Decision-theoretic Defaults

This paper considers defaults as summaries of decision-theoretic deliberations. We investigate the idea that the default e ! a means that a is the optimal action based on all we know (contingently) being e. It is shown how this notion of a default is nonmonotonic and has a preference for more speciic defaults. It has the advantage of defaults can, in principle, be derived from lower level concepts. We thus have a rational basis for determining whether a default is correct or not. One special case considered is where the action is whether to accept some proposition as true, accept it as false or neither. This is needed to allow for conclusions to be used as premises in other defaults. It is shown that when the gain in utility of accepting a proposition depends only only on the truth of the proposition, then the acceptance of q based on evidence e depends only on whether P(qje) exceeds a threshold that is a function of the utilities for accepting q. We also give a bound on the loss (in utility terms) of using an accepted proposition in another derivation.