Abstract
Reinforcement Learning (RL) is known to be sample-inefficient, and RL agents often do not generalise to new environments. Meta-Reinforcement Learning (Meta-RL) is one solution to these issues. Meta-RL is about learning how to reinforcement learn; the model learns parts of the RL algorithm instead of us designing the algorithm ourselves. Ideally, this should allow the model to generalise to unseen similar tasks and learn faster since it has “learned how to learn”. This talk covers the fundamentals of Meta-RL, starting with the definition before moving on to the different categories (such as few-shot and many-shot Meta-RL). Finally, we conclude with how I used Meta-RL in my Master’s research.
Bio
Batsi Ziki is a Research Scientist at Bytefuse. He completed his MSc at the University of Cape Town (UCT) under the supervision of Professor Jonathan Shock and co-supervision of Andries Smit. His master’s research focused on meta-learning the intrinsic reward weighting in curiosity-driven reinforcement learning. Currently, his research interests include continual learning and meta-learning.