Professional and personal website.
Reinforcement Learning (RL) agents should be able to efficiently generalize to novel situations and transfer their learned skills. Without these properties, such agents would always have to learn from scratch, even though they have already mastered primitive skills that could potentially be leveraged to acquire more complex ones.
Autotelic agents—agents that are intrinsically motivated to represent, generate and pursue their own goals—aim at growing their repertoire of skills. This implies that they need not only to discover as many goals as possible, but also to learn to achieve each of these goals. When these agents evolve in environments where they have no clue about which goals they can physically reach in the first place, it becomes challenging to handle the exploration-exploitation dilemma.
Toddlers are the best learners we know to exist. From a very young age, they exhibit impressive behaviors when interacting with their surroundings. Interestingly, these capacities are not only dependent on caregivers, but can also come from autonomous interactions with objects. So how can an infant learn concepts on her own based only on sensorimotor interactions with the physical world ? On this matter, research has been mainly divided in two conceptually and factually different perspectives: the empiricist view, suggesting that knowledge comes primarily from sensory experience, and the nativist view, stating that certain skills and abilities are hard-wired into the brain at birth.