Abstract
I generalize acyclic deterministic structural causal models to thenondeterministic case and argue that this offers an improved semantics forcounterfactuals. The standard, deterministic, semantics developed by Halpern(and based on the initial proposal of Galles & Pearl) assumes that for eachassignment of values to parent variables there is a unique assignment to theirchild variable, and it assumes that the actual world (an assignment of valuesto all variables of a model) specifies a unique counterfactual world for eachintervention. Both assumptions are unrealistic, and therefore I drop both ofthem in my proposal. I do so by allowing multi-valued functions in thestructural equations. In addition, I adjust the semantics so that the solutionsto the equations that obtained in the actual world are preserved in anycounterfactual world. I provide a sound and complete axiomatization of theresulting logic and compare it to the standard one by Halpern and to morerecent proposals that are closer to mine. Finally, I extend these models to theprobabilistic case and show that they open up the way to identifyingcounterfactuals even in Causal Bayesian Networks.