Policy based performance management for highly scalable O-RAN networks
A method for enhanced radio resource management of radio access network (RAN) based on machine-learning-based technique includes: deploying a trained machine-learning-based model trained on a plurality of RAN-related parameters for a selected user equipment (UE) on the RAN, including i) plurality of trained weights for determining scheduling priority of the selected UE, ii) a network operator policy influencing scheduling priority of the selected UE, and iii) at least one of Packet Delay Budget (PDB), a target guaranteed bit rate (GBR), and proportional fair (PF) metric of the selected UE; computing, at the RIC, a difference between two consecutive overall error functions for PDB, GBR, and PF calculated based on corresponding observed values at two consecutive sampling time points; and updating, at the RIC, the plurality of weights based on the difference between the first and second overall error functions.
1 . A method for implementing enhanced radio resource management of radio access network (RAN) based on machine-learning-based technique, comprising:
deploying, from a RAN Intelligent Controller (RIC) to a distributed unit (DU) of the RAN, a trained machine-learning-based model trained on a plurality of RAN-related parameters for a selected user equipment (UE) on the RAN, wherein the trained machine-learning-based model comprises the plurality of RAN-related parameters including i) plurality of trained weights for determining scheduling priority of the selected UE, ii) a network operator policy influencing scheduling priority of the selected UE, and iii) at least one of Packet Delay Budget (PDB), a target guaranteed bit rate (GBR), and proportional fair (PF) metric of the selected UE;
sending, from the DU to the RIC, values of the plurality of RAN-related parameters for the selected UE observed at least at a first sampling time point and a sequentially following second sampling time point;
computing, at the RIC, a difference between 1) a first value of overall error function for PDB, GBR, and PF calculated based on corresponding observed values at the first sampling time point, and 2) a second value of overall error function for PDB, GBR, and PF calculated based on corresponding observed values at the second sampling time point; and
updating, at the RIC, the plurality of weights based on the difference between the first and second values of the overall error function.
2 . The method according to claim 1 , wherein the RIC deploying the trained machine-learning model is a near-real time RIC.
3 . The method according to claim 2 , further comprising:
implementing an RIC subscription procedure between the RIC and the DU to facilitate the sending of the values of the plurality of RAN-related parameters from the DU to the RIC.
4 . The method according to claim 2 , wherein:
i) for UE-experienced delay falling within the PDB, corresponding error term for PDB is considered as zero; and
ii) for UE-experienced delay exceeding the PDB, corresponding error term for PDB is square of deviation of the UE-experienced delay from the PDB.
5 . The method according to claim 4 , wherein:
i) for achieved GBR exceeding the target GBR, corresponding error term for GBR is considered as zero; and
ii) for achieved GBR falling below the target GBR, corresponding error term for GBR is square of deviation of the achieved GBR from the target GBR rate.
6 . The method according to claim 5 , wherein:
i) for allocated PRBs to the selected UE exceeding average PRBs per UE in the network, corresponding error term for PF is considered as zero; and
ii) for allocated PRBs to the selected UE falling below average PRBs per UE in the network, corresponding error term for PF is square of deviation between the number of allocated PRBs to the selected UE and the average PRBs per UE in the network.
7 . The method according to claim 3 , wherein the plurality of RAN-related parameters sent from the DU to the RIC include:
W GBR representing current weight of priority value P GBR for GBR;
W PDB representing current weight of priority value P PDB for PDB;
W PF representing current weight of priority value P PF for PF;
PDB UE representing packet delay budget corresponding to the selected UE's Quality of Service (QOS) flow;
Delay UE representing overall delay experienced by the selected UE;
Delay UE =MidhaulDelay UE +BackhaulDelay UE +QDelay CU-UP +QDealy RLC ;
TarGbr UE representing the target GBR of the UE;
AchvGbr UE representing current achieved bit rate of the selected UE;
Nprb UE representing the number of physical resource blocks (PRBs) allocated to the selected UE in a transmission time interval (TTI);
AvgPrb UE representing the ratio of total PRBs to the number of active UEs in the RAN; and
the network operator policy.
8 . The method according to claim 7 , wherein observed values of the plurality of RAN-related parameters for the selected UE are sent at each one of a plurality of sequential time intervals T fdbk .
9 . The method according to claim 8 , wherein:
the overall error function for PDB, GBR, and PF is a sum of error functions corresponding to PDB, GBR, and PF;
the overall error function for PDB, GBR, and PF is computed over at least two consecutive sampling time points corresponding to at least two consecutive T fdbk intervals; and
W GBR , W PDB , and W PF are updated based on the difference between the computed overall error function values for the at least two consecutive sampling time points.
10 . The method according to claim 9 , wherein:
W GBR , W PDB , and W PF are optimized by minimizing the overall error function for PDB, GBR, and PF over K 1 sampling time points corresponding to K 1 number of T fdbk intervals, as follows:
Min
∑
i
=
1
K
1
[
(
P
D
B
UE
i
-
Delay
UE
i
)
2
+
(
T
a
r
G
b
r
UE
i
-
A
c
h
v
G
b
r
UE
i
)
2
+
(
A
v
g
P
r
b
UE
-
N
P
r
b
UE
i
)
2
]
.
11 . A system for implementing enhanced radio resource management of radio access network (RAN) based on machine-learning-based technique, comprising:
a RAN Intelligent Controller (RIC) configured to deploy, to a distributed unit (DU) of the RAN, a trained machine-learning-based model trained on a plurality of RAN-related parameters for a selected user equipment (UE) on the RAN, wherein the trained machine-learning-based model comprises the plurality of RAN-related parameters including i) plurality of trained weights for determining scheduling priority of the selected UE, ii) a network operator policy influencing scheduling priority of the selected UE, and iii) at least one of Packet Delay Budget (PDB), a target guaranteed bit rate (GBR), and proportional fair (PF) metric of the selected UE;
wherein the DU is configured to send, to the RIC, values of the plurality of RAN-related parameters for the selected UE observed at least at a first sampling time point and a sequentially following second sampling time point;
and wherein the RIC is further configured to:
a) compute a difference between 1) a first value of overall error function for PDB, GBR, and PF calculated based on corresponding observed values at the first sampling time point, and 2) a second value of overall error function for PDB, GBR, and PF calculated based on corresponding observed values at the second sampling time point; and
b) update the plurality of weights based on the difference between the first and second values of the overall error function.
12 . The system according to claim 11 , wherein the RIC configured to deploy the trained machine-learning model is a near-real time RIC.
13 . The system according to claim 12 , wherein the RIC and the DU are configured to implement an RIC subscription procedure between the RIC and the DU to facilitate the sending of the values of the plurality of RAN-related parameters from the DU to the RIC.
14 . The system according to claim 12 , wherein:
i) for UE-experienced delay falling within the PDB, corresponding error term is considered as zero; and
ii) for UE-experienced delay exceeding the PDB, corresponding error term is square of deviation of the UE-experienced delay from the PDB.
15 . The system according to claim 14 , wherein:
i) for achieved GBR exceeding the target GBR, corresponding error term is considered as zero;
ii) for achieved GBR falling below the target GBR, corresponding error term is square of deviation from the target GBR rate.
16 . The system according to claim 15 , wherein:
i) for allocated PRBs to the selected UE exceeding average PRBs per UE in the network, corresponding error term for PF is considered as zero; and
ii) for allocated PRBs to the selected UE falling below average PRBs per UE in the network, corresponding error term for PF is square of deviation between the number of allocated PRBs to the selected UE and the average PRBs per UE in the network.
17 . The system according to claim 13 , wherein the plurality of RAN-related parameters sent from the DU to the RIC include:
W GBR representing current weight of priority value P GBR for GBR;
W PDB representing current weight of priority value P PDB for PDB;
W PF representing current weight of priority value P PF for PF;
PDB UE representing packet delay budget corresponding to the selected UE's Quality of Service (QOS) flow;
Delay UE representing overall delay experienced by the selected UE;
Delay UE =MidhaulDelay UE +BackhaulDelay UE +QDelay CU-UP +QDealy RLC ;
TarGbr UE representing the target GBR of the UE;
AchvGbr UE representing current achieved bit rate of the selected UE;
Nprb UE representing the number of physical resource blocks (PRBs) allocated to the selected UE in a transmission time interval (TTI);
AvgPrb UE representing the ratio of total PRBs to the number of active UEs in the RAN; and
the network operator policy.
18 . The system according to claim 17 , wherein observed values of the plurality of RAN-related parameters for the selected UE are sent at each one of a plurality of sequential time intervals T fdbk .
19 . The system according to claim 18 , wherein:
the overall error function for PDB, GBR, and PF is a sum of error functions corresponding to PDB, GBR, and PF;
the overall error function for PDB, GBR, and PF is computed over at least two consecutive sampling time points corresponding to at least two consecutive T fdbk intervals; and
W GBR , W PDB , and W PF are updated based on the difference between the computed overall error function values for the at least two consecutive sampling time points.
20 . The system according to claim 19 , wherein:
W GBR , W PDB , and W PF are optimized by minimizing the overall error function for PDB, GBR, and PF over K 1 sampling time points corresponding to K 1 number of T fdbk intervals, as follows:
Min
∑
i
=
1
K
1
[
(
P
D
B
UE
i
-
Delay
UE
i
)
2
+
(
TarGb
r
UE
i
-
A
c
h
v
G
b
r
UE
i
)
2
+
(
A
v
g
P
r
b
UE
-
N
P
r
b
UE
i
)
2
]
.