Why "Symmetric" Clipping is Asymmetric in Effect
PPO clips ratio r = π
θ
/ π
old
; the per-step ceiling on π
new
is
multiplicative
: max π
new
= (1+ε)·π
old
ε
high
:
0.28
ε
low
= 0.20 fixed
π
0
:
0.010