Bandwidth part (BWP) and power scheduling method and apparatus for indoor terahertz
View Patent ↗A bandwidth part (BWP) and power scheduling method and apparatus for indoor terahertz are provided. The BWP and power scheduling method uses an intelligent reflecting surface (IRS) and a non-orthogonal multiple access (NOMA) to compensate for a shortcoming, wherein terahertz waves are not suitable for an indoor transmission, convert an indoor environment of an IRS-assisted terahertz propagation into a channel model, and convert an optimal wireless resource scheduling in an indoor IRS-assisted terahertz system into a functional optimization design with random constraints. Policies and corresponding actions are trained by using a proximal policy optimization (PPO) to schedule a power and a BWP. An actor part generates continuous a power scheduling, and a critic part generates a discrete BWP scheduling.
1. A bandwidth part (BWP) and power scheduling method for an indoor terahertz, comprising the following steps:
S 1 : initializing parameters and states of a wireless access device, users, and an intelligent reflecting surface (IRS), wherein the parameters comprises an initial transmit power p 0 , a policy parameter θ 0 :={μ 0 ,σ 0 }, and Lagrange multipliers λ 0 ,ϕ 0 , wherein μ 0 denotes expectation of an initial policy, and σ 0 denotes variance of the initial policy;
S 2 : extracting samples {circumflex over (x)} 1 , {circumflex over (x)} 2 , {circumflex over (θ)} a truncated standard normal distribution;
S 3 : calculating a terahertz path loss L(ƒ,d) and a channel gain h j,k , wherein paths comprises a first transmission path from the wireless access device to each user, and a second transmission path from the wireless access device to the each user through a reflection by the IRS, and wherein ƒ denotes carrier frequency, and d denotes distance, j denotes an index of an antenna of an access point, k denotes an index of a current user;
S 4 : inputting initial conditions to a resource scheduling formula;
S 5 : inputting a channel state obtained from a channel state matrix H JK to a proximal policy optimization (PPO) algorithm, wherein J denotes a number of antennas, and K denotes a number of users;
S 6 : calculating an estimate of a gradient according to an obtained new policy parameter set θ and a generated scheduling action;
S 7 : updating primal-dual variables by using a primal-dual optimization algorithm, until a reward converges;
S 8 : outputting an action obtained after learning, a scheduled BWP factor β and a power p, and a final system performance function ƒ 0 (x), wherein x denotes an ergodic mean;
S 9 : traversing the users, and repeating step S 2 to step S 7 , to complete a scheduling for the users; and
S 10 : terminating the BWP and power scheduling method process, and fetching data.
2. The BWP and power scheduling method according to claim 1 , wherein in step S 3 , channel estimations are integrated:
H j,k ≙H LOS,j,k +H j,r ΦH r,k ,
to unify the first transmission path and the second transmission path;
wherein H j,k is a first channel gain from the wireless access device to the each user, H r,k is a second channel gain from the IRS to the each user, H j,r is a third channel gain from the wireless access device to the IRS, Φ is a phase transition matrix, and H LOS,j,k is a fourth channel gain without passing through the IRS.
3. The BWP and power scheduling method according to claim 1 , wherein in step S 4 , the resource scheduling formula is expressed as a long-term instantaneous system performance function using the ergodic mean x capture system: x≤E[f(p(h),h)], wherein f(p(h),h) is an instantaneous performance function, wherein h denotes a channel state.
4. The BWP and power scheduling method according to claim 1 , wherein in step S 5 , the channel state is inputted to an actor part and a critic part of the PPO algorithm, wherein the actor part generates a new policy function, and selects a Gaussian policy to output a choice for scheduling of the action, and the critic part generates a new action according to an original Q value function, to schedule a BWP factor.
5. The BWP and power scheduling method according to claim 1 , wherein in step S 6 , the estimate of the gradient is calculated according to the obtained new policy parameter set θ and the generated scheduling action:
∇
^
f
0
(
x
0
)
:=
f
^
0
(
x
0
+
α
1
x
^
1
)
-
f
^
0
(
x
0
)
α
1
x
^
1
,
∇
^
f
2
(
x
0
)
:=
f
^
2
(
x
0
+
α
2
x
^
2
)
-
f
^
2
(
x
0
)
α
2
x
^
2
T
,
∇
^
θ
E
[
f
1
(
π
(
h
,
θ
)
,
h
)
]
:=
f
^
1
(
π
(
h
,
θ
0
+
α
3
θ
^
)
,
h
)
-
f
^
1
(
π
(
h
,
θ
0
)
,
h
)
α
3
θ
^
T
;
wherein x 0 ,θ 0 are given points, {circumflex over (x)} 1 , {circumflex over (x)} 2 , {circumflex over (θ)} are sampling points, α 1 ,α 2 ,α 3 are iteration factors of a gradient estimation;
h is the channel state; π(h,θ) is an allocation policy;
(π(h,θ),h) is an observation value of an instantaneous system performance function;
(x 0 ) is an observation value of a constraint function;
{circumflex over (ƒ)} 0 (x 0 ) is an observation value of a target function; and
{circumflex over (∇)}ƒ 0 (x 0 ), {circumflex over (∇)}f 2 (x 0 ), {circumflex over (∇)} θ E[f 1 (π(h,θ),h)] are gradient observation values obtained through a calculation; and
in step S 7 , Lagrangian operators λ k+1 and ϕ k+1 are updated based on step sizes γ 1 ,γ 2 ,γ 3 ,γ 4 , and a Lagrangian operator updating formula is as follows:
θ
k
+
1
=
θ
k
+
γ
1
∇
^
θ
E
[
f
1
(
π
(
h
,
θ
k
)
,
h
)
λ
k
]
,
x
k
+
1
=
x
k
+
γ
2
(
∇
^
f
0
(
x
k
)
+
∇
^
f
2
(
x
k
)
ϕ
k
-
1
)
,
λ
k
+
1
=
λ
k
-
γ
f
(
f
^
1
(
π
(
h
,
θ
k
+
1
)
,
h
)
-
x
k
+
1
)
,
ϕ
k
+
1
=
ϕ
k
-
γ
4
f
^
2
(
x
k
+
1
)
,
θ k is a parameter set of a k th cycle;
x k is an ergodic mean of the k th cycle; and
λ k ,ϕ k are Lagrange multipliers of the k th cycle.