In addition to reasoning about actual states of affairs,humans often reason about what might have been, or counterfactuals. A doctormight ask "if Alice had not been treated with the experimental drug, would shehave survived?'' and a parent might tell a child that "if you had been payingattention, you wouldn't have gotten hurt.'' Bayesian networks have been used toaccount for many aspects of causal reasoning, including inferences aboutcounterfactuals. I will present a Bayes net model of counterfactual reasoningthat generalizes and extends the work of Pearl (2000). The model distinguishesbetween counterfactual observations and counterfactual interventions, and canreason about both backtracking and non-backtracking counterfactuals. Several experiments demonstratethat our model accounts better for human inferences than Pearl's originalproposal and a more recent Bayes net account developed by Rips (2010).

Chris Lucas is a Chancellor's Fellow in the School ofInformatics at the University of Edinburgh. His current research interestsinclude causal inference, preference attribution, function learning, and otherinductive problems that human learners face. He explores these problems usingBayesian models and psychological experiments. Previously, Chris was apostdoctoral researcher at Carnegie Mellon University, after receiving his PhDfrom the University of California, Berkeley.