Learning Subjunctive Conditional Information
Learning Subjunctive Conditional Information
16th International Congress on Logic, Methodology and Philosophy of Science and Technology
"If Oswald had not killed Kennedy, no one else would have." What does a rational agent or artificial intelligence learn from such a subjunctive conditional? To date there is no theory that answers this question. Bayesian philosophers, computer scientists, and psychologists of reasoning alike have almost only tackled the learning of indicative conditionals. As compared to Bayesian approaches, Günther (2018) put forth a promising method of how a rational agent learns indicative conditionals. The method is based on Stalnaker's possible worlds semantics of conditionals and Lewis's updating method called `imaging'. Roughly, an agent learns an indicative conditional by imaging on the corresponding Stalnaker conditional. Here, we extend Günther's method to cover the learning of subjunctive conditionals.Why does Günther's method not apply to subjunctive conditionals? On the face of it, there seems to be no problem. From the above cited subjunctive, his method says, you learn that the most similar possible world at which Oswald did not kill Kennedy is a world at which no one else did. However, it is widely agreed upon that the meaning of this subjunctive is different from its corresponding indicative conditional “If Oswald did not kill Kennedy, no one else did“. If you believe this indicative to be false, you can still believe the corresponding subjunctive to be true. Hence, if Günther's method applies to indicative conditionals, it cannot straightforwardly also apply to subjunctive conditionals. The reason is that, as it stands, Günther's method cannot discern between learning a conditional in the subjunctive and indicative mood. This is a problem in light of pairs like the Oswald-Kennedy conditionals.The extended method allows an agent to learn both an indicative conditional and its "contrary" subjunctive. This shows that we have put forth a formal method that can discern between learning a conditional in the indicative and subjunctive mood. This adds yet another dimension on which Günther's method is more general than Bayesian accounts of learning conditionals. It furthermore enables AI systems to learn subjunctive -- as opposed to indicative -- conditional information.