With the advent of deep learning supported by substantial technological advances, the Artificial Intelligence has taken a decisive step towards the automation of large dimension tasks. Reinforcement learning has been revolutionized thanks to new representation concepts introduced by deep learning. However, the extension of this paradigm application to the real world has triggered new challenges of generalization and optimization associated with higher level of tasks non-stationarity. In this thesis, we are interested in the recent methodological evolution of machine learning towards meta-learning in order to remedy the deep learning limits. The proposed approach is built on the basis of a Markovian formulation gradually evolving along 2 axe...