Enum MemorySamplingStrategy

Memory sampling strategies for experience replay.

public enum MemorySamplingStrategy

Fields

Boundary = 6: Boundary-focused sampling selects examples near decision boundaries. Prioritizes hard-to-classify examples for better discrimination.
ClassBalanced = 3: Class-balanced sampling ensures equal representation of each class. Best for imbalanced datasets to prevent bias toward majority classes.
GradientBased = 7: Gradient-based sample selection.
Herding = 4: Herding-based selection picks examples closest to class means. From iCaRL paper - provides exemplars that well represent the class distribution.
KCenter = 5: K-Center coreset selection maximizes coverage of the feature space. Picks examples that minimize maximum distance to any unselected point.
Random = 1: Random uniform sampling from the dataset. Simple but effective for homogeneous data.
Reservoir = 0: Reservoir sampling - uniform random selection. Each item has equal probability of being selected.
RingBuffer = 2: Ring buffer - FIFO replacement.

For Beginners: When storing examples from previous tasks, the sampling strategy determines how examples are selected and maintained in memory.