Combining unreliable pieces of evidence

The paper discusses the problem of combining unreliable pieces of evidence in a (generalized) probabilistic setting. The popular Dempster's rule for combining evidence in Dempster-Shafer theory is criticized on two accounts. Firstly, Dempster's rule is claimed to be invalid, and secondly, the rule may fail to support intuitively acceptable conclusions. The reason for this is that the application of Dempster's rule implicitly involves making unrealistic assumptions. It is conceded that it in many situations it may be unavoidable to make some assumptions to reach conclusions that are of some practical use, but often intuitively more acceptable assumptions than those used in Demp-ster's rule can be argued for. We illustrate our position by discussing a concrete example of sensor fusion in robotics. Further, an abstract for-malization of belief states is proposed in which assumptions are explicitly mentioned, and the processes conditioning and combining pieces of evidence in the presence of assumptions are discussed. The investigations were carried out as part of the PIONIER-project Reasoning with Uncertainty , subsidized by the Netherlands Organization of Scientiic Research (NWO), under grant pgs-22-262.