Analysis of GAIL hyperparameters

For each environment, we sample 1000 different configurations of hyperparameters at random, and report statistics over the configurations.

Success rate learning curves


Hyperparameters influence breakdown

For each hyperparameter, we aggregate all configurations for each possible value of the hyperparameter and compute the following statistic:
the average of final success rate for the top 50% of configurations (since poorly performing configurations are uninformative). Some
plot bars are "missing", it is because the success rate is 0 for all configurations.

Analysis of discriminator number of layers


Analysis of discriminator number of hidden units per layer


Analysis of regularization


Analysis of discriminator input dropout rate


Analysis of discriminator hidden dropout rate


Analysis of observation normalization


Analysis of discriminator learning rate


Analysis of GAIL reward balance


Analysis of learning rate


Analysis of reward scale


Analysis of n step


Analysis of tau


Analysis of discriminator weight decay