Enseigner des agents autotéliques basés sur des prédicats
Ahmed Akakzia · theses.fr (ABES) · 2022
As part of the quest for designing embodied machines that autonomously explore their environments, discover new behaviors and acquire open-ended repertoire of skills, artificial intelligence has been taking long looks at the inspiring fields of developmental psychology and cognitive sciences which investigate the remarkable continuous and unbounded learning of humans. This gave birth to the field of developmental robotics which aims at designing autonomous artificial agents capable of self-organizing their own learning trajectories based on their intrinsic motivations. It bakes the developmental framework of intrinsically motivated goal exploration processes (IMGEPs) into reinforcement learning (RL). This combination has been recently introduced as autotelic reinforcement learning, where autotelic agents are intrinsically motivated to self-represent, self-organize and autonomously learn about their own goals. Naturally, such agents need to be endowed with good exploration capabilities as they need to first physically encounter a certain goal in order to take ownership of and learn about it. Unfortunately, discovering interesting behavior is usually tricky, especially in hard exploration setups where the rewarding signals are parsimonious, deceptive or adversarial. In such scenarios, the agents’ physical situatedness-in the Piagetian sense of the term-seems insufficient. Luckily, research in developmental psychology and education sciences have been praising the remarkable role of socio-cultural signals in the development of human children. This social situatedness-in the Vygotskyan sense of the term-enhances the toddlers’ exploration capabilities, creativity and development. However, deep \rl considers social interactions as dictating instructions to the agents, depriving them from their autonomy. This research introduces \textit{teachable autotelic agents}, a novel family of autonomous machines that can learn both alone and from external social signals. We formalize such a family as a hybrid goal exploration process (HGEPs), where autotelic agents are endowed with an internalization mechanism to rehearse social signals and with a goal source selector to actively query for social guidance. The present manuscript is organized in two parts. In the first part, we focus on the design of teachable autotelic agents and attempt to leverage the most important properties that would later serve the social interaction. Namely, we introduce predicate-based autotelic agents, a novel family of autotelic agents that represent their goals using spatial binary predicates. These insights were based on the Mandlerian view on the prelinguistic concept acquisition suggesting that toddlers are endowed with some innate mechanisms enabling them to translate spatio-temporal information into an iconic static form. We show that the underlying semantic representation plays a pivotal role between raw sensory inputs and language inputs, enabling the decoupling of sensorimotor learning and language grounding. We also investigate the design of such agents' policies and state-action value functions, and argue that combining Graph Neural Networks (GNNs) with relational predicates provides a light computational scheme to transfer efficiently between skills. In the second part, we formalize social interactions as a goal exploration process. We introduce Help Me Explore (HME), a novel social interaction protocol where an expert social partner progressively guides the learning agent beyond its zone of proximal development (ZPD). The agent actively selects to query its social partner whenever it estimates that it is not progressing enough alone. It eventually internalizes the social signals, becomes less dependent on its social partner and maximizes its control over its goal space.