Master's Thesis at ULB+VUB
Digital version available here.
Please bear in mind that the code in this repository should be refactored quite heavily soon to make it more general than what it is now - quite focused for the experiments made in the thesis.
There have been incredible advances in the field of reinforcement learning in recent years. Computers keep getting closer to the human level benchmark on many tasks, sometimes even outperforming humans at famously complicated tasks such as playing the game of Go or driving cars under certain circumstances. Although many of these breakthroughs are attributed to machine learning, paradoxically, very few attempts have been made to teach a machine to learn, as opposed to teaching a machine to solve a task. This work reviews the state of the art in meta reinforcement learning, which is the art of teaching a machine to learn. An implementation is made available, along with experiments from the literature and their results. The main contributions of this work are the application and analysis of meta reinforcement learning to a new class of continuous problems derived from the CartPole environment, identifying key dynamics and pathologies related to such problems and proposing a simple solution to enable meta learning on problems of this type. Experiments showing the positive effect of meta reinforcement learning on unseen tasks are presented.