Demystifying Deep Reinforcement Learning: A Neuro-Symbolic Framework for Interpretable Open RAN Automation

Abstract: Open Radio Access Networks (O-RAN) are increasingly adopting data-driven control through Deep Reinforcement Learning (DRL) to optimize complex tasks such as network slicing and mobility management. However, the deployment of DRL in carrier-grade networks is hindered by its inherent opacity and stochastic execution, which limit operator trust, auditability, and safe deployment. Existing explainable AI (XAI) approaches primarily provide post-hoc insights and fail to produce executable, interpretable policies suitable for operational environments. In this paper, we present DeRAN, a neuro-symbolic framework that bridges the gap between DRL performance and operational transparency by distilling black-box DRL policies into human-readable symbolic representations. DeRAN introduces a concept-driven abstraction layer that transforms high-dimensional network telemetry into a compact set of semantically meaningful features, enabling interpretable policy learning. Building on the semantically grounded concepts, DeRAN synthesizes symbolic policies using deep symbolic regression (DSR) for continuous control and neurally guided differentiable logic (NUDGE) for discrete decision-making. We implement DeRAN on a live 5G O-RAN testbed and evaluate it on two representative use cases. Experimental results demonstrate that DeRAN achieves 78\% and 87\% of DRL's cumulative rewards in the two use cases, while offering interpretability and auditability by design. Source code is available at https://github.com/Jadejavu/A-Neuro-Symbolic-Framework-for-Interpretable-Open-RAN-Automation

• The paper introduces DeRAN, a neuro-symbolic teacher–student framework that converts DRL policies into specification-grounded concepts, symbolic expressions, and deterministic IF–THEN rules for auditable Open RAN automation.
• DeRAN recovered 78% of teacher reward for resource slicing and 87% for handover while reducing tail latency, using two-stage shielding to enforce constraints and retrieve previously verified-safe actions.
• Live 5G testbed results show that concept bottlenecks are critical: removing them reduced slicing reward from 0.782 to 0.583 and handover reward from 0.871 to 0.543, while symbolic policies met Near-RT RIC timing limits.

Motivation and problem statement

Deep reinforcement learning (DRL) has become the dominant engine for closed-loop control in Open Radio Access Networks (O-RAN), powering xApps for resource slicing, energy optimization, and mobility management. However, deploying DRL in carrier-grade networks is impeded by three structural deficiencies: stochastic action sampling inherited from exploration, opaque neural decision logic that cannot be traced to 3GPP/O-RAN specification vocabulary, and the absence of feasibility guarantees on hard QoS constraints. Existing explainable AI (XAI) approaches address only fragments of this gap. Post-hoc attribution methods such as SHAP and LIME quantify feature importance but leave the black-box model on the critical control path. Network-specific distillers (METIS, EXPLORA, ComTree, SymbXRL, SIA) produce rule sets over hundreds of raw key performance measurements (KPMs) that are mathematically valid but operationally unreadable — a semantic gap between low-level telemetry and the high-level predicates operators actually reason about.

DeRAN addresses this gap with a teacher–student neuro-symbolic framework: a trained DRL policy $\pi_\theta$ continues to explore as a teacher and streams traces $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$ into a rolling buffer; a symbolic student $\pi_\phi$ is continuously fitted to reproduce the teacher's pre-activation outputs and produces every deployed decision as either a closed-form expression or an IF–THEN rule over operator-named concepts grounded in 3GPP specifications (TS 28.552 KPMs; TS 38.331 A3/A5 events). The design targets three properties no prior system simultaneously offers: spec-level auditability, deterministic execution within the Near-RT RIC budget, and enforceable constraint compliance.

Framework design

DeRAN comprises three sequential modules.

Spec-grounded conceptizer. A concept template $$\mathcal{C} = \{(\eta_k, \mathcal{G}_k, \mathcal{M}_k)\}$$ is an expert-committed contract mapping each concept $k$ to its semantic name $\eta_k$, entity subset $\mathcal{G}_k$, and KPM subset $\mathcal{M}_k$. Each concept is computed by a per-concept encoder restricted by architectural masking to only its scoped KPMs and entities, aggregated permutation-invariantly via DeepSets-style summation. Because concept values are latent, supervision is indirect: an auxiliary linear head maps the concept vector back to the teacher's action-parameter space, and the conceptizer is trained to minimize reconstruction error against $\mathbf{z}_t$; the head is discarded after training. Post-hoc auditing uses Integrated Gradients to verify that attribution concentrates inside each concept's declared support and that sign patterns agree with semantics (e.g., BLER drives "URLLC stress" positively while CQI drives it negatively). The masking eliminates off-support leakage by construction rather than by regularization — a notable design choice given that naive encoders over all KPMs would contaminate concepts with unsupported inputs.

Per-dimension symbolic distillation. Joint symbolic search over multi-dimensional actions is combinatorially explosive, so DeRAN exploits the factorized output of DRL policies and distills each dimension independently. For continuous dimensions, deep symbolic regression (DSR) synthesizes expression trees over the concept vector, reformulated as a safe, asynchronous extraction on the telemetry buffer with a behavioral-fidelity reward rather than trial-and-error environment interaction. For discrete dimensions, DeRAN adapts NUDGE: predicates are enumerated from protocol specifications into threshold and comparison categories, conjunctions are realized via the product t-norm, per-action logits are attention-weighted pools of rule activations, and training is offline KL distillation against the teacher's marginals (replacing NUDGE's on-policy PPO). Rules are then compiled into deterministic IF–THEN tables executable in $O(|\mathcal{R}_i|)$ comparisons.

Two-stage symbolic action shielding. Distillation optimizes fidelity, not feasibility. Stage one applies spec-derived correction rules projecting closed-form violations (resource budgets, simplex constraints, bounds). Stage two, safe-decision retrieval, handles state-dependent QoS constraints admitting no static projection: a lightweight risk estimator gates execution, and when triggered, DeRAN replays the nearest previously verified-safe action from a bank indexed in low-dimensional concept space. Notably, this shielding can make the student outperform the teacher on QoS metrics, since retrieval suppresses risky exploratory decisions.

Experimental results

Evaluation runs on a live indoor multi-cell 5G NR testbed (srsRAN, commercial Benetel/Pegatron RUs, n78 band, 100 MHz, $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$0 MIMO, Open5GS core, ORAN-SC Near-RT RIC, 14 smartphones across three slices), with baselines (METIS decision trees, SymbXRL knowledge graphs) distilled from identical teacher traces. Teachers are PPO for slicing and Double DQN for handover.

Resource slicing (continuous control). DeRAN recovers 78% of the teacher's cumulative reward at complexity $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$1, versus SymbXRL's 0.65 reward at $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$2 and METIS's saturation at 0.43 even at $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$3. At the median, DeRAN achieves 87% reward recovery versus 73% (SymbXRL) and 51% (METIS). Strikingly, DeRAN is the only student that beats the teacher on QoS: 95% of steps maintain near-zero throughput violation rate versus 90% for the teacher, attributed explicitly to safe-decision retrieval. Tail latency drops from 3,800 μs (teacher) to 265 μs ($(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$4 speedup) with a $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$5 memory reduction.

UE handover (discrete control). DeRAN recovers 87% of teacher reward at $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$6, achieving median delay of ~50 ms against ~80 ms for the teacher and improving P99 delay by $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$7 over the teacher and up to $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$8 over METIS. The compiled policy is a four-branch IF–THEN table whose branches map legibly onto A3-style margin handover, rescue handover, and QoS-pressure-triggered mobility. P99 latency is 194 μs with 15 MB RSS.

Ablations. Across both tasks the conceptizer is the dominant contributor: removing it (distilling directly over raw KPMs) collapses slicing reward from 0.782 to 0.583 and handover reward from 0.871 to 0.543, because DSR/NUDGE search exhausts its budget in the raw-KPM space. Replacing symbolic distillers with MLP heads recovers modest reward but inflates $(t, \mathbf{s}_t, \mathbf{a}_t, \mathbf{z}_t)$9 by $\pi_\phi$0–$\pi_\phi$1, eliminating auditability. The two shield stages are complementary: rule-based correction primarily protects throughput violations, while safe-decision retrieval accounts for nearly all tail-delay reduction (omitting it degrades handover P99 from 135 ms to 265 ms due to ping-pong handovers under load).

Limitations and open questions

The paper concedes several boundaries. Reward recovery plateaus below parity (78% and 87%), so the framework trades measurable performance for transparency; whether this gap closes on larger-scale or wide-area deployments is untested. The concept templates are expert-committed and time-invariant contracts — their quality depends entirely on correct expert scoping, and the paper does not evaluate robustness to mis-specified or stale concept schemes under severe distribution shift. Safe-decision retrieval presumes that past verified-safe actions remain valid under current conditions, an assumption whose failure modes (e.g., prolonged non-stationarity emptying the effective safe bank) are not characterized. Evaluation is confined to two use cases on a single indoor testbed; scalability of the per-dimension distillation to policies with many coupled action dimensions remains open, as does the sensitivity of the risk-estimator trigger threshold $\pi_\phi$2.

Conclusion

DeRAN demonstrates that DRL-based O-RAN control can be replaced, at deployment time, by intrinsically interpretable symbolic students operating over specification-grounded concept bottlenecks, while retaining most of the teacher's reward and — through symbolic action shielding — sometimes exceeding it on QoS compliance and tail latency. Its principal contribution is empirical validation on a live 5G testbed that concept-bottlenecked symbolic distillation is compatible with Near-RT timing budgets, establishing a concrete mechanism for auditable AI-driven network automation.