; Thagard 1988a). Researchers in artificial intelligence and some philosophers have used the term “abduction” to refer to both the formation and the evaluation of hypotheses. AI work on this kind of inference has concerned such diverse topics as medical diagnosis (Josephson et al. 1987; Pople 1977; Reggia et al. 1983) and natural language interpretation (Charniak and McDermott 1985; Hobbs et al. 1988). In philosophy, the acceptance of explanatory hypotheses is usually called inference to the best explanation (Harman 1973, 1986). In social psychology, attribution theory considers how people in everyday life form hypotheses to explain events (Fiske and Taylor 1984).