Meta Reinforcement Learning
Lilian Weng
Meta-reinforcement learning is a research approach where agents are trained across distributions of related tasks to quickly adapt to new unseen tasks with minimal exposure. The field traces back to 2001 work by Hochreiter et al. on meta-learning, with modern formulations like Wang et al. (2016) and Duan et al. (2017) using recurrent neural networks with memory to internalize task dynamics. Key techniques include LSTM-based policies that incorporate previous actions and rewards, meta-learning algorithms like MAML and meta-gradient RL, and diverse task distributions during training to enable rapid adaptation at test time.
Why it matters
In my earlier post on meta-learning, the problem is mainly defined in the context of few-shot classification. Here I would like to explore more into cases when we try to “meta-learn” Reinforcement Learning (RL) tasks by developing an agent that can solve unseen tasks fast and efficiently.
