Exponentially Weighted Averages
vt=βvt−1+(1−β)θt
v
t
=
β
v
t
−
1
+
(
1
−
β
)
θ
t
=β[βvt−2+(1−β)θt−1]+(1−β)θt
=
β
[
β
v
t
−
2
+
(
1
−
β
)
θ
t
−
1
]
+
(
1
−
β
)
θ
t
=β2vt−2+β(1−β)θt−1+(1−β)θt
=
β
2
v
t
−
2
+
β
(
1
−
β
)
θ
t
−
1
+
(
1
−
β
)
θ
t
=βnvt−n+(1−β)∑i=0n−1βiθt−i
=
β
n
v
t
−
n
+
(
1
−
β
)
∑
i
=
0
n
−
1
β
i
θ
t
−
i
令
n=⌊11−β⌋,
n
=
⌊
1
1
−
β
⌋
,
则
limβ→1βn=limβ→1β⌊11−β⌋=limβ→1β11−β=1e
lim
β
→
1
β
n
=
lim
β
→
1
β
⌊
1
1
−
β
⌋
=
lim
β
→
1
β
1
1
−
β
=
1
e
因此
vt≈1evt−n+1n∑i=0n−1βiθt−i
v
t
≈
1
e
v
t
−
n
+
1
n
∑
i
=
0
n
−
1
β
i
θ
t
−
i
≈1n∑i=0n−1θt−i
≈
1
n
∑
i
=
0
n
−
1
θ
t
−
i
Bias Correction
vt=⎧⎩⎨0,11−βt[βvt−1+(1−β)θt],t=0otherwise v t = { 0 , t = 0 1 1 − β t [ β v t − 1 + ( 1 − β ) θ t ] , otherwise
性质
v1=θ1
v
1
=
θ
1
limt→+∞(1−βt)=1
lim
t
→
+
∞
(
1
−
β
t
)
=
1