Data

We provide the full experiment data. For each experiment a compressed container contains the the folders with the different experimental runs corresponding to different seeds and hyperparameters. Each of these contains the following files:

params.json : experiment configuration file specifying the hyperparameters and seeds of the run
progress.csv : tabular file containing the measurements and quantitative results that are used to generate the plots in the paper
log.txt: log file of the meta-training
params.pkl: serialized dictionary containing the trained policy, baseline, and the environment (pickled with joblib)

Data

Benchmark Study of Meta-Policy Search Methods

Gradient Variance of Curvature Estimators (DICE vs. LVC)

Comparison of Initial Sampling Distributions

Gradient Update Directions of Meta-RL Formulations