In this talk we discuss some arguments that have been raised against the use of probability theory in artificial intelligence. We identify two main elements in these arguments: The first is related to practical issues on the use of probability theory, while the second is clearly of philosophical nature and concerns the contraposition stochastic/deterministic and the status of the stochastic object. We call these two elements technical and methodological, respectively. The methodological element has not been addressed by the AI community, and we intend precisely to fill this gap. In the methodological element we spot the misleading assumption that since probability theory supposes that data are observations of a stochastic object, a probabilistic model is adequate only for dealing with intrinsically stochastic phenomena. In this talk, on the basis of ideas borrowed from the empiricist and instrumentalist epistemology, we point out that the central hypothesis formulated when adopting a probabilistic model, namely the existence of the stochastic object, is simply a working hypothesis. It follows that a probabilistic model is to be evaluated only on the basis of its predicting power, leaving aside any further metaphysical consideration on the truth of the underlying working hypotheses.
Probabilistic models, Artificial intelligence, Epistemology
C. Piscopo and M. Birattari. (2002)
Probability Theory and Intrinsically Non-Stochastic Phenomena in AI: How to reject a metaphysical issue. Technical Report AIDA-2002-04 of FG. Intellektik, FB. Informatik, TU Darmstadt., Darmstadt, Germany.
J. McCarthy and P. J. Hayes. (1969)
Some philosophical problems from the standpoint of artificial intelligence.
In B. Meltzer and D. Michie (ed.)
Machine Intelligence 4. Edinburgh University Press, Edinburgh, UK. pp. 463-502.
P. Cheeseman. (1985)
In defence of probability.
Proceedings of the Ninth International Joint Conference on Artificial Intelligence. Morgan Kaufmann Publisher, San Mateo, CA, USA. pp. 1002-1009.