perc_ema#
- hybrid_aif.train.perc_ema(state, x, perclo, perchi, decay, cap)[source]#
Percentile-EMA return normalization:
(new state, offset, invscale).stateis the(low, high)pair of percentile EMAs, updated from this batch’s returns before use.invscale = max(1/cap, high − low), so a small spread is never amplified.