FAROS: Forecast-Aware Risk-Informed Online Reinforcement Learning for Proactive O-RAN Slicing

1

FAROS: Forecast-Aware Risk-Informed Online Reinforcement Learning for Proactive O-RAN Slicing

arXiv:2611.12638v1 [cs.NI] 9 Nov 2026

Peyman Eftekhari , Student Member, IEEE, Seyyed Kasra Mortazavi , Student Member, IEEE, Reza Farahani , Member, IEEE, Robert Legenstein , Kyandoghere Kyamakya , Member, IEEE, and Martin Gebser , Member, IEEE

Abstract—Proactive Open Radio Access Network (O-RAN) slicing requires decisions before an approaching quality-of-service (QoS) loss becomes difficult to reverse. We present FAROS, a Forecast-Aware Risk-Informed O-RAN Slicing framework that combines an eight-step QoS forecast, training-only PCMCI+supported temporal risk, Multi-Agent Proximal Policy Optimization (MAPPO) under centralized training and decentralized execution (CTDE), symbolic feasibility constraints, and reliability-aware Top-K arbitration. MAPPO supplies the actor proposal, while A5v2 checks admissibility and one-step utility before execution. Evaluation uses ten validation-selected frozen checkpoints, 500 matched start indices per seed under the riskystart protocol, and 150,000 agent-level decisions. FAROS obtains mean reward 0.2778 and ten-step recovery 90.56% across ten seeds. A fixed model-matched diagnostic rule reaches reward 0.3015 and recovery 97.57% under the same transition–reward contract, so we make no superiority claim over direct threshold logic. A same-actor M4 shield control also reaches zero riskystate prohibited actions. In the M5 trace, A5v2 changes 22.61% of actor proposals and reduces the predicate-defined prohibitedaction rate from 16.79% at proposal time to zero at execution; this property is enforced by the arbiter rather than learned by MAPPO. Independent ns-3 experiments test the six-action interface at packet level and remain separate from end-to-end FAROS claims. Index Terms—O-RAN, network slicing, multi-agent reinforcement learning, quality-of-service optimization, forecast-aware control

I. I NTRODUCTION O-RAN slicing can require intervention before service degradation becomes clear in current KPIs. Recent work uses O-RAN programmability and DRL for multi-agent control, VNF reconfiguration, slice orchestration, QoS-aware allocation, offloading, uncertainty, and SLA-aware management [1]– [12]. Reactive control remains difficult when an approaching QoS loss is weak in the observed state. P. Eftekhari and M. Gebser are with the Institute of Artificial Intelligence and Cybersecurity, University of Klagenfurt, Klagenfurt, Austria (e-mails: peyman.eftekhari@aau.at; Martin.Gebser@aau.at). S. K. Mortazavi and K. Kyamakya are with the Institute of Smart Systems Technologies, University of Klagenfurt, Klagenfurt, Austria (e-mails: seyyedkasmor@edu.aau.at; kyandoghere.kyamakya@aau.at). R. Farahani is with the Institute of Information Systems Engineering, Distributed Systems Group, TU Wien, Vienna, Austria (e-mail: r.farahani@dsg.tuwien.ac.at). R. Legenstein is with the Institute of Machine Learning and Neural Computation, Graz University of Technology, Graz, Austria (e-mail: robert.legenstein@tugraz.at). P. Eftekhari and S. K. Mortazavi contributed equally to this work. This research was funded by the Austrian Science Fund (FWF) project 10.55776/COE12. Corresponding author: Reza Farahani (e-mail: r.farahani@dsg.tuwien.ac.at).

Forecasting can expose this loss earlier. Mobile-network studies cover feature selection, spatial context, federated and direct prediction [13]–[16], and predictive information has been connected to policy transfer and proactive O-RAN orchestration [17], [18]. Yet future QoS alone does not determine a command because the appropriate response also depends on resource state and vulnerability. The control problem is therefore to connect future QoS and temporal vulnerability with the learned multi-agent preference and execution-time admissibility. Constraint-aware RL and decentralized wireless control address parts of this problem [19]–[24]. FAROS addresses this execution path while retaining the learned actor. FAROS combines current radio/resource measurements with an eight-step QoS forecast and PCMCI+-supported temporal risk in each cluster-level state. MAPPO learns the action distribution under centralized training and decentralized execution. PCMCI+ is restricted to training data; it supports the retained risk components and their adverse orientation without determining the deployed numerical weights. The final M5 stage applies A5v2 reliability-aware Top-K arbitration. It retains strong actor proposals, filters candidates with frozen admissibility rules, and scores feasible actions using the evaluator’s one-step transition and reward structure together with the frozen candidate-level action costs of Section IV-H. Actor probability and forecast reliability also enter the score; replacement occurs only when a feasible alternative exceeds the fixed intervention margin. Evaluation preserves this separation. Leakage-safe forecasting and later-QoS vulnerability tests precede matched controller rollouts and proposal–execution audits. Sensitivity, runtime, calibration, and explainability characterize the frozen path, while independent ns-3 experiments check the semantics of the six commands outside end-to-end FAROS claims. The main technical contributions are: We formulate FAROS with three cluster-level actors under centralized training and decentralized execution. Each fixed local state combines current network measurements with an eight-step QoS forecast and temporal-risk information, while MAPPO retains decentralized actor proposals during execution. • We integrate training-only PCMCI+ temporal discovery with the controller risk path while keeping it separate from action causality. The reconstructed structure supports the retained risk components and adverse orientation; numerical weights remain fixed implementation parameters. At H = 3, adding temporal risk to the •

2

TABLE I: Positioning of FAROS Against Representative Recent Journal Studies. Work Cai et al. [5] Nagib et al. [17] Nagib et al. [35] Zangooei et al. [36] Staffolani et al. [18] Kavehmadavani et al. [25] Kasuluru et al. [29] Saleh et al. [8] Monaco et al. [27] FAROS

O-RAN / Predictive / Multi-agent Explicit Reliability-aware Independent external / slicing forecast-aware RL constraints arbitration simulation evidence Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

– Yes Yes – Yes Yes Yes – – Yes

forecast raises recall from 0.083592 to 0.325462 and balanced accuracy from 0.540750 to 0.653257. • We develop the final M5 execution stage through the A5v2 reliability-aware Top-K arbiter. It combines symbolic candidate admissibility, one-step model-based utility, actor preference, and forecast reliability. The audit follows all 150,000 agent-level proposal–execution records, separates actor behavior from enforced properties, and finds zero selected-to-executed mismatches. • We evaluate FAROS on 500 matched start indices per training seed under the risky-start protocol. The sameactor M4 shield control reduces the risky-state prohibitedaction rate from 6.67% to zero; M5 has higher mean reward (0.27781 vs. 0.24649) and recovery (90.56% vs. 80.87%) under the same evaluator. A fixed modelmatched diagnostic reference is reported separately; 120 matched ns-3 executions check the semantics of the six actions.

II. BACKGROUND AND R ELATED W ORK Learning-based O-RAN work covers allocation, VNF reconfiguration, slicing, SLA control, traffic steering, offloading, and service placement [2]–[7], [9], [24]–[27]. Forecasting spans spatial, multi-slice, federated, probabilistic, digital-twin, and edge settings [13]–[16], [28]–[31]; predictive policy transfer and proactive orchestration connect forecasts to resource decisions [17], [18]. These studies do not cover the path from early vulnerability to the executed command. Multi-agent wireless control covers distributed resource management, service-function chaining, and RAN slicing [21], [22], [32]–[34]; safety-oriented work uses critics, constrained control, and shielding [19], [35]–[37]. FAROS positions its execution-time intervention within this literature without treating predicate filtering as a safety certificate. PCMCI+ supplies training-only lagged-dependency support [38], while the deployed risk score remains observational. Table I gives the closest functional comparison; entries indicate reported functions rather than equivalent implementations. FAROS links future QoS and temporal risk to a learned multiagent proposal, then applies Top-K retention, admissibility filtering, forecast reliability, and margin-gated arbitration. This structure keeps forecasting, risk, proposals, and execution separately evaluable.

– – – Yes – Yes – – – Yes

– – Yes Yes – – – Yes Yes Yes

– – – – – – – – – Yes

– – – – – – – – – Yes

TABLE II: Six-Action Control Space. Action A0 A1 A2 A3 A4 A5

Control command Keep current configuration Increase resource fraction by 10% Decrease resource fraction by 10% Switch scheduler Switch scheduler and increase resources by 10% Switch scheduler and decrease resources by 10%

III. S YSTEM A RCHITECTURE AND C ONTROL S ETTING Fig. 1 separates the online FAROS path from training-only temporal discovery. It is a logical controller architecture rather than an implemented O-RAN interface stack. A prospective near-real-time RIC xApp can host forecasting, runtime risk, decentralized actors, shielding, and A5v2, while PCMCI+ discovery and critic optimization remain offline. The present implementation does not instantiate E2SM-KPM, E2SM-RC, A1, or O1 signaling or measure RIC/transport latency. A. O-RAN Slicing Environment Let C = {c1 , . . . , cNc } denote the traffic clusters, with Nc = 3 in evaluation. Each cluster executes a decentralized actor over the common six-action space, while the centralized critic uses the concatenated cluster states during training. The environment averages cluster-level rewards into a shared team reward. Its trace-conditioned transition model updates each cluster resource state separately and does not impose an explicit shared physical resource block (PRB) budget or inter-cluster contention constraint. The 10-dimensional local state combines current measurements with an eight-step LSTM forecast from a 64-step QoS history and trainingonly PCMCI+-supported risk. The actor proposal is executed directly, shielded, or reconsidered by M5 /A5v2. Paired evaluations keep the transition model, seeds, and matched starts fixed; ns-3 later checks the same action semantics. PCMCI+ remains outside online action selection. Table II defines the common commands used by every controller variant and by the ns-3 implementation. B. Multi-Agent Control Architecture Under CTDE, the centralized critic receives concatenated cluster states during optimization, while each actor proposes from its local state at execution; agents share the mean cluster

3

Online FAROS Decision Path Forecast Module

O-RAN Slicing Environment

Temporal Risk Module

Local State Assembly

CTDE MAPPO

M5 /A5v2 Execution

Six-Action Interface

Training-Only Temporal Support Training Reservation Trajectories

PCMCI+ Temporal Discovery

Supported Risk Components

Fixed Risk Construction

Fig. 1: FAROS controller architecture with the online decision path and training-only temporal-discovery support path. TABLE III: Main Notation Used in the FAROS Formulation. Symbol

Meaning

Symbol

Meaning

i, t si,t ai,t , at S

Agent index and decision step Local state of agent i at step t Local action and joint action Joint state space

N , N = |N | st A P , R, γdisc

qt c ∆q t ht fθf , θf Eq π θi rt (θi ) τq(M3 ) , τρ(M3 )

qbtfuture ρtcausal qbt+8 zt τ Vϕ bi,t A

aactor i,t zi,t rel ri,t Pi,t (a) Si,t (a) δint

Trace-derived QoS utility at step t Candidate-adjusted future QoS minus current QoS 64-step QoS history used by the forecaster LSTM forecasting function and trained parameters FDR-controlled lagged-link set into QoS Actor policy of agent i PPO probability ratio for agent i Frozen low-QoS and high-risk thresholds of the M3 shield Actor argmax proposal Risky-state indicator used by M5 Forecast-reliability score Actor-preference term used by A5v2 Final A5v2 candidate score Intervention margin

bDL , η UL , preq , pgrant Cres , Csched

DL bitrate, UL SINR, requested PRBs, and granted PRBs Resource-reconfiguration and scheduler-switching costs

Set of agents and number of agents Joint state of all agents Common six-action set {A0 , . . . , A5 } Transition mechanism, reward function, and discount factor Predicted future-QoS term used by the controller Temporal-risk score supplied to the controller Eight-step-ahead QoS forecast target Variable vector used for PCMCI+ temporal discovery Temporal lag used in PCMCI+ discovery Centralized critic with parameters ϕ Normalized advantage estimate Frozen low-QoS and high-risk thresholds of the M5 arbiter Top-K candidate set retained by A5v2 Feasible subset after symbolic filtering One-step model-based utility of candidate action a Forecast-reliability adjustment for candidate a Frozen actor-preference and reliability weights Highest-scoring feasible candidate selected by the arbiter UE count, DL MCS, and UL-buffer occupancy

reward. The evaluator has no explicit shared PRB budget or cross-cluster contention, so no advantage over independent PPO is isolated. In M5 , A5v2 may reconsider the learned proposal before execution. IV. M ODEL AND P ROBLEM F ORMULATION This section formalizes the state, forecast, temporal-risk, MAPPO, and execution components of FAROS. Table III collects the symbols used below.

τq(M5 ) , τρ(M5 ) Ci,t feas Ci,t bi,t (a) R Gi,t (a) wactor , wrel a⋆ i,t nUE , µDL , ubuf L, df , hf , K

Forecast look-back, forecast input dimension, LSTM hidden width, and retained candidate count

These joint symbols collect the three cluster variables used by the CTDE formulation; they do not imply a coupled resource transition or a shared PRB-capacity constraint.

B. State and Action Representation All controller variants use the same normalized 10dimensional local state, UL req grant DL buf future causal si,t = bDL , nUE bi,t , ρi,t i,t , ηi,t , pi,t , pi,t i,t , µi,t , ui,t , qi,t , q



A. Multi-Agent Decision Variables At decision step t, agent i corresponds to traffic cluster ci , receives a local state si,t ∈ R10 , and selects ai,t ∈ A, where A = {A0 , A1 , A2 , A3 , A4 , A5 }. For compact notation, the collections of local states and actions are written as st = (s1,t , s2,t , s3,t ) ,

at = (a1,t , a2,t , a3,t ) .

This is the frozen feature order used by the actor: downlink bitrate, uplink SINR, requested and granted PRBs, UE count, downlink MCS, uplink-buffer occupancy, current QoS, predicted future QoS, and temporal risk. The order is unchanged across paired experiments, and no feature is added during final evaluation. Table II keeps resource changes distinct from scheduler switching, which is necessary for interpreting M5 versus-rule replays when similar mean resource levels arise from different action sequences.

⊤

.

4

TABLE IV: Transition-Model Coefficients. Coefficient

Value

qos_prb_coeff risk_prb_coeff risky_idle_qos_penalty risky_idle_risk_penalty risky_safe_qos_bonus risky_safe_risk_bonus risky_reduce_qos_penalty risky_reduce_risk_penalty

0.08 -0.20 0.08 0.20 0.10 -0.15 0.25 0.60

to predict QoS eight steps ahead, qbt+8 = fθf (ht ),

C. Markov Decision Process Formulation The control process is ⟨N , S, A, P, R, γdisc ⟩, with three agents, the common six-action set, trace-conditioned transition mechanism, reward, and discount factor. For the forecast- and temporal-risk-aware controllers, the controller reward is c t − 0.1 ρtcausal , rt = 1.0 qbtfuture + 0.5 ∆q

(1)

where qbtfuture and ρtcausal denote predicted future QoS and c t = qbtfuture,adj − qt is the temporal risk, respectively, and ∆q candidate-adjusted future QoS relative to the current QoS. Equation (1) favors predicted QoS and improvement while penalizing risk. The versioned implementation fixes the coefficients at 1.0, 0.5, and 0.1; they are protocol parameters rather than optimized weights. Table IV reports the frozen QoS and risk response coefficients of the trace-conditioned evaluator. They are modeling assumptions recovered from the implementation, rather than measured radio parameters; Section VII-F tests their influence. Controller-level gains are conditional on this transition contract. We therefore vary actuator strength while holding the controller fixed, and use ns-3 separately to test the semantics of the six actions outside the trace-conditioned evaluator. The QoS variable qt used by the forecast, temporal-risk, and controller paths is the trace-derived composite utility retained in the controller data. For each constituent xt , let x et = (xt − xmin )/(xmax − xmin + 10−8 ). The resulting utility is e P RB − 0.15R etP RB − 0.05N etU E , (2) qt = 0.45TetDL + 0.35G t where the terms denote normalized downlink throughput, granted PRBs, requested PRBs, and UE load, respectively. Higher throughput and granted resources increase qt , whereas greater requested resources and UE load decrease it. We use this trace-derived utility consistently as the QoS target in the forecasting and controller evaluations. D. Forecasting Formulation The first information path gives the controller a view of near-future QoS before an action is chosen, following recent mobile-network forecasting studies that include recurrent and LSTM-based models [13], [14], [28]. The LSTM uses the 64step history available at decision step t, ht = [qt−63 , qt−62 , . . . , qt ],

where θf denotes the forecasting parameters. The retained controller data establish the 64-step input history, the eightstep target, and the stored qbtfuture values used by FAROS. The original LSTM training source and complete optimization recipe are not preserved, so full-trace statistics from that stored forecast are not interpreted as holdout accuracy. This forecast horizon is separate from the controller-evaluation horizon. Forecast capability is therefore checked independently with a leakage-safe LSTM trained from the retained QoS traces under a frozen chronological protocol. Within each eligible cluster–reservation trace, 64-step inputs and their t + 8 targets are assigned to train, validation, and test partitions without target overlap. Scaling is fitted on training data only, checkpoint selection uses validation MAE only, and the protected test partition is evaluated once after selection. This validation model is used only to test the forecasting capability; it does not replace the stored forecast values consumed by the frozen FAROS experiments. Section VII-A reports MAE, RMSE, R2 , and Pearson correlation against a past-only eight-sample moving average (MovingAverage8) and persistence, and separately tests evaluation-time forecast sensitivity. E. Temporal-Risk Formulation The second information path represents temporal vulnerability from lagged dependencies among the radio and traffic follows the controller implevariables. The notation ρcausal t mentation; throughout the paper it denotes the PCMCI+supported temporal-risk score and is not interpreted as an action-level causal effect. We reconstruct these dependencies with PCMCI+ through Tigramite, following recent journal use of PCMCI+ for time-series dependency analysis [38], and use partial correlation (ParCorr) as the conditional-independence test. The discovery set is  ⊤ req grant DL UL buf , zt = nUE , bt , ηt , µDL t , pt , pt t , ut , qt where the entries correspond to num_ues, sum_requested_prbs, sum_granted_prbs, tx_brate_downlink_mbps, ul_sinr, dl_mcs, ul_buffer_bytes, and qos_score, respectively. The reconstruction uses only the training region and preserves reservation boundaries. Independent reservation trajectories are passed to Tigramite in multiple-series mode, preventing lagged links across reservations. The conditionalindependence test is evaluated for τ ∈ {1, . . . , 8}, with τmin = 1, τmax = 8, and αPC = 0.05. Benjamini– Hochberg false-discovery-rate correction is applied at 0.05, consistent with multiplicity-controlled PCMCI analysis [38]. Temporal-dependency discovery uses the 518,430 training observations (172,810 per cluster) organized as 90 independent reservation trajectories. Validation and final-test observations are excluded.

5

TABLE V: PCMCI+ Temporal-Dependency Discovery Configuration. Property

Configuration

Conditional-independence test Discovery variables Target variable Lag interval PC significance level Multiplicity correction Discovery-fit observations Independent reservation trajectories Cross-reservation lagged links Significant lagged links into QoS Risk components retained by controller Historical component weights

ParCorr UE count, requested PRBs, granted PRBs, DL bitrate, UL SINR, DL MCS, UL buffer, QoS qos_score τ = 1, . . . , 8 0.05 Benjamini–Hochberg FDR, α = 0.05 518,430 90 Disallowed 20 sum_granted_prbs, dl_mcs 0.709609, 0.290391

For a candidate lagged relation zj,t−τ → qt , PCMCI+ returns a partial-correlation coefficient βj,τ and an associated significance value. We denote the retained FDR-controlled set by  Eq = (j, τ ) : pFDR < 0.05 . j,τ

−0.002944 in Q4, with Pearson correlation r = −0.9687. These results are associative; future vulnerability is tested separately in Section VII-B.

The reconstruction identifies 20 significant lagged links into the QoS target. In particular, both variables used by the historical risk index, sum_granted_prbs and dl_mcs, retain significant lagged links to QoS. Granted PRBs are significant at lags {2, 5, 6, 8}, while downlink MCS is significant at lags {2, 3, 4}. Their aggregate signed partial correlations are negative, consistent with the adverse orientation used in the risk construction. The controller does not recompute a graph at every decision. It consumes a scalar temporal-risk index constructed from the two historical PCMCI+-supported components. Let

MAPPO is the learned policy component of FAROS and converts the assembled local state into action preferences under centralized training and decentralized execution, following recent journal formulations of MAPPO and multiagent resource control [21], [36], [39]. Each actor maps 10 → 128 → 128 → 6 with Tanh activations, while the centralized critic maps the concatenated 30-dimensional state through two 256-unit Tanh layers to one value. FAROS adds forecast and temporal-risk inputs to this policy path and governs the actor proposal through shielding or A5v2 arbitration. Since the experiments do not compare MAPPO with independent PPO, we do not attribute controller gains specifically to centralized training. For a policy update, we use the clipped PPO formulation described in [40]. The probability ratio for sample t is

, gt = −pgrant t

dt = −µDL t ,

and standardize each component as xt − µ x x̄t = , x ∈ {g, d}. σx + 10−8 The unscaled risk index is (0)

rt

= 0.709609 ḡt + 0.290391 d¯t .

F. MAPPO Optimization Model

rt (θi ) = (3)

A final standardization gives the temporal-risk score supplied to the controller,

πθi (ai,t | si,t ) . πθiold (ai,t | si,t )

The clipped actor objective is Lclip i (θi ) = h

 i bi,t , clip(rt (θi ), 1 − ϵ, 1 + ϵ)A bi,t , Et min rt (θi )A

(0)

ρcausal = t

rt − µr(0) . σr(0) + 10−8

The weights in (3) are the frozen historical parameters used by the controller. The preserved implementation records this pair directly in the PCMCI+-supported risk construction. We therefore retain the exact operational weights without interpreting them as raw PCMCI+ partial-correlation coefficients. The training-only reconstruction is used to verify the retained components, their lagged links to QoS, and their adverse orientation; Table V records this discovery contract separately from reward and execution logic. For consistency with the retained controller data, the descriptive risk analysis uses the 1,001,456 aligned controller observations. Across numerical quartiles of the temporalrisk score, mean QoS decreases from 0.264621 in Q1 to

bi,t is the normalized advantage estimate. The retained where A MAPPO implementation uses generalized advantage estimation in the standard PPO-based multi-agent optimization role [39], [40], with γdisc = 0.99 and λGAE = 0.95; the PPO clipping parameter is ϵ = 0.20. The critic minimizes  2  V b L (ϕ) = Et Vϕ (st ) − Rt . Actor and critic use Adam at 3 × 10−4 with value-loss coefficient 0.5 and entropy coefficient 0.005. Training uses 512-transition rollouts, five PPO epochs, and 60 iterations, with validation every five iterations on 128 protected starts. For each seed from 0 to 9, checkpoint selection uses the maximum validation score, with the earliest iteration resolving a tie. The ten retained checkpoints have distinct verified hashes

6

TABLE VI: Controller Lineage. Method

Configuration

M0 M1 M2 M3 M4

No-Forecast/No-Risk MAPPO Forecast-Only MAPPO Temporal-Risk-Only MAPPO Shielded Forecast+Temporal-Risk MAPPO Forecast+Temporal-Risk MAPPO; actor prior of M5 M4 actor + M5 risky-state A2 /A5 prohibition only FAROS reliability-aware Top-K arbiter over M4 Independent common-dynamics MAPPO; forecast/risk masked

M4 +Shield M5 Fair MAPPO

and share the same training-configuration and validation-index fingerprints. Protected final-test outcomes are excluded from checkpoint selection and are accessed only after the checkpoints are frozen. Table VI separates the FAROS ablation ladder from the independent Fair MAPPO reference and from the physical actions. M0 is the no-forecast/no-risk ablation within M0 – M5 ; it is not Fair MAPPO. Fair MAPPO is the separately trained Standard_MAPPO_Common_Dynamics_V2 baseline: forecast and risk inputs are masked from actor and critic, the common dynamics and six-action space are retained, and the shield is disabled. Its ten seed-specific policies use 128unit actors and a 256-unit centralized critic and are trained for 30 iterations of 512 rollout steps with Adam 3 × 10−4 , PPO clipping 0.20, entropy coefficient 0.02, value coefficient 0.5, and five PPO epochs. The frozen seed-specific checkpoints are evaluated without retraining on the same 500 matched start indices selected under the risky-start protocol and the same 10-step horizon, yielding 150,000 decisions across ten seeds. A0 –A5 remain the common physical commands. G. Execution Constraints Following the execution-constraint perspective used in safe and constrained RL [19], [20], [35], [36], M3 adds a statedependent shield to the forecast- and temporal-risk-aware policy. In the reconstructed M3 configuration, the risky-state thresholds were estimated from 518,430 training-only observations. The low-QoS boundary corresponds to the 0.35 quantile, τq(M3 ) ≈ 0.079330,

(4)

and the high-risk boundary corresponds to the 0.65 quantile, τρ(M3 ) ≈ 0.198715.

(5)

These reconstructed M3 thresholds remain separate from M5 /A5v2; displayed values are rounded, while evaluation retains machine precision. The final M5 stage follows the frozen A5v2 arbitration configuration used before the protected final evaluation. Its low-QoS and high-risk boundaries are τq(M5 ) ≈ 0.064818,

τρ(M5 ) ≈ 0.442658.

(6)

A5v2 computes these boundaries as the 0.35 predicted-futureQoS and 0.65 temporal-risk quantiles of its reference controller dataframe, whose scope differs from the training-only M3

data. Evaluation retains machine precision; protected final-test outcomes do not estimate or tune either pair. For evaluation, a predicate-defined prohibited action is A2 or A5 while the frozen risky-state predicate is active. A zero executed rate therefore records symbolic enforcement, rather than learned MAPPO safety or formal network safety. Reward, recovery, QoS and risk changes, resource use, and ns-3 assess operational behavior separately. H. Reliability-Aware Arbitration M5 uses the A5v2 arbitration procedure after the M4 Forecast+Temporal-Risk MAPPO actor has produced a probability distribution over the six actions. Let πθi (a | si,t ) denote the actor probability for agent i. The actor proposal is aactor = arg max πθi (a | si,t ), i,t a∈A

and the arbiter retains the K actions with the largest actor probabilities, Ci,t = TopKa∈A πθi (a | si,t ),

K = 3.

The actor argmax is explicitly retained in this candidate set. The risky-state indicator for the final M5 controller uses the frozen A5v2 thresholds in (6), h i future causal zi,t = I qbi,t < τq(M5 ) ∨ ρi,t > τρ(M5 ) . Before candidate scoring, A5v2 applies symbolic admissibility rules. When zi,t = 1, the PRB-reduction actions A2 and A5 are inadmissible. The idle action A0 is also inadmissible when rel the forecast-reliability score ri,t is below 0.42. Reliability is computed from forecast errors that have already matured at decision time. With forecast horizon H = 8 and reliability window W = 64, the implementation shifts the absolute forecast-error sequence by H, averages the available shifted residuals over the last W samples, and uses rel ri,t = clip[0,1] (exp[−40 ēi,t ]) ,

where ēi,t is the rolling mean of these matured absolute residuals. Before any residual has matured, the global median matured error provides the frozen fallback. Thus, the reliability available at step t does not use the unavailable error of the current forecast, and larger values indicate lower historical feas forecast error. These rules define Ci,t ⊆ Ci,t . For each feasible candidate, A5v2 evaluates a one-step outcome with the same trace-conditioned transition and reward equations used by the controller evaluator. A candidate action produces an adjusted future-QoS value qei,t+1 (a) and adjusted temporal-risk value ρei,t+1 (a). The corresponding model-based utility is bi,t (a) = qei,t+1 (a) + 0.50[e R qi,t+1 (a) − qi,t ] − 0.10 ρei,t+1 (a) − Cres (a, zi,t ) − Csched (a). This candidate utility retains the QoS and temporal-risk structure of the controller reward in (1) and additionally subtracts the frozen action costs below. It is used for A5v2 candidate scoring and is distinct from the rollout reward recorded for

7

TABLE VII: Frozen Action Costs in the A5v2 One-Step Candidate Model. Action

PRB change

base Cres

Csched

0 +10% −10% 0 +10% −10%

0.000 0.020 0.010 0.000 0.030 0.010

0.000 0.000 0.000 0.015 0.015 0.015

A0 A1 A2 A3 A4 A5

controller evaluation. The cost terms follow the frozen onestep candidate model. Table VII reports the action-level values used by the implementation. The resource term contains the recorded base cost and, for A1 and A4 , an additional cost of 0.020 when zi,t = 0: base Cres (a, zi,t ) = Cres (a) + 0.020 I[zi,t = 0, a ∈ {A1 , A4 }] .

Scheduler switching contributes Csched (a) = 0.015 for A3 , A4 , and A5 , and zero for the remaining actions. Here, Cres is an action-level resource-reconfiguration cost in the frozen candidate model. Its positive values for the resourcereducing actions A2 and A5 are therefore penalties assigned to taking those commands, rather than signed measures of net PRB consumption. The one-step model uses the frozen risky-state bonuses and penalties reported in Table IV; no transition parameter is refitted during arbitration. Actor preference is represented by
Pi,t (a) = log πθi (a | si,t ) + 10−12 . The forecast-reliability adjustment is  rel ri,t , zi,t = 1, a ∈ {A1 , A3 , A4 },     − 1 − rrel
, z = 1, a = A , i,t 0 i,t Gi,t (a) =  −0.25, z = 0, a ∈ {A  i,t 1 , A4 },    0, otherwise. The 0.020 safe-state term in Cres and the −0.25 safe-state term in Gi,t are separate frozen mechanisms: the former belongs to one-step action cost, while the latter adjusts the candidate score using forecast reliability. Actor preference is the logpolicy term defined above; the 10−12 constant is used only inside the logarithm. The exact candidate score used by A5v2 is bi,t (a) + wactor Pi,t (a) + wrel Gi,t (a), Si,t (a) = R feas a ∈ Ci,t , (7) with frozen values K = 3,

wactor = 0.20,

wrel = 0.15,

and intervention margin

Algorithm 1 M5 (A5v2) Reliability-Aware Top-K Arbitration. 1: Observe si,t and obtain πθi (· | si,t ) 2: Set aactor ← arg maxa πθi (a | si,t ) i,t 3: Form Ci,t ← TopK(πθi , K = 3) future causal 4: Determine zi,t from qbi,t and ρi,t 5: Remove A2 and A5 when zi,t = 1 rel 6: Remove A0 when zi,t = 1 and ri,t < 0.42 feas 7: if Ci,t = ∅ then 8: execute A3 if risky; otherwise execute A0 9: else feas 10: for each a ∈ Ci,t do b 11: compute Ri,t (a) with the frozen one-step model 12: compute Pi,t (a) and Gi,t (a) 13: compute Si,t (a) using (7) 14: end for 15: a⋆i,t ← arg maxa Si,t (a) 16: if aactor is infeasible then i,t 17: execute a⋆i,t 18: else if a⋆i,t ̸= aactor and Si,t (a⋆i,t )−Si,t (aactor i,t i,t ) ≥ 0.015 then 19: execute a⋆i,t 20: else 21: execute aactor i,t 22: end if 23: end if 24: Record the executed action, intervention reason, QoS, risk, recovery, and resource outcomes

three terms as fixed components of the implemented arbitration score, rather than as equally scaled quantities; no weight is retuned from this analysis. Let a⋆i,t = arg max Si,t (a). feas a∈Ci,t

If the actor proposal is infeasible, the arbiter selects a⋆i,t . When the actor proposal is feasible, it is replaced only if a⋆i,t ̸= aactor i,t

and

Si,t (a⋆i,t ) − Si,t (aactor i,t ) ≥ δint .

Otherwise, the actor proposal is retained. If all Top-K candidates are infeasible, the implementation uses A3 as the fallback in a risky state and A0 in a non-risky state. A5v2 remains a fixed execution rule around the learned actor. Algorithm 1 summarizes the fixed selection path; the nominal K = 3, actor weight 0.20, reliability weight 0.15, low-reliability cutoff 0.42, and margin 0.015 are fixed before final evaluation. I. Computational and Time Complexity

δint = 0.015. A nominal score-scale analysis over 440 evaluated candibi,t (a), dates gives mean absolute contributions of 0.1875 for R 0.3150 for wactor Pi,t (a), and 0.0235 for wrel Gi,t (a). The actor component exceeds the reward component in absolute magnitude for 85.0% of these candidates. We therefore interpret the

The computational cost is separated into offline model construction and online decision making because forecastingmodel training, PCMCI+ discovery, and policy optimization are not repeated at each control step. For the online path, N = |N | is the number of agents, L is the forecasting lookback, df is the input dimension at one forecast step, and hf is

8

the LSTM hidden width. One LSTM pass over the forecasting window costs O(Lhf (df + hf )) , with L = 64 in the present controller and a fixed eight-step forecast horizon. Each decentralized MAPPO actor uses the mapping 10 → 128 → 128 → 6, giving
O 10·128 + 1282 + 128·6 per agent and N times this cost for all actors. For A5v2, |A| is the action-space size and K is the retained Top-K count. Bounded Top-K selection followed by candidate filtering and scoring costs O(|A| log K + K) , where |A| = 6 and K = 3. The resulting per-decision online complexity is ! Lhf (df + hf ) + N (10·128 + 1282 + 128·6) . (8) O + |A| log K + K Thus, recurrent forecasting and dense actor inference dominate the evaluated online path, while arbitration operates over a fixed six-action domain. Offline cost depends on the number of optimization updates and conditional-independence tests. Let Bf and Ef denote the number of forecast-training windows processed per epoch and the number of forecast-training epochs. Forecast fitting scales as O(Ef Bf Lhf (df + hf )) up to the corresponding backwardpass constant. For MAPPO, let BRL be the rollout samples processed per optimization epoch and ERL the number of optimization epochs; its optimization cost is O(ERL BRL ) times the actor–critic forward/backward cost. The reported configuration uses BRL = 512 and ERL = 5. For PCMCI+, let NCI be the number of conditional-independence tests performed and CCI the cost of one such test under its conditioning set. The discovery stage is then summarized as O(NCI CCI ). These three terms are offline and do not enter per-decision action selection. The latency microbenchmark in Section VII-I follows the same boundary: it measures the MAPPO actor with the A5v2 arbiter, excluding forecasting, environment execution, and rollout generation. V. FAROS F RAMEWORK AND C ONTROL D ESIGN Forecast and temporal risk enter the state before MAPPO proposes an action. Shielding or A5v2 then governs execution without changing the common action interface. A. Reproducibility-Oriented Design Choices FAROS denotes the complete forecast/risk, MAPPO, and shield/A5v2 controller. The final FAROS configuration is M5 , while A0 –A5 denote the shared physical commands. The reproducibility package preserves the evaluation dataframe, configurations, matched starts, environment source, stored forecast path, and ten hash-verified validation-selected checkpoints. A clean ten-seed evaluation reproduces the reported M5 metrics.

TABLE VIII: Controller Dataset and Evaluation Dimensions. Quantity Controller observations Controller columns Number of agents Unique cluster–reservation series Scheduling configurations cluster_1 observations cluster_2 observations cluster_3 observations Training observations State dimension Action dimension Forecast look-back Forecast horizon Evaluation seeds Matched start indices Matched-risk horizon

Fixed value 1,001,456 50 3 184 2 246,674 421,017 333,765 518,430 10 6 64 steps 8 steps 10 (0–9) 500 10 steps

Because upstream clustering and the original forecast-training lineage are incomplete, Section VII-A validates forecasting with a separate leakage-safe LSTM without replacing the stored controller forecast. VI. E XPERIMENTAL M ETHODOLOGY The evaluation follows the FAROS decision path under protected temporal splitting. Trace-conditioned controller results and independent ns-3 action-level evidence remain separate. Forecast and later-QoS vulnerability tests precede controller comparisons. Variants share paired seeds, matched start indices, common actions, and frozen transition assumptions; sensitivity and external validation are separate. A. Dataset Construction and Temporal Evaluation Setup The controller data come from the public Open RAN Commercial Traffic Twinning Dataset, which contains commercial LTE traffic twinned on an experimental Open RAN platform. The underlying Open RAN experimentation and data-collection setting follows the OpenRAN Gym platform described by Bonati et al. [1]. FAROS uses the retained slicing_1 dataset with three cluster labels that define the MAPPO agents. Scheduler codes 0 and 2 are preserved; because the retained dataset lacks an authoritative code-toname mapping, numerical claims do not assign them to roundrobin or PF. A3 applies the stored scheduler-switch operation. Table VIII summarizes the fixed dataset and evaluation dimensions: the aligned dataframe contains 1,001,456 observations and 50 columns across 184 (cluster,reservation) series, with cluster populations 246,674, 421,017, and 333,765. Upstream cluster construction is not preserved, so reproducibility begins from these retained assignments. The traces represent experimentally twinned demand rather than a live O-RAN deployment. Fig. 2 summarizes the protected temporal split and pairedevaluation protocol. The 64-step context, eight-step forecast, and 10-step rollout are reported in decision steps because no physical sampling period is retained; temporal order and reservation boundaries are preserved. Training uses 172,810

9

Controller Evaluation Dataset 1,001,456 observations, 50 columns

cluster_1 246,674 observations

cluster_2 421,017 observations

cluster_3 333,765 observations

Protected start-index separation Validation Protected Final Test Start overlap = 0 128 starts, indices 172,810–202,239 500 starts, indices 202,290–246,627

Frozen paired-evaluation contract: seeds 0–9; 10-step matched-risk horizon; no finaltest selection; same six-action interface, protected starts, and frozen controller conditions

Fig. 2: Dataset construction and protected temporal evaluation protocol used for the FAROS controller evaluation. TABLE IX: Matched-Risk Evaluation Contract. Property

Value

Seeds Matched start indices Evaluation horizon Agent-level trajectories Recorded decisions M5 /A5v2 low-QoS threshold M5 /A5v2 high-risk threshold A5v2 Top-K Actor weight Reliability weight Low-reliability cutoff Intervention margin

0–9 500 10 steps 15,000 150,000 0.064818 0.442658 3 0.20 0.15 0.42 0.015

observations per agent (518,430 total); validation uses 128 starts at indices 172810–202239 and final evaluation 500 starts at 202290–246627, with no start overlap. Historical context may precede an evaluation start, but validation/test targets and post-start outcomes are excluded from training and checkpoint selection. Seven reset-eligible positions (172801– 172807) remain inside training. B. Matched-Risk Evaluation Contract The frozen evaluation uses seeds 0–9 and the same 500 unique start indices per seed selected under the risky-start protocol, fixed once with matched-start seed 2026; the retained start file and evaluation snapshot are byte-identical. With three agents and a 10-step horizon, each learned-controller evaluation has 15,000 agent trajectories and 150,000 decisions. These conditioned rates do not represent full-trace prevalence. Starts, actions, and transition assumptions remain fixed. Thinning to 490 non-overlapping 10-step starts preserves all eight paired M5 –M4 effect directions. Table IX records the frozen A5v2 parameters; displayed rounding does not alter machineprecision values, and no threshold uses the final test. C. Fixed Model-Matched Deterministic Reference The fixed deterministic reference tests direct threshold logic under the same information and evaluator as M5 . It uses the six-action interface in Table II, with thresholds τq(eval) ≈ 0.064818,

τρ(eval) ≈ 0.442658.

These displayed boundaries are rounded to six decimal places; evaluation retains the same machine-precision thresholds as

Algorithm 2 Fixed Model-Matched Deterministic Reference. 1: Observe qt , qbtfuture , and ρtcausal 2: bq ← (b qtfuture < τq ) 3: bρ ← (ρtcausal > τρ ) 4: bf ← (b qtfuture < qt ) 5: bd ← (qt − qbtfuture ≥ 0.02) 6: if bρ ∧ (bq ∨ bd ) then 7: execute A4 8: else if bq ∨ bρ ∨ bf then 9: execute A1 10: else 11: execute A0 12: end if

final M5 /A5v2, fixed without final-test tuning. The rule has no learned parameters or random seed, so it is evaluated once without an across-seed confidence interval. With current QoS qt , predicted future QoS qbtfuture , and temporal risk ρtcausal , the rule evaluates h i bq = I qbtfuture < τq(eval) , h i bρ = I ρtcausal > τρ(eval) ,   bf = I qbtfuture < qt ,   bd = I qt − qbtfuture ≥ 0.02 . Algorithm 2 applies these conditions in fixed priority and therefore uses only A0 , A1 , and A4 from the common sixaction space. It never issues the resource-reduction commands A2 or A5 ; any decrease-like quantity from another logging view must therefore not be interpreted as a rule-selected reduction action. The rule is a model-matched diagnostic rather than an external or reproduced published baseline. D. Independent ns-3 Validation Protocol The ns-3 study checks the six-action interface rather than replaying FAROS. Twenty matched runs execute A0 –A5 with RngSeed=2026 and matched RngRun, giving 120 successful executions with all ten controller seeds represented in provenance. Balanced action sampling is separate from the natural M5 distribution, and simulator outcomes do not affect training, selection, thresholds, or arbitration. Goodput, PDR,

10

packet loss, and latency test packet-level command effects; the experiment does not reproduce sequential state-conditioned M5 behavior. E. Baselines and Metrics Forecast references are persistence and a past-only eightsample moving average (MovingAverage8). They are controlled references linked to the forecasting literature in Section II, not reproductions of those studies [15], [16], [29]. The independent validation uses seed 42, a one-layer 64-unit LSTM, batch size 1024, learning rate 10−3 , at most 30 epochs, and patience 5. The checkpoint is selected only by validation MAE; the test partition is not used for model selection. The controller comparison fixes 500 protected starts per seed, horizon, actions, and transition–reward equations. M4 executes the Forecast+Temporal-Risk actor directly. M4 +Shield applies only the frozen M5 /A5v2 risky-state A2 /A5 prohibition to the same actors; it does not use the A0 reliability rule, Top-K scoring, or margin-gated arbitration. M5 applies A5v2 to the same-seed actor. Fair MAPPO masks forecast and risk, is trained independently for 30 iterations rather than the FAROS lineage’s 60, and is therefore a common-dynamics reference rather than a training-budgetmatched ablation. Learned controllers use ten seeds with training seed as the inference unit. The deterministic rule is one fixed point, contextualized by proactive and constrained O-RAN control [18], [35] without claiming reproduction. Forecast accuracy uses MAE, RMSE, R2 , and Pearson correlation. Controller statistics are first computed per training seed and then summarized across the ten seeds. For seed s, let (s) (s) (s) Irisk = {n : zn = 1} and Nrisk = |Irisk |. The model-internal action-relative changes are ∆qn = qbnfuture,adj − qbnfuture and − ρpre ∆ρn = ρpost n . The reported risky-state QoS change is n X 1 (s) ∆q risk = (s) ∆qn , Nrisk (s) n∈Irisk

with the analogous risky-state temporal-risk change. For binary indicators pn , hn , and vn denoting proactive execution, prohibited execution, and intervention, P P (s) p (s) h n∈Irisk n n∈Irisk n (s) (s) rpro|risk = , rproh|risk = . (s) (s) Nrisk Nrisk “Proactive” means A1 , A3 , or A4 . Intervention is reported over all decisions and, where stated, conditional on risky states. Recovery is computed per seed over initially risky agent trajectories as the fraction recovered within the fixed ten-step horizon; recovery steps average only recovered trajectories. Reward and realized resource changes use post-transition records, so clipped commands are not credited as realized PRB changes. Proposal, execution, and outcome metrics remain distinct. Across-seed summaries report the sample mean, sample standard deviation, and two-sided 95% confidence interval. Paired M5 –M4 inference uses the training seed as the resampling unit and reports the mean difference, a 95% percentile bootstrap interval over the ten seed pairs, and paired

Fig. 3: Independent leakage-safe H = 8 forecast comparison on 138,212 protected test sequences. The LSTM has the lowest MAE/RMSE and highest R2 /correlation relative to MovingAverage8 and persistence; exact values are shown in the plot.

Cohen’s dz . Paired t and Wilcoxon tests are secondary, with Holm correction across eight controller metrics. For ten nonzero pairs, the minimum two-sided exact Wilcoxon value is 2/210 = 0.001953; interpretation therefore rests on effect estimates and uncertainty intervals. Counts and proposal-toexecution mismatches are checked over the complete 150,000decision trace. VII. R ESULTS Results use the frozen protocol of Section VI: ten checkpoints, 500 matched start indices per seed, a 10-step horizon, and 150,000 decisions per learned controller. The fixed diagnostic has 15,000 decisions. SHA-256 manifests fix the evaluated files; forecast, runtime, and ns-3 tests use 138,212 sequences, 3,000 post-warm-up decisions per method, and 20 matched runs per action, respectively. A. Forecast Accuracy Table X and Fig. 3 make the H = 8 baseline comparison explicit on 138,212 protected test sequences. The validation-selected LSTM has lower MAE than MovingAverage8 (0.014279 vs. 0.014994) and persistence (0.019794), while reaching the highest R2 (0.961723). Across 156 reservation groups, its MAE advantage over MovingAverage8 is 0.000683 with 95% CI [0.000597, 0.000766], and 148 groups favor the LSTM. This supports H = 8 forecasting on the retained traces; it does not validate the historical model that produced the stored FAROS forecast. A paired evaluation-time perturbation compares the fixed nominal package only with a verified shuffled-forecast condition. Shuffling raises mean reward from 0.281455 to 0.319659, while model-internal action-relative QoS change falls from 0.097039 to 0.084670, risk change moves from −0.146634 to −0.129792, and recovery falls from 0.953009 to 0.925131. Because forecast-dependent terms enter (1), the model-based reward is not a standalone operational metric; we interpret it with QoS change, risk change, and recovery. No forecastdisabled M4 checkpoint is available, so this is an evaluationtime sensitivity test rather than a trained ablation.

11

TABLE X: Independent Leakage-Safe Forecast Validation. n

MAE

RMSE

R2

Corr.

138,212 138,212 138,212

0.014279 0.014994 0.019794

0.020560 0.021605 0.027913

0.961723 0.957732 0.929450

0.980721 0.978671 0.964750

Method Independent LSTM MovingAverage8 (past-only) Persistence

TABLE XI: Future-QoS Vulnerability Diagnostic at H = 3. Detector

Precision

Recall

Always positive Forecast only Temporal risk only Forecast + temporal risk

0.537610 0.978940 0.954599 0.952315

1.000000 0.083592 0.312663 0.325462

F1 Balanced acc. 0.699280 0.154031 0.471043 0.485127

0.500000 0.540750 0.647687 0.653257

FPR

FNR

1.000000 0.002091 0.017289 0.018947

0.000000 0.916408 0.687337 0.674538

Fig. 4: Mean QoS across numerical quartiles of the temporalrisk score on the 1,001,456 aligned controller observations. Q1 contains the lowest score values and Q4 the highest; Pearson correlation is r = −0.9687.

Fig. 5: Future-QoS vulnerability comparison at H = 3. Forecast+temporal risk improves recall (0.083592 to 0.325462) and balanced accuracy (0.540750 to 0.653257) over forecast only, with FPR 0.018947.

B. Incremental Vulnerability Information from Temporal Risk

vulnerable; this remains observational.

On the aligned controller population, mean QoS decreases monotonically across the numerical temporal-risk quartiles, from 0.264621 in Q1 to −0.002944 in Q4, with Pearson correlation r = −0.9687 (Fig. 4). This correlation is descriptive and is not treated as independent evidence for the risk construction. The retained risk score uses granted PRBs and downlink MCS, while the trace-derived QoS utility itself contains granted PRBs. Part of the association can therefore arise mechanically from shared state variables. The subsequent future-vulnerability test uses later observed QoS and is reported separately. Table XI and Fig. 5 compare future-vulnerability detectors using subsequently observed QoS. A positive target crosses (M ) τq 5 within the prespecified H = 3 horizon. Among 149,980 targets, prevalence is 53.761%; the always-positive F1 of 0.699280 is non-discriminative because balanced accuracy is 0.5. Relative to forecast only, adding temporal risk raises recall from 0.083592 to 0.325462 and balanced accuracy from 0.540750 to 0.653257, with FPR 0.018947. We therefore emphasize balanced accuracy and error rates rather than F1. The gain supports short-horizon vulnerability discrimination rather than intervention causality. Of 20,671 forecast-only false negatives flagged by temporal risk, 94.3447% later become

C. Matched-Risk Controller Comparison Table XII compares Fair MAPPO, M4 , shield-controlled M4 , M5 , and the fixed model-matched diagnostic under the common matched-risk contract. Fair MAPPO shows substantial seed variability and remains a common-dynamics reference rather than a budget-matched ablation; its model-internal QoS-change quantity is not interpreted as measured network improvement. Applying only the frozen risky-state A2 /A5 prohibition to M4 removes its 6.67% prohibited-action rate. Under the same evaluator, M5 has mean reward 0.277807 versus 0.246486 for M4 +Shield and ten-step recovery 90.56% versus 80.87%. Across ten matched seeds, the paired differences are 0.031322 for reward (95% CI [0.008173, 0.054470]) and 0.096927 for recovery ([0.043174, 0.150680]); recovery steps are essentially unchanged. The shield intervenes on 33,347 of 150,000 M4 decisions (22.23%), separating enforced admissibility from the additional outcome differences associated with A5v2 arbitration. The fixed model-matched diagnostic reaches reward 0.301505 and recovery 97.57%, above FAROS; its evaluatormatched thresholds and action-response equations preclude

12

TABLE XII: Matched-risk controller comparison. Learned controllers report ten-seed mean ± SD; the model-matched rule is one fixed point. Fair MAPPO is the independently trained common-dynamics reference with forecast/risk masked and no shield. Metric

Fixed rule

Mean local reward Prohibited-action rate, risky states Proactive rate, risky states Recovery within ten steps Recovery steps, recovered only

0.301505 0.000000 1.000000 0.975672 3.299213

Fair MAPPO

M4

M4 +Shield

M5 : FAROS

0.037007 ± 0.168395 0.188753 ± 0.076943 0.246486 ± 0.040369 0.277807 ± 0.012672 0.242596 ± 0.196296 0.066729 ± 0.048296 0.000000 ± 0.000000 0.000000 ± 0.000000 0.660099 ± 0.199291 0.822766 ± 0.144611 0.877987 ± 0.123849 0.982330 ± 0.022899 0.547119 ± 0.151451 0.655058 ± 0.135045 0.808707 ± 0.071807 0.905634 ± 0.047492 2.908149 ± 0.411096 2.850026 ± 0.171742 3.205451 ± 0.273738 3.200024 ± 0.076056

TABLE XIII: Actor-proposal and final-execution audit over the frozen M5 trace. The trace contains 30,114 risky decisions in total; prohibited/proactive rate entries are unweighted means of the ten seed-level risky-state rates, consistent with Table XIV, rather than pooled 30,114-decision proportions. The zero executed prohibited rate is predicate-enforced. Metric Actor prohibited rate, risky states Executed prohibited rate, risky states Actor proactive rate, risky states Executed proactive rate, risky states Proposal replacement rate, all decisions Intervention rate, risky states Selected-to-executed mismatch

Value 16.7905% 0.0000% 75.7066% 98.2330% 22.6073% 31.3280% 0.0000%

Audit interpretation Proposal Predicate-enforced execution Proposal Execution 33,911 decisions Arbiter intervention 0 decisions

a FAROS-superiority claim over direct threshold logic. The shield control confirms that zero prohibited actions arise from predicate enforcement. A same-state diagnostic compares A5v2 with the modelmatched diagnostic rule on states encountered along the frozen A5v2 trajectories. Their executed actions differ in 77.73% of paired decisions. This is a decision-level comparison on a common state rather than a trajectory-level disagreement rate, because later states can diverge once different actions are executed. We use the value only to show that A5v2 does not mechanically reproduce the rule, without interpreting disagreement as evidence of better performance. D. Actor-Proposal and Execution Audit Table XIII and Fig. 6 audit each A5v2 actor argmax against its final command, separately from the M4 rollout. Here, “selected” denotes the post-arbitration command, distinct from the actor proposal. The arbiter changes 22.61% of proposals. In risky states, A5v2 execution filtering reduces the prohibitedaction rate from 16.79% at proposal time to 0% at execution, while the proactive-action rate rises from 75.71% to 98.23%. Across all 150,000 executed M5 decisions, 40.27% use a resource-increasing command, 23.26% use a resourcedecreasing command, and 36.46% are resource-neutral. Under the action semantics of Table II, these frequencies correspond to an expected multiplicative resource factor of 1.0170, or an action-implied net change of +1.70%. This is a commandlevel resource proxy. The evaluator does not record physical PRB utilization or over-allocation, so we do not interpret it as measured resource efficiency. E. Across-Seed Execution Stability Table XIV reports proposal-to-execution variability. A5v2 execution filtering blocks prohibited proposals for every seed,

Fig. 6: Actor proposals and executed actions in risky states. A5v2 execution filtering reduces the prohibited-action rate from 16.79% to 0%, while the proactive-action rate increases from 75.71% to 98.23%.

executed proactive actions exceed 94.20%, and no postarbitration selected command differs from execution. Fig. 7 makes the matched-risk comparison visible across four outcomes. Relative to M4 +Shield, M5 increases reward from 0.24649 to 0.27781 and recovery from 80.87% to 90.56%; the fixed model-matched rule remains higher at 0.30151 and 97.57%. F. Local Arbitration-Parameter Sensitivity Table XV tests one-at-a-time sensitivity around the frozen A5v2 setting without retraining; checkpoints, starts, transition– reward model, thresholds, and other parameters remain fixed. Across wactor ∈ {0.15, 0.25} and δint ∈ {0.010, 0.020}, each evaluated over ten seeds, prohibited actions remain zero, reward ranges from 0.275543 to 0.280119, and recovery from

13

TABLE XIV: Across-seed statistics for the actor-to-execution audit. Metric Actor prohibited rate, risky states Executed prohibited rate, risky states Actor proactive rate, risky states Executed proactive rate, risky states Intervention rate, all decisions Intervention rate, risky states

Mean

SD

95% CI

Minimum

Maximum

0.167905 0.000000 0.757066 0.982330 0.226073 0.313280

0.123938 0.000000 0.164647 0.022899 0.077578 0.134644

[0.079245, 0.256565] [0.000000, 0.000000] [0.639285, 0.874847] [0.965949, 0.998710] [0.170578, 0.281569] [0.216961, 0.409599]

0.002766 0.000000 0.484811 0.942007 0.130467 0.104159

0.355324 0.000000 0.961964 1.000000 0.377533 0.518871

Fig. 7: Matched-risk controller comparison under the common frozen contract: (a) reward, (b) model-internal risky-state QoS change, (c) risky-state temporal-risk reduction −∆ρ, and (d) ten-step recovery. M5 improves reward and recovery over M4 and M4 +Shield, while the fixed model-matched rule remains higher. Error bars are 95% CIs for learned controllers; the rule is a point estimate. TABLE XV: Local sensitivity of A5v2 arbitration parameters around the frozen nominal configuration. The nominal row is the protected M5 evaluation; each perturbation varies one parameter at a time over ten seeds without retraining. Setting

Reward

Nominal: wactor = 0.20, δint = 0.015 wactor = 0.15 wactor = 0.25 δint = 0.010 δint = 0.020

0.277807 0.280119 0.275543 0.278371 0.277172

Prohibited, risky Proactive, risky Model QoS ∆, risky Risk ∆, risky Recovery Recovery steps 0.000000 0.000000 0.000000 0.000000 0.000000

0.896927 to 0.910755. These results support local stability around the nominal setting, not global parameter insensitivity.

0.982330 0.982282 0.982116 0.982329 0.982329

0.086186 0.087264 0.084707 0.086304 0.086001

-0.128903 -0.130527 -0.126622 -0.129080 -0.128620

Table XVI records the main integrity checks. Dataset, starts, configuration, source, and seed-specific checkpoint identities are fixed before evaluation. A clean run over all ten frozen checkpoints, without training or parameter selection, reproduces the reported M5 metrics. Seven reset-eligible positions may truncate training episodes, but all remain inside the training region and cannot expose validation or final-test observations. H. Evidence Boundary Table XVII bounds the evidence for the retained trace population under the fixed transition contract. A separate traceconditioned offered-load sensitivity test scales the retained

3.200024 3.206772 3.178952 3.201211 3.200086

TABLE XVI: Integrity Checks for the Fixed Evaluation Protocol. Check

G. Evaluation Integrity and Reproducibility

0.905634 0.910755 0.896927 0.906274 0.904994

Training in evaluation run Seed checkpoints Matched starts Expected decisions per seed Total recorded decisions Total agent-level trajectories Selected/executed mismatches

Verified value No 10 500 unique 15,000 150,000 15,000 0

load representation by 0.25×, 0.50×, 0.75×, 1.00×, and 1.25×, using ten seeds, 256 matched starts, and a ten-step horizon. The observed stability supports only this tested load range within the trace-conditioned model. It does not establish robustness to larger distribution shifts, forecast corruption, severe temporal-risk shift, or growth in the number of agents, slices, or actions.

14

TABLE XVII: Scope of the Reported Evidence. Evidence source

Supported statement

Boundary

Independent leakage- H = 8 forecast accuracy on protected test safe forecast test sequences Matched-risk A5v2 trace Ten-seed outcome and execution audit Offered-load sensitivity Stable behavior over 0.25×–1.25× tested load Symbolic predicate audit No prohibited final actions in risky states Balanced matched ns-3 Packet-level direction of six commands action runs

Fig. 8: Controller-side actor/arbitration runtime for RuntimeCfg2, Runtime-Cfg3, and A5v2. Runtime-Cfg2 and RuntimeCfg3 are configuration names, unrelated to action symbols A2 and A3 . Bars report mean, P95, and P99 latency over 3,000 decisions after 200 warm-up decisions. I. Runtime, Reliability, and Explainability

Validation model is separate from the stored FAROS forecast path Trace-conditioned transition and reward model Trace-conditioned load robustness, not computational scaling Predicate compliance, not formal safety Action-level evidence, not sequential FAROS validation

paired executions with all ten controller seeds represented in provenance. Simulator outcomes do not affect training, selection, thresholds, or arbitration, and comparator policies are not replayed. The evidence therefore concerns packet-level command response under the tested configuration rather than end-to-end FAROS superiority. Table XVIII reports coherent action responses. Relative to matched A0 , resource-increasing A1 and A4 have higher mean goodput and PDR and lower mean delay; the reported 95% intervals for their goodput and mean-delay differences exclude zero. Resource-reducing A2 and A5 move packet-level outcomes in the opposite direction, lowering goodput and increasing delay; A5 mean-delay differences are unfavorable in all 20 matched runs. Scheduler-only A3 has smaller mixed effects. Its goodput difference is −0.01277 Mbps with 95% CI [−0.02453, −0.00101]; the paired t-test gives p = 0.03479, unchanged after Holm adjustment, while Wilcoxon gives p = 0.01531 unadjusted and p = 0.02189 after Holm adjustment. Mean/P95 delay intervals cross zero, and there is no Holmsupported P99 improvement. Fig. 9 expresses the matched latency effects on a common relative scale,

Fig. 8 reports controller-side actor/arbitration runtime for two retained pre-A5v2 configurations, denoted RuntimeCfg2 and Runtime-Cfg3, and for A5v2. After 200 warm-up decisions, 3,000 decisions per method give mean/P95/P99 latency of 0.985050/0.995561/1.056689 ms, 1.074527/1.091840/1.128021 ms, and 1.753396/1.777802/1.807549 ms, respectively. The benchmark excludes E2 transport, telemetry, RIC scheduling, and actuator delay and is not end-to-end O-RAN latency. The 30-row reliability audit relates rrel to forecast error across six conditions and five bins. In 1,500 balanced actor explanations, mean absolute SHAP attribution is highest for temporal risk (0.186897), predicted future QoS (0.146218), downlink MCS (0.097787), and granted PRBs (0.075653), following actor-level explainable-RL practice [41]. Finalcommand interpretation uses a separate 4,500-row candidate audit with 263 interventions. Parameter randomization retains rank correlations 0.804878, 0.721212, and 0.830303 across clusters, so SHAP is treated only as descriptive actor attribution.

where m denotes mean, P95, or P99 delay. This normalization leaves the paired samples, confidence intervals, and statistical tests unchanged. Resource-increasing actions reduce delay, resource-reducing actions increase it, and A3 remains comparatively small and unresolved for mean/P95 delay. A natural-frequency consistency check weights the actionlevel responses by the frozen M5 action frequencies. The goodput difference is +0.01012 Mbps with 95% CI [−0.00127, +0.02151], with Holm-adjusted paired-t/Wilcoxon p = 0.464/0.715. Median delay is the narrow exception at about −0.101 ms with 95% CI [−0.1365, −0.0655]. The broader mixture evidence does not establish packet-level superiority; ns-3 supports only the operational meaning of the commands under matched randomness.

VIII. E XTERNAL NS -3 ACTION -L EVEL VALIDATION

IX. D ISCUSSION

The independent ns-3 study tests the six-action interface outside the trace-conditioned evaluator. Twenty matched RngRun values execute all commands, giving 120 successful

Assumptions and applicability. FAROS is evaluated under the frozen trace-conditioned transition–reward contract used for state evolution and A5v2 scoring. The public dataset

∆rel (Ai , m) = 100

m(Ai ) − m(A0 ) m(A0 )

,

15

TABLE XVIII: Independent packet-level ns-3 check of the six frozen commands. Values are means over 20 matched RngRun values per action; final columns give paired 95% confidence intervals relative to A0 . Action

Command

A0 A1 A2 A3 A4 A5

Keep configuration Increase PRBs 10% Decrease PRBs 10% Switch scheduler Switch scheduler + increase PRBs Switch scheduler + decrease PRBs

Goodput (Mbps)

PDR

Mean delay (ms)

∆ Goodput vs. A0 [95% CI]

∆ Mean delay vs. A0 [95% CI] (ms)

5.6113 5.6697 5.5810 5.5985 5.6538 5.5730

0.9300 0.9377 0.9284 0.9272 0.9348 0.9266

86.41 59.46 119.09 84.06 55.47 127.63

– +0.0584 [+0.0261, +0.0907] −0.0303 [−0.0528, −0.0078] −0.0128 [−0.0245, −0.0010] +0.0425 [+0.0117, +0.0734] −0.0382 [−0.0617, −0.0147]

– −26.95 [−43.02, −10.87] +32.68 [+16.59, +48.76] −2.35 [−6.59, +1.89] −30.95 [−51.13, −10.76] +41.22 [+17.75, +64.68]

full-trace forecast statistics are not treated as holdout evidence, and the separate leakage-safe experiment provides the disjointtarget forecast test. X. C ONCLUSION AND F UTURE W ORK

Fig. 9: Normalized ns-3 latency response for A1 –A5 relative to matched A0 ; markers show mean, P95, and P99 latency changes, and error bars are 95% confidence intervals over the same 20 matched runs.

twins commercial LTE traffic on an experimental Open RAN platform rather than a live deployment, and offered-load tests cover 0.25×–1.25×. The independent ns-3 study checks the packet-level semantics of the six commands and does not evaluate closed-loop sequential M5 behavior. Fig. 1 therefore represents a logical deployment architecture. E2SM-KPM, E2SM-RC, A1/O1 signaling, and end-to-end RIC/transport latency are outside the implemented scope. Threats to validity. The zero executed prohibited-action rate follows from predicate enforcement and does not constitute evidence of learned or formally verified safety. Because the evaluator has no explicit shared PRB budget or inter-cluster contention, the experiments do not isolate a CTDE advantage over independent PPO. The model-matched deterministic rule is a diagnostic reference, while Fair MAPPO uses a different training budget. PCMCI+ provides training-only lagged conditional-dependence support for the temporal-risk path. It does not establish action-level causal effects or determine the historical risk weights. SHAP explains the pre-arbitration actor, whereas final executed commands are covered by the proposal–selection–execution audit. Reproducibility is supported by the fixed evaluation dataset, MAPPO configurations, matched starts, ten checkpoint hashes, the 150,000-decision trace, analysis outputs, and the ns-3 package. A clean ten-seed evaluation reproduces the reported M5 metrics. Seven late training starts cannot reach validation or final-test states. Incomplete upstream clustering provenance and the unavailable original LSTM training lineage limit reconstruction of those historical stages; accordingly, historical

FAROS connects QoS forecasting and temporal risk to MAPPO under CTDE and governs execution through symbolic admissibility and reliability-aware Top-K arbitration. The independent H = 8 test reaches MAE 0.014279 and R2 = 0.961723. At H = 3, temporal risk raises vulnerability recall from 0.083592 to 0.325462 and balanced accuracy from 0.540750 to 0.653257. The same-actor shield removes the 6.67% risky-state prohibited-action rate, while M5 reaches reward 0.27781 and recovery 90.56%. The model-matched rule remains higher at 0.30151 and 97.57%, so no superiority over direct threshold logic is claimed. A5v2 changes 22.61% of proposals, while the zero executed prohibited-action rate is enforced by the predicates. The 120 ns-3 executions provide action-level evidence only. Future work should use state-synchronized closed-loop O-RAN/ns-3 evaluation and recalibrate temporal risk under substantial distribution shift. R EFERENCES [1] L. Bonati, M. Polese, S. D’Oro, S. Basagni, and T. Melodia, “OpenRAN Gym: AI/ML Development, Data Collection, and Testing for O-RAN on PAWR Platforms,” Computer Networks, 2023, DOI: 10.1016/j.comnet.2022.109502. [2] E. Amiri, N. Wang, M. Shojafar, M. Q. Hamdan, C. H. Foh, and R. Tafazolli, “Deep Reinforcement Learning for Robust VNF Reconfigurations in O-RAN,” IEEE Transactions on Network and Service Management, 2024, DOI: 10.1109/TNSM.2023.3316074. [3] A. Filali, Z. Mlika, and S. Cherkaoui, “Open RAN Slicing for MVNOs With Deep Reinforcement Learning,” IEEE Internet of Things Journal, 2024, DOI: 10.1109/JIOT.2024.3365665. [4] J. Dai, L. Li, R. Safavinejad, S. Mahboob, H. Chen, V. V. Ratnam, H. Wang, J. Zhang, and L. Liu, “O-RAN-Enabled Intelligent Network Slicing to Meet Service-Level Agreement (SLA),” IEEE Transactions on Mobile Computing, 2025, DOI: 10.1109/TMC.2024.3476338. [5] Y. Cai, P. Cheng, Z. Chen, M. Ding, B. Vucetic, and Y. Li, “Deep Reinforcement Learning for Online Resource Allocation in Network Slicing,” IEEE Transactions on Mobile Computing, 2024, DOI: 10.1109/TMC.2023.3328950. [6] S. Mhatre, F. Adelantado, K. Ramantas, and C. Verikoukis, “Intelligent QoS-Aware Slice Resource Allocation With User Association Parameterization for Beyond 5G O-RAN-Based Architecture Using DRL,” IEEE Transactions on Vehicular Technology, 2025, DOI: 10.1109/TVT.2024.3483288. [7] W. Liu, M. A. Hossain, N. Ansari, A. Kiani, and T. Saboorian, “Reinforcement Learning-Based Network Slicing Scheme for Optimized UEQoS in Future Networks,” IEEE Transactions on Network and Service Management, 2024, DOI: 10.1109/TNSM.2024.3368294. [8] F. Saleh, A. O. Fapojuwo, and D. Krishnamurthy, “eSlice: Elastic Inter-Slice Resource Allocation for Smart City Applications,” IEEE Transactions on Network and Service Management, 2025, DOI: 10.1109/TNSM.2025.3604352.

16

[9] M. Sulaiman, A. Moayyedi, M. Ahmadi, M. A. Salahuddin, R. Boutaba, and A. Saleh, “Coordinated Slicing and Admission Control Using MultiAgent Deep Reinforcement Learning,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2022.3222589. [10] Y. Chiang, C.-H. Hsu, G.-H. Chen, and H.-Y. Wei, “Deep Q-LearningBased Dynamic Network Slicing and Task Offloading in Edge Network,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2022.3208776. [11] A. Gharehgoli, A. Nouruzi, N. Mokari, P. Azmi, M. R. Javan, and E. A. Jorswieck, “AI-Based Resource Allocation in End-toEnd Network Slicing Under Demand and CSI Uncertainties,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3243837. [12] W. Fan, X. Li, B. Tang, Y. Su, and Y. Liu, “MEC Network Slicing: Stackelberg-Game-Based Slice Pricing and Resource Allocation With QoS Guarantee,” IEEE Transactions on Network and Service Management, 2024, DOI: 10.1109/TNSM.2024.3409277. [13] J. S. Mendoza, I. de la Bandera, D. Palacios, and R. Barco, “Forecasting Framework for Mobile Networks Based on Automatic Feature Selection,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3249470. [14] J. Zhou, D. Ding, Z. Wu, and Y. Xiu, “Spatial Context-Aware TimeSeries Forecasting for QoS Prediction,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3250512. [15] H. P. Phyu, R. Stanica, and D. Naboulsi, “Multi-Slice Privacy-Aware Traffic Forecasting at RAN Level: A Scalable Federated-Learning Approach,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3267725. [16] V. Perifanis, N. Pavlidis, R.-A. Koutsiamanis, and P. S. Efraimidis, “Federated Learning for 5G Base Station Traffic Forecasting,” Computer Networks, 2023, DOI: 10.1016/j.comnet.2023.109950. [17] A. M. Nagib, H. Abou-Zeid, and H. S. Hassanein, “Accelerating Reinforcement Learning via Predictive Policy Transfer in 6G RAN Slicing,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3258692. [18] A. Staffolani, V.-A. Darvariu, L. Foschini, M. Girolami, P. Bellavista, and M. Musolesi, “PRORL: Proactive Resource Orchestrator for Open RANs Using Deep Reinforcement Learning,” IEEE Transactions on Network and Service Management, 2024, DOI: 10.1109/TNSM.2024.3373606. [19] Q. Yang, T. D. Simao, S. H. Tindemans, and M. T. J. Spaan, “SafetyConstrained Reinforcement Learning with a Distributional Safety Critic,” Machine Learning, 2023, DOI: 10.1007/s10994-022-06187-8. [20] X. Hou, M. Gan, W. Wu, C. Wang, Y. Ji, and S. Zhao, “Merging Planning in Dense Traffic Scenarios Using Interactive Safe Reinforcement Learning,” Knowledge-Based Systems, 2024, DOI: 10.1016/j.knosys.2024.111548. [21] X. Du, T. Wang, Q. Feng, C. Ye, T. Tao, L. Wang, Y. Shi, and M. Chen, “Multi-Agent Reinforcement Learning for Dynamic Resource Management in 6G in-X Subnetworks,” IEEE Transactions on Wireless Communications, 2023, DOI: 10.1109/TWC.2022.3207918. [22] Z. Ji, Z. Qin, and X. Tao, “Meta Federated Reinforcement Learning for Distributed Resource Allocation,” IEEE Transactions on Wireless Communications, 2024, DOI: 10.1109/TWC.2023.3345363. [23] F. W. Murti, S. Ali, G. Iosifidis, and M. Latva-aho, “Deep Reinforcement Learning for Orchestrating Cost-Aware Reconfigurations of vRANs,” IEEE Transactions on Network and Service Management, 2024, DOI: 10.1109/TNSM.2023.3292713. [24] M. A. O. Alsenwi, E. Lagunas, and S. Chatzinotas, “Distributed Learning Framework for eMBB-URLLC Multiplexing in Open Radio Access Networks,” IEEE Transactions on Network and Service Management, 2024, DOI: 10.1109/TNSM.2024.3451295. [25] F. Kavehmadavani, V.-D. Nguyen, T. X. Vu, and S. Chatzinotas, “Empowering Traffic Steering in 6G Open RAN With Deep Reinforcement Learning,” IEEE Transactions on Wireless Communications, 2024, DOI: 10.1109/TWC.2024.3396273. [26] A. Ndikumana, K. K. Nguyen, and M. Cheriet, “Federated Learning Assisted Deep Q-Learning for Joint Task Offloading and Fronthaul Segment Routing in Open RAN,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3245544. [27] D. Monaco, A. Sacco, and G. Marchetto, “CROWN: Cross-Attention Reinforcement Learning for O-RAN Wireless Networks,” Computer Networks, 2026, DOI: 10.1016/j.comnet.2026.112276. [28] M. Di Mauro, G. Galatro, F. Postiglione, W. Song, and A. Liotta, “Multivariate Time Series Characterization and Forecasting of VoIP Traffic in Real Mobile Networks,” IEEE Transactions on Network and Service Management, 2023, DOI: 10.1109/TNSM.2023.3295748.

[29] V. Kasuluru, L. Blanco, and E. Zeydan, “Enhancing Open RAN Operations: The Role of Probabilistic Forecasting in Network Analysis,” IEEE Transactions on Network and Service Management, 2025, DOI: 10.1109/TNSM.2025.3565268. [30] H. Navidan, C. Martin, V. Maglogiannis, D. Naudts, M. Diaz, I. Moerman, and A. Shahid, “An End-to-End Digital Twin Framework for Dynamic Traffic Analytics in O-RAN,” IEEE Transactions on Network and Service Management, 2026, DOI: 10.1109/TNSM.2025.3628756. [31] A. Tundo, F. Filippini, F. Regonesi, M. Ciavotta, and M. Savi, “Decentralized Edge Workload Forecasting With Gossip Learning,” IEEE Transactions on Network and Service Management, 2025, DOI: 10.1109/TNSM.2025.3570450. [32] Y. Ai, H. Li, H. Ruan, H. Liu, and X. Wang, “Achieving Topology Generalization Without Retraining in Service Function Chaining Through Cooperative Multi-Agent Reinforcement Learning,” Computer Networks, 2025, DOI: 10.1016/j.comnet.2025.111211. [33] Y. Chai, Y. Zhang, Z. Zhang, D. Guo, and Y. Teng, “Research on Joint Game Theory and Multi-Agent Reinforcement Learning-Based Resource Allocation in Micro Operator Networks,” Computer Networks, 2025, DOI: 10.1016/j.comnet.2025.111333. [34] D. J. Birabwa, D. Ramotsoela, and N. Ventura, “Multi-Agent Deep Reinforcement Learning for User Association and Resource Allocation in Integrated Terrestrial and Non-Terrestrial Networks,” Computer Networks, 2023, DOI: 10.1016/j.comnet.2023.109827. [35] A. M. Nagib, H. Abou-Zeid, and H. S. Hassanein, “Safe and Accelerated Deep Reinforcement Learning-Based O-RAN Slicing: A Hybrid Transfer Learning Approach,” IEEE Journal on Selected Areas in Communications, 2024, DOI: 10.1109/JSAC.2023.3336191. [36] M. Zangooei, M. Golkarifard, M. Rouili, and R. Boutaba, “Flexible RAN Slicing in Open RAN With Constrained Multi-Agent Reinforcement Learning,” IEEE Journal on Selected Areas in Communications, 2024, DOI: 10.1109/JSAC.2023.3336156. [37] K.-C. Hsu, A. Z. Ren, D. P. Nguyen, A. Majumdar, and J. F. Fisac, “Sim-to-Lab-to-Real: Safe Reinforcement Learning with Shielding and Generalization Guarantees,” Artificial Intelligence, vol. 314, p. 103811, 2023, DOI: 10.1016/j.artint.2022.103811. [38] Y. Tian, D. Giaquinto, G. D. Capua, J. N. Claassen, J. Ali, H. Li, and C. D. Michele, “Historical Changes in the Causal Effect Networks of Compound Hot and Dry Extremes in Central Europe,” Communications Earth & Environment, vol. 5, p. 764, 2024, DOI: 10.1038/s43247-02401934-2. [39] S. S. Hassan, Y. M. Park, Y. K. Tun, W. Saad, Z. Han, and C. S. Hong, “Satellite-Based ITS Data Offloading & Computation in 6G Networks: A Cooperative Multi-Agent Proximal Policy Optimization DRL With Attention Approach,” IEEE Transactions on Mobile Computing, vol. 23, no. 5, pp. 4956–4974, 2024, DOI: 10.1109/TMC.2023.3300314. [40] M. Zheng, J. Zhang, C. Zhan, X. Ren, and S. Lü, “Proximal Policy Optimization with Reward-Based Prioritization,” Expert Systems with Applications, vol. 283, p. 127659, 2025, DOI: 10.1016/j.eswa.2025.127659. [41] X. Xu, F. Wu, M. Bilal, X. Xia, W. Dou, L. Yao, and W. Zhong, “XRL-SHAP-Cache: An Explainable Reinforcement Learning Approach for Intelligent Edge Service Caching in Content Delivery Networks,” Science China Information Sciences, vol. 67, p. 170303, 2024, DOI: 10.1007/s11432-023-3987-y.

Related documents

• A Patient World Model for Early Forecasting of Digital Health Campaign Outcomes: Capabilities and Limits#1049076
• Long-Horizon RL Needs Muon#1355185
• Long-Horizon RL Needs Muon#1336045
• RL Starts before RL: On Policy Distillation for Better Reinforcement Learning#1064689
• Multipath Extension for the Optimized Link State Routing Protocol Version 2 (OLSRv2)#399333