Apprentissage causal humain : faire le lien entre l'apprentissage par renforcement et l'inférence causale pour modéliser et tester les processus de prise de décision

Causal learning is a fundamental aspect of decision-making. Deciding to perform an action to get an outcome requires the belief that this action can causally produce this outcome. However, most computational models of human decision-making neglect causality. These models most of the time rely on the reinforcement learning theory or on Bayesian inference. Both theories are interested in how humans estimate the probability of obtaining an outcome, but they assume causal links between actions and outcomes without describing how humans appraise it. Recently, a new field called causal reinforcement learning has emerged in the statistics and machine learning communities. It proposes formal tools to disentangle causality from correlation. This thesis attempts to strengthen the links between these different paradigms. On one hand, it proposes computational models of reinforcement learning and Bayesian decision-making with a causal approach. On the other hand, it compares different models to demonstrate the importance of considering the causal aspect in the formulation of research on human decision-making. Finally, a new series of experimental tasks based on causal bandit models is developed to better understand the use of causal information in human decision-making. More precisely, after an analysis of the state of the art on causality problems in decision-making, different algorithms with and without causal information are compared in a series of simulations with increasing difficulty. From these simulations, a new human behavioural protocol is designed to test model predictions, and developed in a pilot study. A reinforcement learning model associated with an intrinsic reward mechanism directly linked to the level of causal curiosity is proposed. Finally, an improvement of a Bayesian model of human decision-making is proposed using knowledge from machine learning.