Quantile Regression versus Naïve Discretization
(Effect of discretization)
More sophisticated discretization schemes such as quantile regression improve performance on tasks where precise actuation is important, such as the highest running speed of the HalfCheetah benchmark.
Naïve discretization, while simpler to implement, performs approximately as well on all other benchmark tasks.
We compare both Transformer Trajectory Optimization (TTO) discretization schemes to the concurrently-developed Decision Transformer [Chen et al., 2021].