focus on communication and queueing systems formulated problem: optimize the time averages of certain quantities subject to time average constraints on other quantities.
Example opportunistic scheduling problem
Fig. 1 The 2-user wireless system.
Channel-aware scheduling is called opportunistic scheduling. We assume the network controller can observe
S
(
t
)
\mathcal{S}(t)
S(t) at the beginning of each slot
t
t
t before making a transmission decision. Queueing dynamics are given by
Q
k
(
t
+
1
)
=
max
[
Q
k
(
t
)
−
b
^
k
(
p
(
t
)
,
S
(
t
)
,
0
]
+
a
k
(
t
)
∀
k
∈
{
1
,
2
,
⋯
}
Q_k(t+1)=\max[Q_k(t)-\hat{b}_k(p(t),\mathcal{S}(t),0]+a_k(t)\quad \forall k\in\{1,2,\cdots\}
Qk(t+1)=max[Qk(t)−b^k(p(t),S(t),0]+ak(t)∀k∈{1,2,⋯}
EXAMPLE PROBLEM 1: MINIMIZING TIME AVERAGE POWER SUBJECT TO STABILITY
Denote
p
ˉ
k
\bar{p}_k
pˉk as the time average power expenditure.
The objective is to minimize the time average power expenditure, and the problem is subject to queue stability, which is formulated as The designed algorithm based on our theory is utilized to make decisions
p
(
t
)
∈
P
\bm{p}(t)\in \mathcal{P}
p(t)∈P every slot
t
t
t, without requiring a priori knowledge of the probabilities associated with the arrival and channel processes
a
(
t
)
\bm{a}(t)
a(t) and
S
(
t
)
\bm{S}(t)
S(t). Decouple? Furthermore,
O
(
V
)
O(V)
O(V) is a tradeoff for the average queue backlog and delay.
EXAMPLE PROBLEM 2: MAXIMIZING THROUGHPUT SUBJECT TO TIME AVERAGE POWER CONSTRAINTS
We assume that the arrival process
a
(
t
)
\bm{a}(t)
a(t) can be controlled by a flow control mechanism. Thus the decision vectors are the power allocation vector
p
(
t
)
\bm{p}(t)
p(t) and the data admission vector
a
(
t
)
\bm{a}(t)
a(t). Denote
a
ˉ
k
\bar{a}_k
aˉkas the time average admission rate (in bits/slot) of user
k
k
k.
A problem that aims to maximize a weighted sum of throughput subject to average power constraints is formulated as
EXAMPLE PROBLEM 3: MAXIMIZING THROUGHPUT-UTILITY
SUBJECT TO TIME AVERAGE POWER CONSTRAINTS
The objective is to maximize a concave function of throughput, i.e., utility function (效用函数). 效用函数是严格凹函数并且非递减函数比如
g
(
a
)
=
log
(
1
+
a
)
g(a)=\log(1+a)
g(a)=log(1+a)。这意味着随着
a
a
a 的增大,效用函数值增加的速度是放缓的,即受到了一个diminsbing returns。在本问题中,当
a
ˉ
1
<
a
ˉ
2
\bar{a}_1 < \bar{a}_2
aˉ1<aˉ2 时,增大
a
ˉ
1
\bar{a}_1
aˉ1 比 增大
a
ˉ
2
\bar{a}_2
aˉ2 使目标函数值上升更多。效用函数的构造涉及比例公平。
一般随机优化问题
上述举例的问题仅考虑了在平均时间约束下优化平均时间,现在我们提出更一般的随机优化问题。考虑一个随机网络及多个个离散的时隙,网络能被描述为队列积压集合,即
Q
(
t
)
=
{
Q
1
(
t
)
,
⋯
,
Q
K
(
t
)
}
\mathcal{Q}(t)=\{Q_1(t),\cdots,Q_K(t)\}
Q(t)={Q1(t),⋯,QK(t)}。
K
=
0
K=0
K=0 表示系统无队列。首先定义三个属性向量 其中,
ω
(
t
)
\omega(t)
ω(t) 是在时隙
t
t
t 被观测到的随机事件(例如新数据包的到达或信道状态),
α
(
t
)
\alpha(t)
α(t) 是动作。令
x
ˉ
m
,
y
ˉ
l
,
e
ˉ
j
\bar{x}_m, \bar{y}_l, \bar{e}_j
xˉm,yˉl,eˉj 分别代表
x
m
(
t
)
,
y
l
(
t
)
,
e
j
(
t
)
x_m(t),y_l(t),e_j(t)
xm(t),yl(t),ej(t) 的平均时间。