Centrala begrepp
Increasing state-action coverage has a greater impact on data efficiency than reward density in reinforcement learning with dynamics-invariant data augmentations.
Sammanfattning
Data augmentation (DA) is a valuable technique in reinforcement learning (RL) to improve data efficiency. This study isolates three aspects of DA - state-action coverage, reward density, and augmented replay ratio - to understand their impact on performance. Increasing state-action coverage is more beneficial than increasing reward density for data efficiency. Decreasing the augmented replay ratio significantly improves learning outcomes. Experimental results demonstrate the importance of these factors in successful DA implementation.
Statistik
"increasing state-action coverage often has a much greater impact on data efficiency than increasing reward density"
"decreasing the augmented replay ratio substantially improves data efficiency"