You are on page 1of 6

2022 IEEE Wireless Communications and Networking Conference (WCNC)

Joint Selection of Local Trainers and Resource


Allocation for Federated Learning in Open RAN
Intelligent Controllers
2022 IEEE Wireless Communications and Networking Conference (WCNC) | 978-1-6654-4266-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/WCNC51071.2022.9771700

Amardip Kumar Singh, Kim Khoa Nguyen


Synchromedia Lab , École de Technologie Supérieure, Montreal, Canada
amardip-kumar.singh.1@ens.etsmtl.ca, kim-khoa.nguyen@etsmtl.ca

Abstract—Recently, Federated Learning (FL) has been applied collected periodically, RAN performance can be improved by
in various research domains specially because of its privacy pre- incorporating AI capabilities. O-RAN is a newly introduced
serving and decentralized approach of model training. However, radio access network architectural framework that supports
very few FL applications have been developed for the Radio
Access Network (RAN) due to the lack of efficient deployment different use cases of 5G services through network slicing and
models. Open RAN (O-RAN) promises a high standard of disaggregation. The key element of O-RAN is its intelligent
meeting 5G services through its disaggregated, hierarchical, and controllers (RICs) that monitor and support the guaranteed per-
distributed network function processing framework. Moreover, it formance for different slice user groups [3]. On the other hand,
comes with built-in intelligent controllers to instill smart decision O-RAN operates on multi-vendor and shared resource system.
making ability into RAN. In this paper, we propose a framework
named O-RANFed to deploy and optimize FL tasks in O-RAN Therefore, it has to function under minimal resource cost and
to provide 5G slicing services. To improve the performance of yet with guaranteed service delivery. These constraints impose
FL we formulate a joint mathematical optimization model of heavy challenges posed for the O-RAN RICs.
local learners selection and resource allocation to perform model In this paper, we investigate the possibility of FL model
training in every iteration. We solve this non-convex problem training in O-RAN to provide 5G slicing services. 5G services
using the decomposition method. First, we propose a slicing based
and deadline aware client selection algorithm. Then, we solve the are governed by slicing of the physical network into several
reduced resource allocation problem by using successive convex logical isolated self-adaptable networks hosted by general
approximation (SCA) method. Our simulation results show the processors located at cloud-based data centres [4]. 5G defines
proposed model outperforms the state-of-the-art FL methods three classes of services based on their quality of services
such as FedAvg and FedProx in terms of convergence, learning (QoS) metrics: ultra-Reliable Low Latency Communication
time, and resource costs.
Index Terms—Federated Learning, O-RAN, 5G, Resource
(uRLLC), extreme Mobile Broad Band (eMBB), and massive
Allocation, RAN Intelligent Controller, Network Slicing, RIC Machine Type Communications (mMTC). Accordingly, the
physical RAN infrastructure is also divided into three logi-
I. I NTRODUCTION cal slices of network elements that are assigned and main-
tained dynamically [1]. The two kinds of O-RAN Intelligent
Federated Learning (FL) is a new approach fusing the Controllers namely, near-realtime RAN Intelligent Controller
concepts of distributed computing and Artificial Intelligence (near-RT-RIC) and Non-RT-RIC coordinate with each other to
[1]. Unlike the traditional machine learning in which models select the slice specific local training points and then train the
are trained on a centralized data set, FL does not require the FL models.
data sets to be located at one central point. It collaborates with In Fig. 1, we map the FL framework via RICs’ specifications
the local computing nodes through a global aggregation point of O-RAN [5]. The slice operational data is collected and
by transferring only the model update vectors in a periodic saved into distributed databases through E2 interface. O1 inter-
communication mode. Therefore, it respects the privacy of face transfers data to near RT-RICs for local processing. The
the local data set, distributes the burden of computing, and interaction of FL parameters is enabled by the A1 interface
still train a centralized model [1]. These features enable between Non-RT-RIC and near-RT-RICs. Since the FL tasks
many applications of FL in edge computing where users’ are distributed over edge cloud sites in O-RAN RICs, and
devices serve as local trainers. However, it is yet to be well shared by various operators, resources required to facilitate this
investigated in carriers’ networks due to the lack of a standard learning process need to be optimized. In addition, FL model is
implementation model. The characteristics of FL pave way for trained iteratively, therefore learning time is also an issue. Due
its uses in the domain of 5G RAN, particularly to enhance the to the stochastic nature of this distributed learning process,
radio resource management policies [2]. minimizing the learning time while guaranteeing the accuracy
5G network services promise to deliver ever fast and reliable of the global model is challenging. Hence, the resource usage
user experience that too on a massive scale. This comes with cost reduction and FL time minimization should be dealt
a burden on improvements in how the access network is together for the O-RAN intelligent controllers.
managed. Using the operational and maintenance data that is A plethora of recent works have focused on the adaptive

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on May 18,2023 at 10:12:20 UTC from IEEE Xplore. Restrictions apply.
978-1-6654-4266-4/22/$31.00 ©2022 IEEE 1874
2022 IEEE Wireless Communications and Networking Conference (WCNC)

Service & Management Orchestration Functions II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
ML Model
O1 Non-RT-RIC Host O2 uRLLC Consider an O-RAN system with a single regional cloud and
slice
Local model
update vector A1
Global model
aggregated vector eMBB
a set M of M distributed edges cooperatively performing an
broadcast
uploading slice
ML Model Application FL algorithm. In this FL setup, each edge cloud uses its locally
mMTC
Near- Near-
Near-
Near-
Near-
near-
RT-RIC
Near-
near-
RT-RIC
RT-RIC
Near-
near-
RT-RIC
slice collected training data to train a local FL model. The Non-
RT-RIC
RT-RIC RT-RIC RT-RIC
RT-RIC
RT-RIC
E2 RT-RIC at the regional cloud integrates the local FL models
RAN Data Analytics & AI Platform
from participating edge clouds and generates an aggregated FL
Data
Data
Data
Data Base
Base Data
Data
Data Base
Base Data
Data
Data Base
Base
model. This aggregated FL model is further used to improve
O1 Base Base Base
Base
E2 E2 E2 local FL models of each near-RT-RIC enabling the local
O-CU-CP O-CU-CP O-CU-CP models to collaboratively perform a learning algorithm without
E1 E1 E1 transferring its training data. We call this aggregated FL model
O-CU-UP O-CU-UP O-CU-UP

F1 F1 F1
generated by using the local FL models as the global FL
O-DU O-DU O-DU model. As illustrated in Fig. 1, the uplink from near-RT-RICs
O-RU O-RU O-RU
to the Non-RT-RIC is used to send the local FL model update
parameters and the downlink is used to broadcast the global
FL model in global rounds of the training.
Fig. 1. Federated Learning set-up for O-RAN Intelligent Controllers
A. The Learning Model
In this model, each near-RT-RIC collects a dataset Di =
optimization of FL under communication constrained edge [xi,1 , ....., xi,Si ] of input data where Si is the number of the
computing. Prior work in [6], [7] and [8] have investigated input samples collected by near-RT-RIC i and each element xis
different resource constrained federated learning approaches. is the FL model’s input vector. Let yis be the output of xis . For
Their objective is to minimize the energy usage of edge simplicity, we consider an FL model with single output, which
devices involved in FL training by allocating optimal trans- can be readily generalized to a case with multiple outputs. The
mission power. The work in [8] is too general and can only output data vector for training the FL model of near-RT-RIC i
be implemented in a single edge (with a single base station) is yi = [yi,1 , ....., yi,Si ]. We assume that the data collected by
and cannot be used for a carrier network of multiple edges. each near-RT-RIC is different from the other near-RT-RICs
Moreover, the problem of resource allocation from the per- i.e. (xi 6= xj ; i 6= j, i, j ∈ M). So, each local trainer
spective of the System Management and Orchestration (SMO), will train the model using a different dataset. This is in line
which is important for O-RAN, has not been considered so with the real scenario as each local near-RT-RIC collects the
far. In this paper, we have taken into account new parameters operational data from the corresponding slice specific users.
of O-RAN architecture to design an algorithm to select local We define a vector gi to capture the parameters related to the
trainers, then allocate resources for the selected trainers, and local FL model that is trained by Si and yi . gi determines
propose an aggregation method. Our contributions in this paper the local FL model of each near-RT-RIC i. For example, in a
can be summarized as below: linear regression prediction algorithm, xi mT .yi represents the
• A mathematical formulation of the joint optimal resource output, and gi determines the prediction accuracy. The training
allocation and local trainers’ selection problem for the O- process of an FL model is done in a way to solve:
RANFed learning tasks. Then, we propose a solution for M Si
1 XX
this non-convex optimization problem using the decom- min f (gi , xis , yis ) (1)
g1 ,......,gM S i=1 s=1
position method.
• An O-RAN slicing based and deadline aware algorithm s.t. g1 = g2 = ..... = gM = g ∀i ∈ M (1a)
to select representative instances of near-RT-RIC as local PM
model participants in each global iteration of FL. where S = i=1 Si is the total size of training data of
• A FL algorithm, so called O-RANFed, for O-RAN slicing all near-RT-RICs. g is the global FL model generated by
services where near RT-RIC hosts the local training the Non-RT-RIC and f (gi , xis , yis ) is a loss function that
instances and Non-RT-RIC hosts the global aggregation captures the FL prediction accuracy. Different FL algorithms
point of the ML model. use different loss functions. Constraint (1a) ensures that, once
the FL model converges, all of the near-RT-RICs and the Non-
To the best of our knowledge, this is the first work proposed RT-RIC will transmit the parameters g of the global FL model
to optimize ML training through federated settings in O- to its connected near-RT-RICs so that they train their local FL
RAN. The remainder of this paper is organized as follows. In models. Then the near-RT-RICs will transmit their local FL
Section II, the system model and the problem formulation are models to the Non-RT-RIC to update the global FL model. The
presented. In Section III, we describe our proposed solution update of each near-RT-RIC i’s local FL model gi depends on
approach. In Section IV, we present the numerical results to all near-RT-RICs’ local FL models. The update of the local FL
evaluate the performance of our proposed solution. Finally, we model gi depends on the learning algorithm. For example, one
conclude the paper and discuss our future work. can use gradient descent or randomized coordinate descent to

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on May 18,2023 at 10:12:20 UTC from IEEE Xplore. Restrictions apply.
1875
2022 IEEE Wireless Communications and Networking Conference (WCNC)

update the local FL model. The update of the global model g C. FL Accuracy Model
is given by: The target for each of these local models is to attain a
M θ  (0, 1) level of accuracy, defined as below:
X Si .gi
g= . (2)
i=1
S t
||∇fm t
(gm t
)|| ≤ θ||∇fm t−1
(gm )||, ∀ m  {1, 2, 3, .., M } (5)

Since, we are considering wireless transmissions through A1 A near-RT-RIC takes several iterations, called local iterations,
interface between near-RT-RICs and the Non-RT-RIC [5], to attain this accuracy. In the global model placed at the non-
there is a resource constraint on the communication model RT-RIC, the target is to attain the optimal model weights to
which in turn affects the performance of FL learning algo- reach  level of global accuracy, defined as below:
rithm. Therefore, we need to jointly consider these two aspects.
|F (g t ) − F (g)| ≤  ∀ t ≥ T (6)

B. FL Resource Model Constraint (6) states that the g is the optimal model parameter
i.e. for every global round beyond T , the difference between
In each global interaction, the O-RAN system has to decide the loss function values falls within the defined accuracy level.
which local training points i.e. which near-RT-RIC to partic- Here, F (.) denotes the global loss function, defined over all
ipate. This is because at each time interval only a limited the local loss functions as:
number of clients can participate due to delay constraints M
originating from the control loops of O-RAN. Therefore,
X
∇F (g) := (|Si |/S)fi (gi ) (7)
the selected clients upload their local FL models updates i=1
depending on the wireless media. We define a binary variable
atm ∈ {1, 0} to decide whether or not the trainer m is In [9], it is proven that the number of global iterations required
selected in round t, and at = (at1 , ......, atM ) collects the to attain a level of global accuracy  and local accuracy θ can
overall trainers’ selection decisions. A selected near-RT-RIC be upper bounded by:
in round t i.e. (atm = 1), consumes compute resources to O(log(1/))
K(, θ) = (8)
train locally with the collected data. At the same time, these (1 − θ)
selected trainers also consume bandwidth resources to transmit
We use this relationship among the local accuracy level, global
update vectors. We consider the orthogonal frequency division
model accuracy, and the upper limit on the number of required
multi access (OFDMA) for local model uploading with a total
global rounds to model the FL time. In order to ensure
bandwidth B. Let btm ∈ [0, 1] be the bandwidth allocation
the convergence of the gradient descent approximation, the
ratio for trainer m in round t, hence its allocated bandwidth
following assumptions are considered on the loss functions at
is btm B.PLet bt = (bt1 , ......., btM ). Bandwidth allocation must
each near-RT-RIC training point:
satisfy m∈M btm = 1 ∀t. Clearly if atm = 0, namely trainer
(i) Fi (g) is convex.
m is not selected in round t, then no bandwidth is allocated to 0 0
(ii) Fi (g) is ρ-Lipschitz, i.e. ||Fi (g) − Fi (g )|| ≤ ρ||g − g ||,
it i.e. btm = 0. On the other hand, if atm = 1, then we require 0
for any g, g .
at least a minimum bandwidth bmin is to be allocated to the 0
(iii) Fi (g) is β-smooth, i.e, ||∆Fi (g) − ∆Fi (g )|| ≤ β||g −
trainer m i.e. btm ≥ bmin . To make the problem feasible, we 0 0
1 g ||, for any g, g .
assume bmin ≤ M . Therefore, total resource cost for using
(iv) For any g and i, the difference between the global gradient
communication bandwidth is:
and local gradient can be bounded byP
S .δ
M
X T
X ||∆Fi (g) − ∆F (g)|| ≤ δi , and δ := i S i i .
Rco = co
Rm = atm btm Bptr (3) These assumptions are in line with the recent works [10], [11],
m=1 t=1 [6], [12] on convergence analysis of FL.
for T global rounds where ptr is the unit cost of bandwidth D. Latency Model
cp
usage. For each near-RT-RIC m, let Rm denote its local
We consider synchronous communication, in other words
training compute resource cost in every round which depends
all the near-RT-RICs send their local update vectors to the
on its computing host and dataset. To process the local dataset
Non-RT-RIC before the tth round of global aggregation starts.
each near-RT-RIC uses the CPU cycle frequency of the host.
Therefore, before entering this communication round, all
Let the CPU power of mth host be fm cycles/s and the per
the near-RT-RICs must finish its local ML processing. In
unit time usage cost be pc . Then the total compute resource
each of the global round, the FL tasks are spanned over
cost is:
M T
three operations: (i) computation, (ii) communication of local
X X D m cm updates to the Non-RT-RIC using uplink, and (iii) broad-
Rcp = Rmcp
= atm pc (4)
m=1 t=1
fm cast communication to all the involved near-RT-RICs using
downlink. Let the computation time required for one local
where cm is the CPU cycles required for processing a bit of round for mth near-RT-RIC be Tm cp
, and there be Kl local
data. iterations in each interval of the global communication. Then,

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on May 18,2023 at 10:12:20 UTC from IEEE Xplore. Restrictions apply.
1876
2022 IEEE Wireless Communications and Networking Conference (WCNC)

cp
the computation time in one global iteration round is Kl Tm . each global round, (12e) imposes this criteria. The relationship
Let the communication time required in transferring the local between local accuracy and the number of global rounds
update vectors from mth near-RT-RIC to the Non-RT-RIC be is stated in (12f) where µ is a multiplication factor. (12h)
co
Tm in the uplink phase. Let dm be the datasize of the update represent the defining domain of the decision variable.
vector of mth trainer. Therefore, the learning time in one
III. P ROPOSED S OLUTION
global round of FEDL for the mth local FL model trainer
is: (12) is a non-convex optimization problem because of the
cp co
Tm = Kl .Tm + Tm ;m ∈ M (9) non-convex objective function and constraints (12e)-(12g). So,
co we decompose the problem into two sub-problems and then
Where Tm is calculated as:
use iterative solution to reach the optimal solution. We first
co dm solve the problem of trainers’ selection and then use this
Tm = ;m ∈ M (10)
btm .B solution to allocate resources optimally to these selected local
In the downlink phase, we do not consider the delay because trainers. Fig. 2 shows the scheme of the proposed solution.
it is negligible as compared to the uplink delay as a result Due to the variation in traffic patterns for different kinds
of high speed downlink communication. Let K be the total Solution Approach
Slicing based
number of global rounds to attain the global accuracy  as and deadline Solving (13) using
established in (8). Therefore, the total learning time can be aware client Algorithm 1
Original problem (12): selection (13)
modeled as: Resource cost and FL Using the
time minimization, solutions to
total sub. to optimal client implement
T = K .Tmax = K . max{Tm ; ∀m ∈ M} (11) selection and Algorithm 2
bandwidth allocation Resource
E. Problem Formulation Allocation for Solving (14) using
selected near- SCA method
Our goal is to jointly minimize the resource cost and the RT-RICs (14)
learning time under the constraints of model accuracy, and
available compute and bandwidth resources. This can be done Fig. 2. Schematic Diagram of the Proposed Solution
by optimizing the selection of local trainers i.e. near-RT-RICs,
bandwidth allocation, and number of local training rounds, as of slicing services of O-RAN, the local FL model might en-
formulated in the optimization model (12). counter inconsistency problem. This may lead to a degradation
in accurate prediction. We take into account this differentiation
P: min {(1 − ρ)Rtotal + ρ.T total } (12) and propose a trainers’ selection algorithm that respects the
at ,bt ,θ,Kl
formation of slices in O-RAN while maintaining a deadline
subject to:
awareness.
0 < θ < 1, (12a)
A. Local Trainers’ Selection
M
X
atm .btm .B ≤ B, (12b) According to the specifications defined by O-RAN Alliance,
m=1 the collected RAN operational data can be separated based on
XM their slice-user groups. Each near-RT-RIC is then fed with
btm = 1, (12c) slice specific network data. The selection of a near-RT-RIC
m=1 corresponding to a slice must be incorporated in each iteration
bmin ≤ btm ≤ 1 ; ∀mM, (12d) of gradient descent training of the model. However, not all the
cm .Dm co local models can be accommodated in each iteration because
max{ + Tm } = Tmax , (12e) of the deadline constraint (13a) and limited computational and
m fm
log(1/) bandwidth resources to be assigned for this learning task. So,
K = µ. , (12f) we propose Algorithm 1 for this selection in alignment with
(1 − θ)
the O-RAN slice definition. In this algorithm, we categorize
atm ∈ {1, 0}. (12g)
the set of near-RT-RICs into three classes corresponding to
The objective function (12) has two components balanced by eMBB, uRLLC, and mMTC slicing services.
a trade-off parameter ρ because the two goals are conflicting. Our objective in this trainers’ selection algorithm is to
The total resource cost, Rtotal = Rcp + Rco and the FL maximize the number of near-RT-RICs to participate in each
training time, T total as given by (11). Minimizing the resource global round and allow the non-RT-RIC to aggregate all
cost naturally leads to higher learning time and vice-versa. received data. This is based on the idea that a larger fraction
Constraint (12a) limits the local accuracy line. Constraint of trainers in each round saves the total time required for a
(12b) bounds the total bandwidth allocated for the FL tasks. global FL model to attain the desired accuracy performance
Constraint (12c) presents the definition of btm i.e, the sum of [13]. Let N (⊆ M) be the set of selected near-RT-RICs, tround
bandwidth fractions must be 1. (12d) denotes the boundary of be the deadline for each global round, t1 be the elapsed time to
the bandwidth fractional allocation. Since we have assumed perform Algorithm 1, and tagg be the time taken in aggregating
the synchronous communication mode of update vectors in the update parameters at the Non-RT-RIC. Therefore, the

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on May 18,2023 at 10:12:20 UTC from IEEE Xplore. Restrictions apply.
1877
2022 IEEE Wireless Communications and Networking Conference (WCNC)

Algorithm 1 : Deadline aware and Slicing based Local determined experimentally as required in attaining the local
Trainers’ Selection accuracy value θ. We solve this problem using Successive
1: Input: M : Set of all near-RT-RICs Convex Approximation (SCA) method.
2: Initialize N u , N e , N m = Φ
3: for tiround defined for i ∈ {N u , N e , N m } do C. Federated Training in O-RAN RICs(O-RANFed)
4: while |N | > 0 do Using the solutions of trainer selection and resource al-
5: x ← arg minn∈N 12 .(tk−1n + α.tkn (estimated)) location in (14), we train the FL model as described in
6: t ← t1 + tagg. + tkn Algorithm 2. In each global round, a subset of participating
7: N \ {x} local trainers are selected first followed by resource allocation,
8: if t < tiround then and then interaction of local Fl models with the global Fl
9: t ← t + tkn model. This loop continues for K iterations, which is the
10: end if maximum number of global rounds required to attain the
11: end while prefixed accuracy of the model.
12: end for
13: Output: N = N u ∪ N e ∪ N m Algorithm 2 : ORANFed
1: Initialize: Untrained local model at each near-RT-RIC i ∈
mathematical optimization problem for the trainer selection M;
becomes: 2: for k ≤ K ( accuracy) do
max{|N |} (13)
N 3: Non-RT-RIC uses Alg. 1 for client selection;
1 4: compute and bandwidth resources are assigned to se-
s.t. t1 + tagg. + (tk−1 + α.tkn ) ≤ tround . (13a)
2 n lected near-RT-RICs (N );
(13) is a combinatorial optimization problem which makes it 5: Each near-RT-RIC trains using local data till it achieves
non-trivial. So, we employ a greedy heuristic to solve this an accuracy θ and obtains gi,k ;
problem as shown in Algorithm 1. We repeat the steps in 6: Model update parameters edge clouds sent to the Non-
each global round until we get the desired accuracy. Here, the RT-RIC;
constraint (13a) restricts the violation of the deadline for every 7: Non-RT-RIC aggregates the local weights through (2);
near-RT-RIC in each global round. The deadline is assigned 8: Non-RT-RIC broadcasts the aggregated parameters;
separately for each slice-user groups while the total deadline 9: Non-RT-RIC calculates the global accuracy attained (6);
in each round is varied experimentally to observe its impact 10: end for
on overall learning time of FL model. 11: Finally trained model is sent to SMO for deployment
B. Resource Allocation
From the trainer selection phase, we obtain at i.e. a binary D. Complexity Analysis
valued vector of selected trainers in k th global round. The O-RANFed consists of trainers’ selection in step 3, and
next phase is to allocate the compute and bandwidth resources assignment of resources in step 4. Steps 5 to 9 trains the FL
to support the local training, parameters uploading, model model iteratively. So, its complexity can be analysed in two-
aggregation, and broadcast of updated model weights. For part. In the first part, (13) is solved using Algorithm 1 having
this we solve the optimization problem (12) with known time complexity of O(L) where L is the cardinality of the set
variable at . Still, (12) is a non-convex optimization prob- M. In the second part, (14) is solved using SCA approach
lem, exact solution of which is infeasible using traditional having complexity O(JSCA ) [14], where JSCA is the total
methods. Therefore, we employ an approximation approach number of iterations within the SCA algorithm.
with equivalent surrogate functions. The multiplication factor IV. N UMERICAL R ESULTS
µ in (12f) is chosen such that the whole numerator part is 1.
(12d) is replaced by an inequality preserving the same lower TABLE I
S IMULATION S ETTINGS
bound on Tmax value. With these changes and substituting the
defining expressions, the optimization problem (12) reduces Parameter Value Parameter Value
N 50 B 1M Hz
the following mathematical form: cm 15cycles/bit fm ∼ U (1, 1.6)GHz
T ptr 1 pc 1
Dm ∼ U (5, 10)M B d 1
n X
P1: tmin (1 − ρ) atm .btm .B.ptr bmin 0.1M Hz ρ (0, 1)
b ,θ,Kl
t=1
T
X Dm .cm  1 o
Federated Learning Task: We trained a prediction model
+ Kl . atm . .pc + ρ. .Kl .Tmax (14)
fm 1−θ where each near-RT-RIC processes a time series data con-
t=1
taining the volume of traffic requested by its corresponding
slice in a period of one month. The Dataset represents hourly
subject to: (12a), (12b), (12c), (12d), and (12e)
operational data of lower level network traffic. Using this
The number of local iterations (Kl ) in each global round is dataset, the trained model predicts the requirement of amount

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on May 18,2023 at 10:12:20 UTC from IEEE Xplore. Restrictions apply.
1878
2022 IEEE Wireless Communications and Networking Conference (WCNC)

of traffic in the next hour. We used Long Short Term Memory the two other methods. This helps O-RANFed in saving FL
(LSTM) based neural network with 4 layers to train this time as well as resources.
regression model. We run this training on Intel(R) Core(TM) In terms of the objectives costs (resource and time), O-
i5-8265U CPU. The model attains 96.3% (approx.) accuracy RANFed performs better than FedAvg and FedProx. Fig. 5 and
in centralized ML model. Therefore, the global accuracy  of 6 show learning time and resources consumed by each method.
the FL model is taken as 0.96. In these figures the behaviour of FL methods is plotted against
Wireless Network: In order to compare our proposed model the pareto co-efficient (ρ). We can see that the learning time
with state-of-the-art FL methods, we considered a compatible of O-RANFed is the lowest, and it is much lower for higher
wireless network setting as described in Table 1. For simplic- value of ρ. Moreover, resource cost required by O-RANFed is
ity, all near-RT-RICs have the same data processing rate cm . the lowest, and it is lower for smaller value of ρ.
We used a uniform distribution random generator for assigning
V. C ONCLUSION
CPU frequency (fm ) of the host. The maximum bandwidth
capacity (B) is 1 MHz whereas the minimum (bmin ) is kept In this paper, we proposed a federated learning method
at 0.1. To better present result and without loss of generality, designed for O-RAN slicing environment. Our model takes
the communication and compute cost (ptr , pc ) are set as unit into account the importance of slice specific local trainers as
value. The local dataset size is distributed uniformly in the well as the resource allocation for performing FL tasks. The
range of (5, 10) MB. For benchmark, we consider FedAvg simulation results show FL can be implemented to predict
[13] algorithm with fixed number of clients (N = 50), which data traffic of different slices in O-RAN. Our proposed model
is the maximum number of near-RT-RICs in each global round. outperforms state-of-the-art FL methods in terms of learning
Another prominent FL method is FedProx [15], which takes a time and resource cost in the simulations. Therefore, it can
probability distribution to select the number of clients in each be deployed in the control loops of O-RAN to guarantee the
global round. These two methods are suitable for comparative slice QoS. In future, we will investigate the location of the
analysis as one sets the upper limit on the client selection, distributed data collection points to improve O-RANFed in a
the later follows variable selection policy that differs from highly distributed environment.
our proposed ORANFed. Fig. 3 presents a comparison of the ACKNOWLEDGMENT
60 The authors thank Mitacs, Ciena, and ENCQOR for funding
0.9 O-RANFed
50 0.8 FedProx this research under the IT13947 grant.
FedAvg
0.7 R EFERENCES
Local Accuracy ( )
Number of Clients

40 0.6
0.5 [1] S. Abdulrahman and et al., “A survey on federated learning: The journey
30
0.4 from centralized to distributed on-site learning and beyond,” IEEE IoT
20 0.3 Journal, vol. 8, no. 7, pp. 5476–5497, 2021.
O-RANFed 0.2 [2] Z. Zhao and et al., “Federated-learning-enabled intelligent fog radio
10 FedProx 0.1 access networks: Fundamental theory, key techniques, and future trends,”
FedAvg 0.0 IEEE MCW, vol. 27, no. 2, pp. 22–28, 2020.
0
50 75 100 125 150 175 200 225 250 0 25 50 75 100 125 150 175 200 [3] S. K. Singh and et al., “The evolution of radio access network towards
Time elapsed Global Rounds open-ran: Challenges and opportunities,” in IEEE WCNCW, 2020, pp.
Fig. 3. Trainer Selection pattern Fig. 4. Accuracy convergence 1–6.
[4] O-RAN Alliance, “O-RAN-WG1.OAM-Architecture-v02.00,” 2019.
[5] H. Lee and et al., “Hosting ai/ml workflows on o-ran ric platform,” in
230 O-RANFed 240 2020 IEEE GCWkshps, 2020, pp. 1–6.
FedAvg [6] H. H. Yang and et al., “Scheduling policies for federated learning in
220 FedProx 220
Resource Usage Cost
Learning Time Cost

200
wireless networks,” IEEE TCOMM, vol. 68, no. 1, pp. 317–333, 2020.
210 [7] W. Shi and et al., “Joint device scheduling and resource allocation for
180 O-RANFed
200 FedAvg latency constrained wireless federated learning,” IEEE TWC, vol. 20,
160 FedProx no. 1, pp. 453–467, 2021.
190 140 [8] C. T. Dinh and et al., “Federated learning over wireless networks: Con-
120 vergence analysis and resource allocation,” IEEE/ACM TNET, vol. 29,
180
100 no. 1, pp. 398–409, 2021.
170 [9] J. Konečnỳ and et al., “Federated optimization: Distributed machine
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Pareto co-efficient ( ) Pareto co-efficient ( ) learning for on-device intelligence,” preprint arXiv:1610.02527, 2016.
Fig. 5. Resource Cost comparison Fig. 6. Learning Time Cost [10] Z. Yang and et al., “Energy efficient federated learning over wireless
communication networks,” ITWC, vol. 20, no. 3, pp. 1935–1949, 2021.
[11] S. Wang and et al., “Adaptive federated learning in resource constrained
three FL approaches in terms of the number of clients selected edge computing systems,” IEEE JSAC, vol. 37(6), pp. 1205–1221, 2019.
in each global rounds with respect to the total learning time [12] M. Chen and et al., “A joint learning and communications framework for
elapsed in the training process. FedAvg serves as the baseline federated learning over wireless networks,” IEEE TWC, vol. 20, no. 1,
pp. 269–283, 2021.
keeping a constant value. O-RANFed gradually attains the [13] B. McMahan and et al., “Communication-efficient learning of deep net-
maximum number of clients as the time progresses which works from decentralized data,” in Artificial intelligence and statistics.
shows its efficiency over FedProx. Then, we compare the PMLR, 2017, pp. 1273–1282.
[14] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge
accuracy achieved after each global round by each of the FL university press, 2004.
methods. In Fig. 4, O-RANFed takes significantly less number [15] T. Li and et al., “Federated optimization in heterogeneous networks,”
of global rounds to achieve the same accuracy compared to arXiv preprint arXiv:1812.06127, 2018.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on May 18,2023 at 10:12:20 UTC from IEEE Xplore. Restrictions apply.
1879

You might also like