Institut des Systèmes Intelligents
et de Robotique

Title : Doctorant.e
No longer in the unit

Title: Computational modelling of reinforcement learning regulation and variability

Abstract: Reinforcement learning is a family of computational algorithms which draws inspiration from behavioural psychology, neuroscience and machine learning, designed to learn state or action values on the basis of simple positive or negative feedback from the environment. In the Mammal brain, this algorithm is implemented by the dopaminergic system which signals the necessary reward prediction errors. The aim of my work is to study the variability of this learning system both within and between individuals through modelling of experimental data. Within individuals, I am interested at how learning and decision-making parameters can vary through time. First of all, I have established that dopamine sets the current exploration level in a given individual, and secondly, I have proposed a meta-learning model which regulates learning parameters to explain long-term improvements in a three-armed bandit task. As for inter-subject variability, I am continuing the work of Florian Lesaint which deals with sign- vs goal-tracking behaviour: in pavlovian conditioning, sign-trackes are more likely to interact with the predictive stimulus while goal-trackers are attracted towards the location of reward delivery, a phenomenon which has ben explained in terms of the relative balance given to model-free versus model-based learning systems. The proposed model, which already explains a range of experimental findings, led to the formulation of predictions which are currently being tested experimentally, and which I must verify.