TV-AIL achieves a horizon-independent imitation gap of O(min{1, sqrt(|S|/N)}) via stage-coupled dynamic programming analysis on locomotion-abstracted MDPs.
The implementation of TV-AIL is based on an existing AIL algorithm DAC [Kostrikov et al., 2019] (https://github.com/google-research/ google-research/tree/master/value_dice)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2022 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Understanding Adversarial Imitation Learning in Small Sample Regime: A Stage-coupled Analysis
TV-AIL achieves a horizon-independent imitation gap of O(min{1, sqrt(|S|/N)}) via stage-coupled dynamic programming analysis on locomotion-abstracted MDPs.