Conférencier :  Tatsuji Takahashi : "Biconditional probability in reasoning, judgment and decision-making"

Date : Jeudi 21 mars de 16H 15-18H (salle B 106)


Résumé :
The word “if” plays quite crucial roles in reasoning, judgment and decision–
making, especially under uncertainty. Conditionals have recently begun
modeled by the conditional probability, the probability of conditional event q|p,
which was proposed in the theory of subjective probability by de Finetti.
Baratgin, Over, and Politser (in press) empirically confirmed its validity in a
three–valued framework. In the theoretical framework, biconditional (such as
“if p then q, and if q then p”) is expressed by biconditional event q||p. We show
that this biconditional explains quite many experimental results in psychology
of reasoning that have not been given enough interpretations. We also prove
that the biconditional, named pARIs (proportion of assumed--‐‑to--‐‑be instances)
describes the result in causal induction the best (Takahashi, Kohno, & Oyo,
2010; Takahashi & Yokokawa, submited). This suggests a possibility of
integrating deductive and inductive reasoning with the pivot of biconditional
event. We also show that it models people’s biconditionals in truth table tasks
and probability judgment. If time permits, we will also briefly talk about the
adaptive rationality of another biconditional, material equivalence (ME). In the
two--‐‑armed bandit problems, which form the most fundamental class of
iterative decision--‐‑making and machine learning, we see the optimal
performance when the biconditional is utilized as a value function. again
despite its extreme simplicity, ME simultaneously implements three cognitive
properties: Satisficing (Simon, 1956), risk atitude (in Prospect theory by
Kahneman & Tversky, 1984), and comparative valuation of actions (Tversky &
Kahneman, 1974). These properties combined together overcome the
exploitation–exploration dilemma and speed–accuracy tradeoff that are
essential in decision--‐‑making and learning under uncertainty.