Learning to Adapt: Meta-Learning for Model-Based Control

版权声明：Copyright reserved to Hazekiah Wang ([email protected]) https://blog.csdn.net/u010909964/article/details/84949014

sudden changes in environment cause failure
if encounter pertub in past experience, can in pri. learn to adapt
study model-based online adaptation
sample efficient than model-free
alleviate a challenge: global model for all state
prior adaptive control works in linear
contrib meta+model-based -> fast a
a global model using recent exper. to fast adapt

model-free large data impractical for real
model-learning challenge: global model acc
model inacc: not generalize to parts
online adaptation

model-based rl, opt dynamics model via log-like max
data distribution mismatch?

update rule
a task: a traj
learn a prior of a environment distribution