XAI-Guided Conservative Decentralized Execution for Offline Multi-Agent Network Slicing

Abstract: The recent advances toward sixth-generation (6G) and beyond-6G networks have accelerated the need for intelligent resource management mechanisms capable of supporting heterogeneous services under shared infrastructures in network slicing. However, resource allocation in network slicing naturally forms a resource-coupled cooperative optimization problem with competing slice demands. Slices compete for limited resources to minimize individual latencies while coordinating to avoid conflicts and underutilization. Although multi-agent reinforcement learning (MARL) has shown promising performance in such settings, existing online formulations remain costly, unsafe, and difficult to deploy due to their reliance on environmental interactions and communication among agents. In this work, we present explainable artificial intelligence (XAI)-guided conservative decentralized execution (X-CODE). X-CODE is an explainable offline MARL that operates offline without environmental interaction, nor inter-agent communication. It exploits explainability-aware reward shaping to modify the relative preference among joint offline transitions during centralized training to improve decentralized resource-allocation behavior. In deployment, the agents operate independently without signaling exchange among the agents. Simulation results demonstrate that the proposed approach achieves zero observed resource-conflict events in the evaluated test episodes while minimizing per-slice latencies. Moreover, the proposed framework exhibits lower signaling overhead and reduces effective inference latency by 88 % under the considered communication-delay model compared to the online baselines. Source codes and datasets are available through: https://github.com/Eslam211/xcode-ran-slicing.

• The paper introduces X-CODE, combining conservative Q-learning, VDN, and cross-fitted SHAP reward relabeling to guide decentralized CPU allocation without execution-time communication.
• X-CODE achieved zero observed resource conflicts, matched the median latency of a communication-assisted online baseline, reduced modeled inference latency by 88%, and used zero signaling bits per step.
• Ablation results show that moderate conservatism and attribution-based shaping are both essential, while excessive penalties or reward bonuses cause OOD overestimation, underutilization, or degraded latency.

Problem setting and motivation

This paper addresses CPU resource allocation in edge-enabled RAN slicing, formulated as a resource-coupled cooperative multi-agent optimization problem with competing slice demands. Three slices—eMBB, mMTC, and URLLC—share a common edge server with a nominal budget of 40 Gcycles/s, and each slice agent selects a discrete CPU allocation from $\mathcal{A}=\{1,5,10,15,20,25,30\}$ Gcycles/s to minimize its own end-to-end latency while jointly avoiding resource conflicts, defined as joint allocations exceeding the shared budget. The tension is structural: latency minimization pushes each agent toward aggressive allocation, while conflict avoidance rewards conservatism that can degenerate into underutilization.

The authors position the work against two deployment constraints of existing MARL. Online CTDE methods require continuous environmental interaction and inter-agent signaling during training and execution, incurring cost, delay, and safety risks from early exploration. Offline MARL removes the interaction requirement but suffers from distributional shift: with narrow-coverage datasets (here, deliberately a random-walk behavioral policy), TD maximization over out-of-distribution (OOD) actions yields over-optimistic value estimates, and the absence of execution-time communication increases the risk of converging to sub-optimal or saddle-point coordination outcomes. The paper's stated gap is that explanation-guided offline MARL remains unexplored, particularly for high-stakes 6G slicing scenarios.

System model

The environment is an episodic POMDP over $T=40$ decision intervals of duration $\tau = 0.01$ s. Each slice maintains coupled computation and transmission queues; the computation queue evolves under the applied CPU allocation via service rate $S_{k,t} = \tau U a_{k,t}$ with computation efficiency $U = 5\times 10^{-3}$ bits/cycle, and per-slice average latency is computed via Little's law from accumulated queue lengths. A key modeling choice is that when a conflict occurs ($\sum_k a_{k,t} > f_{\max}$), the simulator does not apply the infeasible request; instead it falls back to a default feasible allocation vector $\boldsymbol{f}^{\mathrm{th}}=[15,15,10]$ Gcycles/s while recording and penalizing the conflict event through the indicator $\chi_t$. The objective P1 minimizes weighted cumulative latency plus conflict penalties subject to discrete action feasibility. Each agent observes only its local traffic load $\Phi_{k,t}$, previous applied-allocation gap $\beta_{k,t}$, and its two queue lengths; the reward combines a latency penalty ($T=40$0), a small allocation incentive ($T=40$1), and a strong conflict penalty ($T=40$2).

The X-CODE method

X-CODE is built on CQL within a VDN-based CTDE architecture. The centralized value function is decomposed additively as $T=40$3, trained on a team reward (the mean of per-agent rewards) with a Bellman term plus a per-agent CQL regularizer penalizing high values on actions unsupported by the dataset. At deployment, each agent extracts its policy by argmax over its local Q-function alone, requiring no inter-agent communication.

The central algorithmic contribution is XAI-guided reward relabeling. The procedure uses cross-fitting: the offline dataset is split into $T=40$4 disjoint folds, a preliminary CQL-VDN critic is trained on $T=40$5 folds, and SHAP attributions are computed only on held-out transitions to reduce bias from explaining samples used to fit the critic. Absolute SHAP values are mapped through a softmax into a feature-attribution distribution per agent, whose normalized entropy $T=40$6 yields an attribution-concentration score aggregated across agents:

$T=40$7

The team reward is then relabeled as $T=40$8, where centering preserves the empirical reward scale while re-ranking joint transitions toward those whose critic outputs are driven by concentrated, interpretable feature subsets. The intuition is that transitions with concentrated attributions provide more reliable guidance signals for decentralized policy extraction under limited data coverage.

The authors are explicit about an important theoretical caveat: this shaping is not potential-based and therefore does not preserve the optimal policy of the original environment reward. Centering the bonus preserves its empirical mean but not policy invariance, so $T=40$9 governs a genuine trade-off between the original latency-conflict objective and attribution-based preference modification. This is an honest limitation rather than a hidden assumption, and the ablation confirms that large $\tau = 0.01$0 degrades performance by letting the bonus dominate the environment reward.

Experimental results

The evaluation uses a random-walk behavioral dataset of 10,000 transitions with near-uniform action coverage—a deliberately low-quality dataset lacking any coordination structure—to stress-test offline learning. Baselines include online DQN without communication, online DQN with emergent communication codes and an information bottleneck (DQN Comm + IB), XAI-guided independent CQL (X-CQL), CFCQL, QMIX, and plain VDN.

The headline claims are strong and worth stating precisely:

Result	Value	Comparison
Resource-conflict events (test episodes)	Zero observed	Only method achieving this
Effective inference latency reduction	88%	vs. DQN (Comm + IB) under considered delay model
Execution signaling overhead	100% reduction (zero bits/step)	vs. communication-based baselines

On latency, X-CODE matches the median latency of the online DQN (Comm + IB)—a notable result given that the offline method never interacts with the environment—while both online baselines exhibit nontrivial conflict rates. Against offline MARL baselines, the contrast is sharper: QMIX records the worst latency CDF, and CFCQL, QMIX, and X-CQL all converge to conservative policies that underutilize the shared CPU to avoid conflicts. X-CODE progressively increases CPU utilization during training toward the high-utilization region while maintaining zero observed conflicts, indicating that the relabeling mitigates the pessimism-induced underutilization that direct CQL adaptations exhibit in this resource-coupled setting.

The ablations support the mechanism. With $\tau = 0.01$1 (no conservatism), OOD overestimation produces worse latency tails; $\tau = 0.01$2 is optimal; $\tau = 0.01$3 over-constrains the policy to the dataset. With $\tau = 0.01$4 (no relabeling, reducing to VDN), the policy becomes conservative and latencies rise; $\tau = 0.01$5 achieves the best latency distribution; large $\tau = 0.01$6 lets the bonus dominate and degrade performance. These sweeps confirm that both components—the conservative penalty and the XAI bonus—are necessary and that neither should be pushed to extremes.

The SHAP analysis provides interpretability beyond performance reporting. CPU-gap features $\tau = 0.01$7 and traffic loads $\tau = 0.01$8 dominate the centralized value estimates, while queue features contribute small corrections. Notably, the dominant features differ per agent—$\tau = 0.01$9 and $S_{k,t} = \tau U a_{k,t}$0 dominate agents 1 and 3, whereas $S_{k,t} = \tau U a_{k,t}$1 dominates agent 2—indicating that X-CODE learns slice-dependent valuation patterns rather than a uniform attribution profile.

Limitations and open questions

Several limitations bound the strength of these results. First, the zero-conflict claim is empirical and scoped to "the evaluated test episodes" under the simulated traffic profiles; it is not a guarantee, and the fallback mechanism means conflicts are penalized rather than physically prevented. Second, the evaluation scale is modest: three agents, seven discrete actions, a single edge server, and simulated traffic. Scalability of SHAP-based cross-fitted relabeling to larger agent counts and continuous action spaces is untested, and the computational cost of the multi-fold preliminary critic training plus SHAP computation is acknowledged but not quantified against baselines. Third, because the shaping is not policy-invariant, there is no theoretical characterization of how far the optimized policy can deviate from the optimum of P1; the choice of $S_{k,t} = \tau U a_{k,t}$2 is justified empirically only. Fourth, the comparison to DQN (Comm + IB) at matched median latency relies on a specific communication-delay model, so the 88% inference-time reduction is model-dependent. Finally, the paper leaves open whether the learned policies can be fine-tuned online after deployment (offline-to-online transfer) and whether the approach holds in larger-scale slicing scenarios—both explicitly identified as future work.

Conclusion

This paper contributes a concrete instantiation of explanation-guided offline MARL for network slicing: cross-fitted SHAP-based reward relabeling layered on CQL-VDN, yielding decentralized policies that require no inter-agent signaling at execution time. Empirically, it demonstrates that moderate attribution-based guidance counteracts the conservative-underutilization failure mode of offline CQL methods in resource-coupled settings, achieving zero observed conflicts with latency matching a communication-assisted online baseline, at the cost of additional offline training complexity and a non-policy-invariant shaping objective whose trade-off is controlled only empirically.