Comparison of Various Meta-Learning Paradigms in Few-Shot Preference Based Reinforcement Learning

As part of our course project for EECS 545: Machine Learning at the University of Michigan, my peers and I tackled the problem of Few-Shot Preference-Based Reinforcement Learning, primarily looking into Joey Hejna’s work on Few-Shot Preference-Based RL. We replicate their work as well as look into comparing various meta-learning paradigms for this task, including REPTILE, Model-Agnostic Meta-Learning (MAML), and an iterated variant of MAML. Details of our work can be found in our poster and our report. Our code is also published on GitHub.