hero working papers

Multinomial logit processes and preference discovery: inside and outside the black box

Number: 615
Year: 2017
Author(s): S. Cerreia-Vioglio, F. Maccheroni, M. Marinacci, and A. Rustichini

We provide both an axiomatic and a neuropsychological characterization of the dependence of choice probabilities on time in the softmax (or Multinomial Logit Process) form (see below picture) MLP is the most widely used model of preference discovery in all fields of decision making, from Quantal Response Equilibrium to Discrete Choice Analysis, from Psychophysics and Neuroscience to Combinatorial Optimization. Our axiomatic characterization of softmax permits to empirically test its descriptive validity and to better understand its conceptual underpinnings as a theory of agents'rationality. Our neuropsychological foundation provides a computational model that may explain softmax emergence in human behavior and that naturally extends to multialternative choice the classical Diffusion Model paradigm of binary choice. These complementary approaches provide a complete perspective on softmaximization as a model of preference discovery, both in terms of internal (neuropsychological) causes and external (behavioral) effects.

Keywords: Discrete Choice Analysis, Drift Diffusion Model, Luce Model, Metropolis Algorithm, Multinomial Logit Model, Quantal Response Equilibrium