Implicit Quantile Networks are a distributional RL method that model the distribution of returns using quantile regression. They were introduced in the paper [arxiv:1806.06923] and replaced the fixed parametrization of the quantile q-function of Quantile-Regression DQN [arxiv:1710.10044] with uniformly sampled quantile fractions.
For the generation of equally spaced quantile fractions as in QR-DQN in coax have a look at coax.utils.quantiles
. For uniformly distributed quantile fractions as in IQN there is the coax.utils.quantiles_uniform
function.
iqn.py
iqn.py