Conceptio › Archive › arXiv CS

arXiv CSopen access

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

2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

Open Source ↗Direct PDF ↓

distributed-systems internet networking protocols

networking, internet, protocols, distributed systems

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

arXiv:2605.10648v1 [cs.NI] 11 May 2026

Jie Lu†

Peihao Yan† Pang-Ning Tan† Y. Thomas Hou‡ Huacheng Zeng† † Michigan State University ‡ Virginia Tech

Abstract

Black-box DRL

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 here.

Action

DRL Policy O-RAN Telemetry

QoS ? Opaque Hard to audit

Operator

DeRAN Load Latency Interference

Concepts

PRB = α · log(Load) + β

QoS ✓ Interpretable ✓ Auditable ✓

Deep Symbolic Regression

IF (Signal < ε) ∧ (Load > δ) THEN Handover Neurally-Guided Differentiable Logic

Operator

Figure 1: DeRAN transforms a black-box DRL model into a white-box symbolic policy.

inherent opacity of deep neural networks. Modern DRL agents operate as black-box controllers, mapping high-dimensional network telemetry to control actions through opaque nonlinear transformations. For network operators, deploying such unauditable systems introduces operational risks. Transparent and interpretable policies are essential to verify safety constraints and enable root-cause analysis when Quality of Service (QoS) degradation occurs. Consequently, realizing the full potential of DRL in O-RAN requires resolving the fundamental tension between performance and interpretability. Existing eXplainable AI (XAI) techniques fall short of providing the level of transparency required in operational RAN environments. Post-hoc attribution methods, such as SHAP [13] and LIME [19], quantify feature importance but do not recover the underlying decision logic of the neural network, leaving the opaque model on the critical control path. Network-specific distillation approaches, including METIS [15], EXPLORA [6], and ComTree [11], as well as symbolic auditing frameworks such as SIA [10], attempt to extract rules directly from hundreds of raw key performance measurements (KPMs). Although these rule sets are mathematically valid, they are often large, fragmented, and operationally unintelligible, exposing a fundamental semantic gap between low-level telemetry and the high-level predicates used by network operators. To date, no framework simultaneously delivers interpretable, auditable, and executable policies while preserving DRL performance and adaptability.

Introduction

Open Radio Access Networks (O-RAN) are rapidly evolving toward data-driven, automated control planes [18]. Deep Reinforcement Learning (DRL) has emerged as a powerful paradigm for optimizing complex, high-dimensional tasks such as dynamic network slicing [3, 24] and mobility management [14]. By learning directly from network interactions, DRL agents can achieve remarkable performance in highly dynamic environments. However, their deployment in production, carrier-grade networks faces a critical barrier: the 1

In this paper, we present DeRAN, a neuro-symbolic interpretation framework that addresses these challenges by distilling black-box DRL policies into compact, executable, and humaninterpretable representations, as shown in Fig. 1. Unlike conventional XAI methods that provide descriptive explanations, DeRAN produces intrinsically interpretable policies that can directly replace the original neural controller. The framework treats the trained DRL agent as a high-capacity teacher and extracts its decision logic into a lightweight symbolic student model, enabling deterministic and low-latency inference while preserving near-optimal performance. This transition from opaque function approximation to structured reasoning enables systematic auditing, validation, and safe deployment of AI-driven control in O-RAN systems. A key component of DeRAN is its concept-driven abstraction layer, which bridges the semantic gap between raw network telemetry and human reasoning. Instead of operating on hundreds of low-level key performance measurements (KPMs), the framework discovers a compact set of high-level, semantically meaningful concepts that capture the essential network state. This abstraction is derived through a principled pipeline combining attribution analysis and ablation-driven pruning, ensuring that each concept is both interpretable and causally relevant. By projecting high-dimensional telemetry into this structured concept space, DeRAN significantly reduces policy complexity and enables the synthesis of concise, operator-understandable decision rules. Building on these semantically grounded concepts, DeRAN synthesizes symbolic policies tailored to different action spaces. For continuous control tasks, such as resource slicing, the framework employs deep symbolic regression (DSR) [17] to derive closed-form mathematical expressions that approximate the teacher policy. For discrete decision-making tasks, such as mobility handover, DeRAN leverages neurally guided differentiable logic (NUDGE) [4] to construct compact FirstOrder Logic (FOL) rules that capture conditional decision structures. These symbolic policies are both interpretable and computationally efficient, making them suitable for deployment within O-RAN’s near-real-time control loops. To ensure safe and robust operation in dynamic network environments, DeRAN incorporates a two-stage symbolic action shielding mechanism. The first stage performs rule-based action projection, correcting closed-form violations using domain knowledge derived from 3GPP/O-RAN specifications. The second stage introduces a safe-decision retrieval module, which replays previously verified actions when a dedicated risk estimator detects state-dependent risk. Together, this mechanism enables DeRAN to rapidly respond to network dynamics while ensuring constraint compliance under uncertainty. We implement DeRAN on a live multi-cell 5G NR ORAN testbed and evaluate it on two representative DRL-based xApps: (i) resource slicing (continuous control) and (ii) UE handover (discrete control). Experimental results show that

DeRAN achieves 78% of the teacher DRL’s cumulative reward on the resource slicing task and 87% on the UE handover task, while delivering interpretability and auditability by design. More importantly, DeRAN also outperforms state-ofthe-art XAI baselines on both control performance and policy simplicity. To the best of our knowledge, DeRAN is the first XAI framework that has been implemented and validated in an operational O-RAN environment, rather than performing offline post-hoc interpretations. This work advances the state of the art as follows: • DeRAN is the first neuro-symbolic framework for O-RAN that distills black-box DRL policies into intrinsically interpretable and executable symbolic representations. • We design a learnable conceptizer that maps raw network telemetry into operator-aligned, semantically meaningful concepts. We further develop a per-dimension symbolic distillation pipeline to generate symbolic solutions for DRL policies. • We implement DeRAN on a real-world 5G O-RAN testbed and validate its effectiveness on representative DRL xApps under realistic network dynamics.

2 2.1

Preliminaries O-RAN and Its DRL Policies

O-RAN introduces a disaggregated and modular architecture that decouples traditional base station functionality into distinct components, including the Radio Unit (RU), Distributed Unit (DU), Centralized Unit (CU), and the Core network. A key advantage of O-RAN lies in its native support for datadriven and AI-enabled control. Through open interfaces such as the E2 interface, learning-based agents can continuously collect fine-grained network telemetry in the form of KPMs, process these observations to make control decisions, and enforce actions back onto the RAN. This capability is realized through programmable control entities, namely xApps in nearRT RIC and rApps in the non-RT RIC. Together, xAPPs and rAPPs establish a closed-loop, AI-driven control framework that enables continuous learning and adaptation in complex and dynamic wireless environments. System Model. We consider a DRL-based xApp in O-RAN that observes the network state st and outputs a control action at at time step t. Leveraging the open and standardized interfaces of O-RAN, the xApp can collect KPMs from all entities in the network, forming a global view of the system state. Let G denote the set of network entities, with cardinality G = |G |, where each entity could correspond to a user equipment (UE), a cell, or a slice, depending upon a specific task. Let M denote the set of KPMs extracted from each entity, with M = |M |. Then, the global network state is represented as st = [st,g,m ] ∈ RG×M , where st,g,m denotes the m-th KPM of entity g at time t. Table 1 provides an example of 2

State ID

Metric

Link

Description / Context

Slicing

s0 , s1 s2 , s3 s4 , s5 s6 , s7 s8 , s9−11

CQI, SNR UE PRB Throughput Delay Vol, Slice PRB

DL/UL DL/UL DL/UL DL/UL DL

Channel quality, PUSCH SNR Per-UE PRB usage Average data rate RLC SDU air-interface delay Transmitted Vol, Slice PRB usage

Handover

Table 1: KPM data samples serving as the network state. Task

s12 s13−15 s16−18 s19−22 s23,24 s25,26 s27−30

Cell Index Srv. RSRP/Q/I Nbr. RSRP/Q/I Thrpt, Delay CQI, SNR HARQ NACK Srv/Nbr PRB

– DL DL DL/UL DL/UL DL/UL DL/UL

Primary serving cell ID Serving cell power, quality, SINR Neighbor cell power, quality, SINR Avg. data rate and RLC delay Channel quality and PUSCH SNR HARQ Nack rate PRB usage for Srv. and Nbr. cells

controllers that map high-dimensional network telemetry to actions through opaque nonlinear transformations. (iii) QoS constraint violation: DRL lacks explicit mechanisms to guarantee feasibility, and its end-to-end neural decisions may violate QoS requirements or other system-level constraints.

2.2

Existing XAI methods can be broadly categorized into (i) post-hoc attribution techniques (e.g., SHAP [13], LIME [19]), which estimate feature importance for individual predictions, and (ii) symbolic approaches that produce inherently interpretable, rule-based representations. While attribution methods provide local, descriptive insights, they do not expose the underlying decision logic or offer a consistent global view. In contrast, symbolic approaches reconstruct explicit, structured policies that are transparent, verifiable, and better suited for debugging, auditing, and deployment in safety-critical systems such as O-RAN. At a high level, symbolic methods approximate complex black-box models with interpretable surrogates that capture their input–output behavior in a structured form. This is typically achieved through techniques such as symbolic regression, program synthesis, or differentiable logic learning, which search over human-readable expressions to fit a trained model. For DRL agents, this paradigm is particularly suitable, as policies define mappings from states to actions that can often be expressed through conditional logic or low-dimensional functional relationships. By distilling these policies into symbolic forms, one can recover the implicit decision boundaries and control rules encoded in neural networks. In the design of DeRAN, we will employ the following two symbolic representations. • Mathematical Expression Trees: For continuous control tasks (e.g., power control and resource slicing), policies are represented as algebraic expressions structured as trees, where leaf nodes correspond to input variables and internal nodes represent mathematical operators (e.g., +, −, ×, max) [26]. Evaluating the tree yields a deterministic continuous action.

KPMs collected from a single entity (i.e., a UE) under two representative tasks. Let π be a stationary decision-making policy. The objective is to find an optimal policy π∗ that maximizes the expected cumulative discounted reward: " # T

max Eτ∼π π

∑ γ t r(st , at ) ,

(1)

t=0

where at ∈ A is the control action, τ denotes the state-action trajectory induced by π, γ ∈ [0, 1) is the discount factor, and T is the decision horizon. The step-wise reward is defined as r(st , at ) = U(st , at ) − β Λ(st , at ), where U(·) captures the service utility reward, Λ(·) captures operator-defined QoS violations to penalize the reward, and β ≥ 0 is a coefficient. DRL-Based xApp. Deep reinforcement learning (DRL) provides a principled framework for online decision-making in the Near-RT RIC of O-RAN. The DRL policy πθ is typically implemented as a deep neural network (DNN) parameterized by θ. Given the network state st , the policy outputs a vector of action distribution parameters zt ∈ Rdz , where dz denotes the number of parameters required to specify the policy distribution (e.g., concatenated logits for discrete actions or mean and variance parameters for continuous actions). The decision-making process is defined as zt = πθ (st ),

at ∼ p(· | zt ),

A Primer on Symbolic Representations

(2)

• First-Order Logic (FOL): For discrete decisions such as handover, policies are expressed as conditional IF–THEN rules. FOL formalizes these rules using atomic predicates (e.g., CQI < 5) and logical operators such as AND (∧) and OR (∨) [8], enabling transparent, verifiable, and easily modifiable decision logic.

where zt parameterizes the action distribution p(· | zt ), and at is sampled accordingly. Although DRL has demonstrated strong performance and adaptability in O-RAN, its deployment in carrier-grade networks is hindered by inherent opacity and stochastic execution, which limit operator trust, auditability, and safe deployment. We attribute it to the following reasons. (i) Stochastic process: DRL policies inherently balance exploration and exploitation; however, exploration (e.g., sampling-based action generation) introduces randomness, leading to instability and unpredictability. (ii) Decision-making opacity: Most DRL approaches rely on DNNs, functioning as black-box

Problem Statement

DeRAN adopts a teacher-student structure. The teacher is the DRL policy πθ : it continues to explore the environment as in any standard deployment and produces the high-reward decisions we want to preserve. The student πφ is a transparent 3

Concept ID

Physical Meaning (ηk )

KPM Sets (Mk )

Slicing

c0 c1 c2 c3

eMBB Demand URLLC stress Slice load Channel quality

{s4 , s8 } {s0 , s6 } {s9−11 } {s0 }

c4 c5 c6 c7 c8

Srv. Signal Quality Tgt. Signal Quality Srv. Cell Load Tgt. Cell Load QoS Degradation

{s13−15 , s23 } {s16−18 } {s27,28 } {s29,30 } {s19 , s21,22 , s25 }

Teacher DRL πθ

O-RAN Environment

1 N d zt , ẑt , ∑ N t=1

DSR (continuous)

DSR

…

Φξ

× α

Action Shield Rule correction

IF violate THEN project

ât

β log

ẑ = α · log(ct,1 ) + β

ct,k ct,K

ât*

Rule Correction Safe Retrieval

NUDGE

ct,1 ct,2

Action Shield

ẑt

NUDGE (discrete)

Conceptizer

Safe retrieval Yes

AND

ât'

IF ct,2 < ε AND ct,1 > δ THEN Handover

Safe bank

ât*

Figure 2: Overview of DeRAN. where d(·, ·) is a distribution divergence between the teacher and student over the current buffer window N = |D |. The optimization is performed in two stages, i.e., a spec-grounded conceptizer followed by per-dimension symbolic distillers, which are detailed in §4.1. The student symbolic policy πφ is designed to satisfy the following requirements. (i) Deterministic execution. During real-time operation, the student generates actions via a fixed component-wise head g(·): ẑt = πφ (st ; C ),

(3)

ât = g(ẑt ),

(5)

where g(·) is a deterministic function, corresponding to argmax over categorical logits for discrete actions and the distribution mean for continuous actions. (ii) Spec-level auditability. The decision process is expressed as either a closedform expression or a set of IF–THEN rules over expertdefined concepts in C , making each control decision traceable to underlying KPMs and specification events. (iii) Constraint compliance. A shielding mechanism is applied to the generated action, projecting each candidate ât onto the feasible region defined by network constraints.

where K = |C | denotes the number of concepts, determined by the protocol specifications. For each concept k, ηk is the semantic meaning that operators can directly audit against, i.e., "Target-Cell Load", "Per-Slice Demand". Gk ⊆ G is the subset of network entities related to concept k (e.g., the subset of UEs belong to the eMBB slice k), and Mk ⊆ M is the subset of KPMs related to concept k. At each step t, DeRAN represents the concept-level of the network by a concept vector ct = [ct,1 , . . . , ct,K ]⊤ ∈ [0, 1]K , whose k-th entry ct,k is the activation value. Once the control task and its spec basis are chosen, C is an expert-committed and time-invariant contract that does not change during deployment. In contrast, ct is computed at runtime from the current state st by the conceptizer (§4.2), whose parameters are learned from the buffer D . Experts define each concept template and its dependencies once, while DeRAN learns to compute its value at runtime. Distillation objective. Given concept template C and data trace buffer D , DeRAN trains the symbolic student policy πφ to track the teacher’s pre-activation outputs by minimizing min Ldistill (φ) =

Conceptizer

Φξ

policy that runs on the Near-RT RIC. The teacher streams its traces (t, st , at , zt ) into a rolling buffer D , and the student is continuously fitted to D = {(t, st , at , zt )} so that its distribution ẑt tracks the teacher’s zt . This paradigm lets the teacher keep its learning flexibility while the student provides the three properties the teacher cannot: spec-level auditability, deterministic execution, and constraint enforcement. Concept Template. A concept is an expert-defined and task-level abstraction that aggregates a subset of related KPMs into a single semantically meaningful indicator. 3GPP/ORAN specifications (e.g., 3GPP TS 28.552 [1]) enumerate tens of KPMs per task, but the logic reasoning operates on only a few higher-level predicates, e.g., "target cell is strong" or "target slice is overloaded". Each derived from several KPMs and their interactions (e.g., a servingvs-neighbor RSRP comparison). Denote C as the concept template of a given control task. Then, it can be formulated as a three-element tuple:

C = {(ηk , Gk , Mk )}Kk=1 ,

Symbolic Distiller

…

Task

Handover

Table 2: Concept examples of resource slicing and handover.

DeRAN: Design

4.1

Overview

Fig. 2 illustrates the overall architecture of DeRAN and its interaction with the teacher DRL model at runtime. DeRAN consists of three sequential modules. The spec-grounded conceptizer compresses st into a concept vector ct . The symbolic distiller then maps each dimension of the teacher output zt to a human-readable form (either a closed-form expression for continuous control or an IF–THEN rule set for discrete control) over ct . Finally, the action shielding module projects the

(4)

candidate action into the feasible region defined by operator and specification constraints before applying it to the RAN. The design of DeRAN faces three key challenges. First, spec-grounded concept inference: operators define concepts in C but do not provide their numerical values, requiring the conceptizer to infer ct from the teacher’s behavior using only scoped KPMs. We address this by combining architectural masking in per-concept encoders with indirect supervision via an auxiliary linear head, avoiding off-support regularization (details in §4.2). Second, multi-dimensional symbolic distillation: jointly distilling high-dimensional actions leads to combinatorial explosion. To address this, we exploit the factorized structure of the teacher policy and distill each dimension independently, using deep symbolic regression (DSR) for continuous outputs and neurally guided differentiable logic (NUDGE) for discrete ones (§4.3). Third, constraint enforcement: distillation prioritizes fidelity to the teacher rather than feasibility, so the resulting actions may violate system constraints. We address this with a two-stage symbolic action shielding mechanism, combining rule-based projection for closed-form constraints with safe-decision retrieval for statedependent risks (§4.4). Fig. 3 illustrates the logic flow of the proposed action distillation process, along with the key notation and a handover example.

4.2

Figure 3: The logic flow of the proposed action distillation. zt . Specifically, we introduce an auxiliary linear head (A, b) with A ∈ Rdz ×K that maps the concept vector back to the teacher’s action-parameter space z̃t = A ct + b. The conceptizer is trained by minimizing the fidelity loss min Lfid = ξ,A,b

∑ |B | t∈ B

z̃t − zt 2 ,

(7)

which requires ct to preserve sufficient information to reconstruct the teacher’s decisions on a batch B ⊆ D . After training, the auxiliary head (A, b) is discarded, since it serves only as a differentiable surrogate for supervising the conceptizer. The conceptizer parameters ξ are then frozen, and the symbolic distillers in §4.3 take ct as input. Post-hoc auditing. Post-hoc auditing addresses two questions that operators can examine via Integrated Gradients (IG) [23]: (i) whether the KPMs within Mk influence ct,k in a manner consistent with the semantic name ηk , and (ii) which KPM contributes most to the concept value at a given sample. Let sbase denote a baseline state. Let IGk (st ) ∈ R|Gk |×|Mk | be a tensor-valued attribution matrix. Its entry (g, m) quantifies how strongly KPM [st ]g,m influences the value of concept ct,k . Computed over Nig interpolation steps, this attribution is given by 1 Nig ∂ct,k s(ℓ) [IGk (st )]g,m = [st ]g,m − [sbase ]g,m · ∑ ∂[s(ℓ) ]g,m , (8) Nig ℓ=1

Spec-Grounded Conceptizer

A key idea of DeRAN is that the symbolic student operates on operator-named concepts rather than raw KPMs. To build semantic connection between KPMs st and concept ct , we define Φξ (·) as the conceptizer that maps each state st to a concept vector ct = Φξ (st ) ∈ [0, 1]K . Per-concept encoder and prediction head. For the k-th concept, denote Mk as the set of its associated KPMs and Gk as the set of its associated entities. Rather than processing the entire global state st , the k-th concept operates only on the KPMs in Mk for the entities in Gk . Specifically, the concept value ct = [ct,1 , ct,2 , . . . , ct,K ] is computed by ct,k = ρk ∑ hk [st,g ]Mk , (6) g∈Gk

where [·]Mk returns the subvector indexed by Mk ; hk (·) : R|Mk | → Rdh is a parameterized per-concept encoder; and ρk (·) : Rdh → [0, 1] is a parameterized per-concept prediction head that maps the aggregated embedding into the concept value ct,k . Here, dh is a shared hyperparameter across all concepts, hk and ρk are instantiated as lightweight feedforward networks. Note that the sum in (6) is permutation-invariant following DeepSets [25]. When |Gk |=1, the sum reduces to a single term and ρk processes the entity’s embedding directly. ξ is the collection of all ρk (·) and hk (·) functions in Eq. (6). Training. Since concept values are latent, supervision is provided indirectly from the teacher’s pre-activation outputs

where g ∈ Gk , m ∈ Mk , and s(ℓ) = sbase + Nℓig (st − sbase ) is the ℓ-th point on the linear path between sbase and st .

4.3

Symbolic Distillation

Based on the semantically-grounded concepts, DeRAN is then to extract the domain knowledge embedded within the blackbox policy πθ and synthesize an executable symbolic policy. RAN control typically involves a large multi-dimensional action space, and directly searching for a joint symbolic policy across all action dimensions simultaneously leads to a combinatorial explosion in the hypothesis space. To address this 5

con where ẑt,i = πcon i (ct ). Maximizing J(πi ) drives the search toward more faithful expressions to reproduce the highperformance continuous decisions of the black-box DRL policy πθ . To keep the generated formulas human-readable, we prune candidate formulas whose tree depth exceeds the maximum tree depth Dmax during search. The highest-reward expression is adopted as the interpretable initialization policy πcon for the action under consideration. An Example. To illustrate this procedure, consider a dynamic resource slicing task in O-RAN, where the teacher policy allocates a fractional PRB ratio to each network slice. Focusing on a slice i (e.g., an eMBB slice), the teacher outputs a scalar zt,i denoting its PRB-ratio allocation. Suppose the conceptizer has produced two concepts: ct,1 : "Aggregate Cell Load" summarizes the cell-wide congestion level, and ct,2 : "Slice Queue Backlog" captures the aggregate queuing pressure across slices. The DSR generator combines these variables and operators to synthesize a transparent symbolic expression for action dimension i, for example: 1 log(1 + ct,2 ). This expression admits a direct opẑt,i = ct,1 erational meaning: the allocated bandwidth share grows logarithmically with the slice’s queue backlog and scales inversely with the cell-wide load. More generally, the synthesized tree yields the deterministic continuous control output ẑt,i = πcon i (ct ), making the teacher’s decision on action dimension i explicit and auditable.

issue, we leverage the factorized architecture of DRL policies, which parameterize independent marginal distributions, yielding a concatenated global parameter vector zt . Specifically, we decompose the distillation into parallel procedures along each action dimension. The physical correlations and joint constraints across dimensions are subsequently enforced a posteriori by an action shielding module (§4.4). At each step t, the teacher policy πθ outputs a vector zt , which factorizes along the action dimension i as zt,i ∈ Rdi . Each zt,i parameterizes the marginal distribution of at,i : • Continous Actions: When Ai ⊆ R is continuous, we treat the teacher as a deterministic policy and distill the mean only. Thus di = 1 and zt,i reduces to a scalar i that specifies the mean of the policy distribution; • Discrete Actions: When Ai is a discrete set, di = |Ai | and zt,i holds the categorical logits over Ai . The teacher’s per-dimension action at,i is then sampled from this distribution governed by zt,i . 4.3.1

Distillation for Continuous Actions

For continuous-action tasks such as transmit power control and resource slicing, DeRAN adapts DSR [17] to generate the mathematical expressions that approximate the teacher’s decision function over ct . The core idea of DSR is to employ an autoregressive neural network, referred to as the expression generator, to search the discrete space of mathematical formulas, thereby transforming equation discovery into a sequence generation problem. Expression Construction. We instantiate the dimensionspecific continuous student policy πcon directly as an expresi sion tree, whose leaf variables are the entries of the concept vector ct and the internal nodes are mathematical operators. K Equivalently, πcon i : [0, 1] → R is a deterministic scalar function that the expression generator synthesizes as a sequence of operators and variables. Evaluating the tree on the current concept vector yields the student’s prediction ẑt,i = πcon i (ct ). To satisfy the timing requirements for Near-RT RIC execution, we restrict the operator set to the set below:

4.3.2

Distillation for Discrete Action

Data-Driven Structural Search. The original DSR formulation relies on extensive trial-and-error interaction with the environment, which would bring unacceptable QoS violations in production networks. We therefore reformulate DSR into a safe and asynchronous extraction procedure that operates entirely on the streaming telemetry buffer D = {(t, st , zt , at )}. Each candidate tree is evaluated by a behavioral fidelity reward against the teacher’s target zt,i :

For discrete actions such as beam selection or UE handover, DeRAN distills the teacher’s marginal zt,i ∈ R|Ai | into a set of IF-THEN rules over the concept vector ct . The challenge is that the student should be symbolic for operators to audit, yet differentiable to fit the teacher’s distribution on the buffer D = {(t, st , at , zt )}. To this end, we build on the NUDGE framework [4], which jointly optimizes rules via gradient descent and compiles them into deterministic policies at deployment. The discrete distiller consists of three modules: (i) a spec-grounded vocabulary P derived from C , keeping every predicate traceable to protocol specifications; (ii) a differentiable rule layer that composes predicates into conjunctions and pools their activations into per-action logits via attention; and (iii) offline KL distillation that fits ẑt,i to the teacher and compiles the rules into a deterministic IF-THEN table. Predicate vocabulary. We enumerate a predicate vocabulary P from the protocol specifications related to C , with P = |P |. Each predicate p ∈ P falls into one of two categories: • Threshold predicates (i.e., “ct,k is either low or high”) for magnitude conditions in KPM-driven policies, e.g., load comparisons as specified in 3GPP.28.552 [1];

N 2 −1 1 J(πcon ) = 1 + ẑ − z , t,i t,i ∑ i N

• Comparison predicates (i.e., “ct,k exceeds ct,k′ by offset”) for spec-named measurement events, e.g., the A3/A5 events specified in 3GPP.38.331 [2].

O = {+, −, ×, ÷, log(·), exp(·)}.

(9)

(10)

t=1

Distillation training. Trained under the KL objective, the attention within each Ri,n sharpens onto the rule(s) that best explain the teacher’s decisions for candidate action n. Unlike NUDGE, which updates rule weights via on-policy PPO, DeRAN trains the rule layer by offline distillation on D . With temperature κ > 0, the teacher and student distributions are jointly optimized by z ẑ 1 N Ldis (W, ψ) = ∑ DKL softmax( t,i ) ∥ softmax( t,i ) . N t=1 κ κ (15) The distillation can be run asynchronously on an rApp within the non-real-time loop, avoiding resource competition with the tasks within the near-real-time control loop. Compilation to deterministic IF-THEN. For each candidate action logit n ∈ Ai , the dominant rule u⋆n = arg maxu∈Ri,n αu is preserved, and rules whose in-subset attention αu falls below a pruning threshold Wth are discarded. The result is a plain IF-THEN table, i.e., one deterministic branch per candidate action, that operators can audit line by line and that runs in O(|Ri |) comparisons on the Near-RT RIC. This compiled program is the student policy πdis . After per-dimension distillation, DeRAN concatenates the dimension-wise outputs {ẑt,i }i into the full student parameter vector ẑt , and applies the deterministic head g(·) defined in Eq. (5) to obtain the candidate student action ât = g(ẑt ). Because each dimension is distilled independently from the teacher’s factorized marginals, ât encodes the student’s perdimension intent but does not yet enforce joint operator/spec constraints. It is therefore handed to the symbolic action shield (§4.4), which produces the final executed action ât⋆ .

Boolean-valued predicates have been widely used in classical FOL rules (e.g., p is hard True or False). However, it is not effective and informative in our problem. Therefore, we introduce a confidence vector, defined as vt = [vt,1 , vt,2 , . . . , vt,P ] ∈ [0, 1]P , to quantify the predicate valuation. For example, suppose P contains prsrp : High Target RSRP and pload : Low Target Load; then vt,prsrp =0.9 and vt,pload =0.2 indicate 90% confidence that “RSRP is very high” and 20% confidence that “the target cell has low load”. A single affine mapping network parameterized by ψ = (Wψ , bψ ), with Wψ ∈ RP×K and bψ ∈ RP , maps ct to vt . For each predicate p ∈ P , let w p ∈ RK denote the p-th row of Wψ and b p ∈ R the p-th entry of bψ . Then, we have vt,p = σ w⊤ p ct + b p , where w p is masked so each predicate depends only on the concepts it is defined over, and b p is a learnable threshold. Rule body. For each action i, we associate a finite rule set Ri , where each rule u ∈ Ri is a conjunction of predicates from P proposing a discrete candidate action in Ai (e.g., IF p ∧ p′ ∧ p′′ THEN n, with p, p′ , p′′ ∈ P and n ∈ Ai ). This conjunction is realized by the product t-norm: vt,u = ∏ vt,p ,

(11)

p∈u

which transforms the non-differentiable “AND” operator to differentiable “product” operator. Continuing the above example, the rule u: “IF prsrp AND pload THEN handover” yields vt,u = vt,prsrp · vt,pload = 0.18, which means rule u has only a weak impulse of 0.18 to be executed under the current network state. Action head. A rule activation vt,u captures how strongly a rule u activates, but not which candidate action the rule should trigger. Recall that the teacher’s pre-activation vector zt,i ∈ Rdi has di = |Ai | logits for discrete actions. Let n ∈ {1, . . . , |Ai |} index the logits of zt,i . To align rules u with logit n, we partition rule set Ri by the logit, let

Ri,n = {u ∈ Ri : head(u) = n}

4.4

Since the above distillation focuses on behavioral fidelity rather than feasibility, the candidate student action ât may occasionally violate operator-defined or system-level constraints. We design a two-stage symbolic action shielding mechanism to produce the final action ât⋆ . The first stage is rule-based correction, which handles constraints expressed in closed form. The second stage is safe-decision retrieval, which handles state-dependent QoS constraints by replaying a previously verified decision. Rule-based correction. Constraints tractable in closed form, e.g., resource budgets, admissibility sets, per-dimension bounds, are encoded as a finite set of spec-derived correction rules R̃ = {ũ}. Each rule ũ couples a closed-form violation function ν with a closed-form correction operator Π:

(12)

collect the rules proposing candidate action logit n, so that S Ri = n∈Ai Ri,n . DeRAN assigns each rule u ∈ Ri a learnable weight wu ∈ R. Within each subset Ri,n , these weights are normalized into an attention distribution αu =

exp(wu ) , exp(wu′ ) ∑ i,n u′ ∈R

u ∈ Ri,n ,

Symbolic Action Shielding

(13)

which is unambiguously defined since each rule belongs to exactly one subset. Here, αu is read as the fraction of candidate action n’s score attributed to rule u. The n-th logit of ẑt,i is then the attention-weighted sum of activations within that subset: [ẑt,i ]n = ∑ αu · vt,u . (14)

ũ : IF ν(ât , st ) > 0 THEN ât ← Π(ât , st ),

(16)

where ν is positive when (ât , st ) violates the constraint encoded by ũ (e.g., the resource budget is exceeded), and Π projects the action back into the feasible region, for instance, via simplex renormalization, clamping to per-entry bounds,

u∈Ri,n

or rounding to the nearest admissible value. Starting from the candidate ât , rules are applied in spec-prioritized order; the result is denoted as ât′ . Safe-decision retrieval. Rule-based correction handles only constraints expressible in closed form and acts pointwise on the candidate ât . The corrected action ât′ can still violate state-dependent QoS constraints introduced in §2.1, e.g., per-slice throughput or delay targets, whose feasibility depends on how the radio environment subsequently evolves under ât′ and therefore admits no static projection. DeRAN closes this gap by replaying past decisions that have already been observed to satisfy QoS in operationally similar contexts, rather than synthesizing a new feasible action online. DeRAN maintains an online safe-decision bank as Dsafe = (c j , a⋆j ) , (17)

Figure 4: A multi-cell indoor 5G NR O-RAN testbed. • Q3 System overhead. Does DeRAN satisfy the Near-RT RIC latency and memory requirements? • Q4 Component contribution. How does each DeRAN module (conceptizer, symbolic distiller, action shield) contribute to the end-to-end result?

collecting past executed actions whose subsequently observed T ′ -step QoS-violation cost Λ(st , ât⋆ ) stayed within operator operator-tolerated bounds, with T ′ ≤ T . Entries are collected only after post-hoc verification, so every element of Dsafe is known-safe in the context in which it was executed. A lightweight risk estimator q : [0, 1]K ×A → [0, 1], instantiated as a two-layer MLP, scores the constraint risk of ât′ and acts only as a QoS-violation gate. We train q offline on logged triples (ct , at , yt ), where yt ∈ {0, 1} flags whether Λ stayed within tolerance over the subsequent T ′ steps. Once q is trained, we record the empirical distribution of its scores {q(c j , a⋆j )} j∈Dsafe on the training buffer and fix the trigger threshold δ based on the distribution of the score. Retrieval then executes when q(ct , ât′ ) > δ. Upon this trigger, the shield replaces ât′ with the knownsafe entry whose concept vector is closest to the current one, k⋆ = arg

min

k: (ck ,·)∈Dsafe

∥ct − ck ∥2 ,

ât⋆ ← âk⋆ .

5.1

Fig. 4 shows our 5G NR O-RAN testbed deployed within a building. The protocol stack is built on srsRAN 25.10 [22] with Split 7.2 fronthaul and Split Option 2 midhaul. Commercial RUs from Benetel and Pegatron are integrated into the deployment, operating on the 5G n78 band with 100 MHz bandwidth and 4 × 4 MIMO. The core network is implemented on Open5GS. To emulate realistic network dynamics, the testbed incorporates 14 smartphones of mixed vendors (OnePlus Nord AC2003, Motorola G54, Samsung Galaxy A15). Traffic is generated via iperf to produce continuous and heterogeneous flows across three network slices with timevarying loads. The ORAN-SC Near-RT RIC [21] is containerized and deployed on a server with NVIDIA RTX PRO 6000 Blackwell GPU. DeRAN’s fast-loop (conceptizer inference and shield) runs inside an xApp co-located with the RIC, while the slow-loop (conceptizer and symbolic distiller training) runs asynchronously on an rApp host. GitHub link to the source code will be provided if the paper is accepted for publication.

(18)

Proximity is measured in concept space rather than in raw state space, so the metric is low-dimensional and the O(|Dsafe |·K) query is cheap. Otherwise (when q(ct , ât′ ) ≤ δ), the rule-corrected action passes through directly: ât⋆ ← ât′ .

5.2 5

5G NR O-RAN Testbed

Experiments

Performance Metrics and Baselines

Baselines. We compare DeRAN against its DRL teacher and two SOTA XAI baselines, Metis [15] and SYMBXRL [5]. To isolate the effect of the distillation method from that of the teacher, both baselines are trained the same teacher traces as DeRAN from the buffer D ; once distilled, every student policy is frozen and deployed on the testbed for online evaluation without further teacher involvement. • Teacher. We use Proximal Policy Optimization (PPO) for continuous resource slicing and Double DQN for discrete UE handover. They serve as the performance upper bound.

We evaluate DeRAN on two representative O-RAN control tasks: (i) resource slicing representing continuous action space (§5.3), and (ii) UE handover representing discrete action space (§5.4). Both tasks are evaluated on a live indoor multi-cell 5G NR testbed shown in Fig. 4. The evaluation seeks to answer four questions: • Q1 Interpretability. Are the concept and the student policy spec-auditable, and how simple are they relative to SOTA XAI baselines under a complexity metric Ω? • Q2 Control performance. Does DeRAN preserve the teacher’s reward and QoS guarantees on live 5G traffic?

• Metis [15]. We use its decision-tree mode, which fits a regression tree or a classification tree (depending on the 8

task type) directly to the teacher’s trace (st , at ) over the raw KPMs.

Use Case 1: Resource Slicing

5.3.1

Task Formulation

DeRAN (Ours) SYMBXRL Metis

Reward

0.6

0.4

s4 −0.5

−1.0

0.0

0.5

Attribution Score

1.0

-1

Figure 5: IG attribution from KPMs to slicing concepts.

200

400

600

Complexity

Figure 6: Reward vs. symbolic complexity.

zeMBB = log(1 + c0 ) + 1.32 c3 + 0.62 c0 c3 − 1.86 c2 − 0.45 c1 , zURLLC = 1.74 c1 + 1.26 c21 + 0.62 c1 c2 − 0.74 c3 , zmMTC =

1 − 0.36 log(1 + c2 ) + 0.23 c3 . 1.42 + c0 + c1

Figure 7: An instance of the distilled formulas for slicing.

construction. The sign pattern also agrees with the semantic names: s6 (BLER) drives c1 (URLLC stress) positively while s0 (CQI) drives it negatively, consistent with “high stress = poor channel + high error rate.” Complexity. Fig. 6 reports the reward-complexity Pareto comparison of the three XAI methods under a sweep of hyperparameters (number of leaves for Metis, KPM change threshold for SYMBXRL, expression-tree depth for DeRAN. At its highest-reward operating point, DeRAN achieves mean step reward 0.71 at Ω=56, while SYMBXRL requires Ω=206 to reach 0.65 and Metis saturates at 0.43 even at Ω=532. Compared to SYMBXRL, DeRAN achieves 1.09× higher reward with 3.7× fewer nodes; relative to Metis, the gap widens to 1.65× reward at 9.5× lower complexity. The gap is driven by the concept bottleneck: DeRAN distills over four operatornamed variables, whereas Metis and SYMBXRL must recover the same structure from |M | = 12 KPMs. Symbolic policy auditing. The distilled DeRAN student for each slice is a single closed-form expression ẑt,i = πcon i (ct ) over four concepts, which we provide in Fig. 7. Each expression is operationally readable: the eMBB logit rises with eMBB demand (log(1 + c0 )) and channel quality (c3 ), and is penalized by competing slice load (c2 ) and URLLC stress (c1 ); the URLLC logit grows super-linearly in c1 (a near-quadratic reflex to rising URLLC stress) and is inversely coupled to channel quality; the mMTC logit is a residual claim on PRBs that relaxes when either eMBB or URLLC demand grows.

QoS penalty [p]

where nt,g is the PRBs allocated to UE g, rt,g = max((µ̃t,g − [d]

µt,g )/µ̃t,g , 0), and rt,g = max((ωt,g − ω̃t,g )/ω̃t,g , 0) are the QoS penalties, where µt,g , ωt,g are the achieved throughput and delay, (µ̃t,g , ω̃t,g ) are the corresponding QoS targets, weighted by β1 , β2 > 0. In addition, since the throughput and delay demands vary across slices, we report the per-step violation rates Vthp (t) = |Gt |−1 ∑g∈Gt 1(µt,g < µ̃t,g ) and Vdly (t) = |Gt |−1 ∑g∈Gt 1(ωt,g > ω̃t,g ). The concept scheme CRS is in the Table 2, i.e., eMBB demand, URLLC stress, slice load, channel quality. 5.3.2

c1 c0

We consider a case where a cell serves a set of time-varying UEs. Here, Gt is the set of UEs, with |Gt | ∈ [9, 14]. The UEs are partitioned into three disjoint slice-level subsets so that Gt = GeMBB ∪ GURLLC ∪ GmMTC . Denote NPRB as the total number of physical resource blocks (PRBs). The state st is the per-UE KPM row {st,g }g∈Gt of Table 1, and the action at = (at,eMBB , at,URLLC , at,mMTC ) is a continuous per-slice PRB ratio on the simplex. The PPO teacher maximizes h i [p] [d] 1 rtrs = ∑ . (19) − β r + β r 1 2 t,g t,g 1+∑g∈G nt,g g∈Gt {z } | | {z t } Resource efficiency

s9 s10 s11

Value

Concept

• SYMBXRL [5]. This approach discretizes the continuous states into first-order-logic (FOL) predicates and constructs a knowledge graph over the predicates and actions. Interpretability metric Ω. We quantify the complexity of a policy by Ω, defined as the number of atoms in its symbolic representation. An atom is the smallest indivisible syntactic unit of the representation: nodes of the expression tree (operators, variables, constants) for mathematical expressions; logical connectives, comparison operators, predicates, variables, and constants for FOL rules; nodes for decision trees; subjects, relations and objects in the knowledge graphs. We report Ω jointly with the mean step reward as a Pareto curve. Training and reporting. Each baseline is evaluated across five independent runs of 104 decision steps, with each step triggered by a KPM report from the E2 node every 500 ms.

5.3

0.8

Interpretability 5.3.3

Concept auditing. Fig. 5 shows the normalized IG values of each KPM for each concept over the aggregated UEs. For every concept ck ∈ CRS , the attribution concentrates inside the support Mk in by Table 2: c0 (eMBB demand) loads on s4 , s8 ; c1 (URLLC stress) on s0 , s6 ; c2 (slice load) on s9 , s10 , s11 ; and c3 (channel quality) on s0 . This confirms that the perconcept masking in §4.2 eliminates off-support leakage by

Control Performance

Fig. 8 reports the CDF of reward together with Vthp and Vdly for DeRAN, Metis, SYMBXRL, and the teacher over five independent runs. Reward. In Fig. 8a, DeRAN follows the teacher closely across the entire reward distribution: the median reward of DeRAN is 0.65 against the teacher’s 0.75 (a 87% recovery), 9

0.5

0.0

0.5

1.0

Reward

1.00

0.5

0.0

CDF

1.0

DeRAN (Ours) Teacher SYMBXRL Metis

CDF

1.0

DeRAN (Ours) Teacher SYMBXRL Metis

0.0

0.5

Table 4: Ablation study of DeRAN on the resource-slicing task. Reward is normalized w.r.t. the teacher; Vthp and Vdly are calculated for per-UE step violation rates.

0.95 0.90

DeRAN (Ours) Teacher SYMBXRL Metis

0.85

1.0

0.0

0.5

1.0

Delay Violation Rate

Thp Violation Rate

(a) CDF of reward. (b) CDF of throughput (c) CDF of delay vioviolation rate Vthp . lation rate Vdly .

Figure 8: Control performance comparison on the resourceslicing task. Table 3: System overhead on the resource-slicing task. P50 latency (µs)↓ P99 latency (µs)↓ max (µs) ↓ RSS (MB) ↓

Method Teacher Metis SYMBXRL DeRAN (Ours)

2,100 45 380 95

3,800 110 950 265

7,200 340 2,100 520

410 12 85 18

Option

Reward ↑

Vthp % ↓

Vdly % ↓

Ω↓

Teacher

–

1.000

7.4

1.5

—

DeRAN (full)

–

0.782

3.6

0.6

Conceptizer

w/o conceptizer (DSR) w/o per-concept masking

0.583 0.755

15.2 8.1

12.4 7.3

78 62

Distiller

DSR → linear on ct DSR → MLP on ct

0.743 0.868

11.2 2.9

9.8 1.8

34 3,217

Shield

w/o rule-based correction w/o safe-decision retrieval w/o both shields

0.763 0.752 0.622

12.0 6.8 17.6

5.9 10.7 15.8

56 56 56

Ω↓ — 532 206 56

5.3.5

Ablation Studies

Table 4 reports the contribution of each DeRAN module on the slicing task. Three findings stand out. (i) The conceptizer is the dominant driver of reward performance: replacing it with DSR directly over the raw KPMs drops the reward from 0.782 to 0.583 and increases Vthp and Vdly by 4.2× and 20.7× respectively, because DSR on a ∼12-dimensional KPM vector exhausts its search budget. Removing only the per-concept masking (but keeping the concept scheme) costs 0.03 reward but doubles Vthp . (ii) DSR outperforms linear distillation and achieves comparable performance to a MLP with lower complexity: replacing DSR by a linear head on ct loses 0.04 reward, while replacing it by a dense MLP recovers 0.086 reward over full DeRAN but increases Ω by 57×, eliminating auditability. (iii) The two shield stages are complementary: removing rule-based correction primarily damages Vthp (driven by simplex violations), while removing safe-decision retrieval primarily damages Vdly (driven by state-dependent QoS risk); removing both is strictly worse than either ablation and brings the student close to an un-shielded symbolic controller.

and the lower tail at CDF= 0.1 is within 0.02 of the teacher. SYMBXRL lags the teacher by 0.20 at the median (73% recovery), and Metis collapses to a 0.38 median (51% recovery) because the hard splits in the tree generation cannot express the smooth multi-slice trade-off that the teacher learns. QoS. DeRAN is the only student that outperforms the teacher. Fig. 8b shows that 95% of DeRAN’s decision steps maintain Vthp ≈ 0, against 90% for the teacher, 80% for SYMBXRL, and 68% for Metis; the crossover between DeRAN and the teacher at the low tail is produced by the safe-decision retrieval stage of the action shielding mechanism (§4.4), which replaces the candidate action whenever the risk estimator flags throughput risk. The delay-violation distribution in Fig. 8c follows the same ordering: DeRAN (97% nearzero) ≈ teacher (94%) > SYMBXRL (92%) > Metis (83%). Aggregated across all steps, DeRAN reduces the mean Vthp by 3.9× versus SYMBXRL and 5.8× versus Metis, and the mean Vdly by 2.3× and 5.1× respectively. 5.3.4

Variant

System Overhead

Table 3 compares the 50th-percentile (P50) lantecy, 99thpercentile (P99) latency, and the memory footprint measured by peak resident-set size (RSS) of the teacher, Metis, SYMBXRL, and DeRAN. The teacher operates on raw KPMs through a DNN forward pass, while DeRAN inference on the concept vector ct followed by a per-dimension closedform expression on ct . As a result, DeRAN reduces the 99thpercentile latency from 3,800 µs (teacher) to 265 µs (a 14.3× speedup) and the RSS from 410 MB to 18 MB (22.8×). Compared to SYMBXRL, which must traverse a knowledge graph at every step, DeRAN is 3.6× faster in terms of the 99thpercentile latency. Although Metis yields the lowest P99 latency (110 µs), it comes at the expense of a 9.5× larger Ω=532 overhead and a severe performance drop to a 0.38 median reward. Overall, DeRAN achieves a Pareto improvement over all baselines regarding the reward-latency trade-off.

5.4

Use Case 2: UE Handover

5.4.1

Task Formulation

A target UE g in the overlap of two cells G = {serv, tgt} decides each step to stay on the serving cell (at =0) or switch to the target cell (at =1), with zt ∈ R2 holding the corresponding Q-values. The state KPMs st are shown in Table 1. The teacher maximizes [p] [d] rtho = − β3 rt,g + β4 rt,g , (20) | {z } QoS penalty

where β3 , β4 > 0 are weights, and the two penalty terms for throughput and delay deviations are defined identically to (19). The concept scheme CHO is in the Table 2, i.e., serving/target signal quality and load (c4 −c7 ) and QoS-degradation status (c8 ), with |Gk |=1 for the signal-quality and load concepts. 10

s30

s28 s s1816

−1.0

s27

−0.5

0.0

0.5

Attribution Score

0.4 DeRAN (Ours) SYMBXRL Metis

0.2

s17 s s 13 s14 15 s23

0.6

s29

1.0

-1

Figure 9: IG attribution from KPMs to handover concepts.

0.0

100

0.8

0.5

0.6 0.4

DeRAN (Ours) Teacher SYMBXRL Metis

0.2 0.0

0.0 0.0

150

0.5

1.0

200

(a) CDF of reward.

Figure 10: Reward vs. symbolic complexity.

400

600

Delay (ms)

Reward

Complexity

(b) CDF of delay.

Figure 12: Control performance on the UE-handover task. Table 5: System overhead on the UE-handover task.

if c5 − c4 > 0.12 and c7 < 0.47 then switch else if c4 < 0.32 and c5 > 0.58 then switch else if c8 > 0.50 and c5 > 0.63 and c7 < 0.42 then switch else stay end if

P50 latency (µs)↓ P99 latency (µs)↓ max (µs) ↓ RSS (MB) ↓

Method Teacher Metis SYMBXRL DeRAN (Ours)

1,800 30 290 52

3,100 70 820 194

5,400 180 1,800 420

180 9 75 15

Ω↓ — 103 152 25

and lightly loaded. The table runs in four comparisons and is line-by-line auditable against the 3GPP TS 38.331 A3/A5 vocabulary.

Figure 11: An instance of the distilled rules for handover task. 5.4.2

1.0

DeRAN (Ours) Teacher SYMBXRL Metis

CDF

1.0

0.8

s21 s25

CDF

s22

Reward

s19

Value

Concept

Interpretability 5.4.3

Concept auditing. Fig. 9 reports the IG attribution of the five handover concepts across the handover KPM set in Table 1. The attribution pattern matches the semantic names: c4 (serving signal quality) loads on {s13−15 , s23 } (serving RSRP/RSRQ/SINR/CQI), c5 (target signal quality) on {s16 , s17 , s18 }, c6 , c7 split onto serving- and target-side PRB utilization (s27 , s28 vs s29 , s30 ), and c8 (QoS degradation) loads positively on BLER/delay-burst KPMs while loading negatively on s19 (achieved throughput), i.e., degradation grows when throughput drops. Crucially, the serving- and target-cell supports of c4 and c5 are disjoint in the IG map: no KPM from the target cell contaminates the serving-cell concept or vice versa, which is a direct consequence of the per-concept masking of §4.2. Complexity. Fig. 10 plots the reward-complexity Pareto for the three XAI methods. DeRAN’s highest operating point is (Ω=26, reward=0.71), against (Ω=152, reward=0.55) for SYMBXRL and (Ω=103, reward=0.38) for Metis. Relative to SYMBXRL, DeRAN achieves 1.29× higher reward with 6.0× fewer nodes; relative to Metis, 1.87× reward at 4.0× lower complexity. The gap is again attributable to the concept bottleneck: the distiller reasons over five named concepts rather than the 19 raw KPMs available to the baselines, so the higher reward target is reachable with far simpler predicates. Symbolic policy auditing. The compiled DeRAN student (Fig. 11) is a four-branch IF-THEN table in which every threshold is a concept comparison (rather than a raw KPM). The first branch encodes the canonical A3 semantics at the concept level (“target signal exceeds serving by a margin and target is not overloaded”); the second branch triggers a rescue handover when the serving cell has degraded in absolute terms; the third branch binds mobility to QoS pressure, firing only when c8 is elevated and the target is both strong

Control Performance

Fig. 12 reports the per-step reward and delay CDFs for DeRAN, the teacher, Metis, SYMBXRL. Reward. DeRAN demonstrates competitive performance to the teacher : the median reward is 0.70 against the teacher’s 0.76, a 92% recovery. SYMBXRL and Metis underperforms at medians of 0.55 and 0.45. QoS. DeRAN outperforms the three baselines in terms of the delay: the median per-step delay of DeRAN is ∼50 ms, against ∼80 ms for the teacher, ∼150 ms for SYMBXRL, and ∼350 ms for Metis. At P99, DeRAN reaches 140 ms, improving the teacher by 1.2×, SYMBXRL by 2.9×, and Metis by 4.1×. The performance improvement over the teacher policy comes from the action shielding stage. This mechanism replays a previously verified handover whenever the risk estimator is triggered due to potential ping-pong or congestiondriven delay spike, effectively mitigating the tail of the delay. 5.4.4

System Overhead

As shown in Table 5, DeRAN achieves a P99 latency of 194µs and a peak RSS of 15 MB. Compared to the teacher policy’s 3,100 µs and 180 MB, this represents a 16× reduction in tail latency and a 12× decrease in memory footprint. Furthermore, avoiding expensive knowledge-graph traversals makes DeRAN 4.2× faster than SYMBXRL at the 99th percentile. Although Metis evaluates faster (70 µs), it comes at a 4.1× higher structural complexity (Ω = 103 vs. 25), as analyzed in §5.4.2. Ultimately, while all XAI students satisfy the Near-RT processing budget, DeRAN is the only method that uniquely strikes the optimal balance: it combines sub-millisecond execution and the lowest structural complexity (Ω = 25) while maintaining 87% of the teacher’s reward. 11

Table 6: Ablation study of DeRAN on the UE-handover task. Variant

Option

Teacher

–

P99 Delay (ms) ↓

Ω↓

1.000

172

—

Work

Paradigm

Output

METIS [15] EXPLORA [6] AICHRONO. [7] ComTree [11] SymbXRL [5] SIA [10] inRAN [27] DeRAN (Ours)

Post-hoc Post-hoc Post-hoc Post-hoc Post-hoc Post-hoc Intrinsic Intrinsic

Tree Graph Attribution Tree Graph Graph Formula Formula&Rule

DeRAN (full)

–

0.871

135

Conceptizer

w/o conceptizer (raw-KPM) w/o per-concept masking

0.543 0.782

388 342

44 38

Distiller

NUDGE → linear on ct NUDGE → MLP on ct

0.794 0.868

172 148

30 2,105

Shield

w/o safe-decision retrieval

0.805

265

5.4.5

Table 7: Comparison of existing XAI work for O-RAN.

Reward ↑

Ablation Studies

✗ ✓ ✓ ✓ ✓ ✓ ✗ ✓

✓ ✓ ✗ ✓ ✓ ✓ ✗ ✓

✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓

a Semantic high-level representations (✓) rather than raw KPMs (✗). b Enable Near-Real-Time (Near-RT) execution (✓) otherwise (✗).

We evaluate the same components as in the slicing scenario, but exclude the simplex correction because it is inapplicable to binary handover actions. Table 6 presents our experimental results. We observe two findings consistent with the slicing case and one unique to the handover: (i) The conceptizer is the primary contributor to the reward performance. Replacing it with raw-KPM predicates degrades the reward from 0.871 to 0.543 and increases the P99 delay from 135 ms to 388 ms. This degradation occurs because NUDGE’s rule search suffers from state-space explosion in the raw-KPM space and fails to effectively distinguish between serving and target cells. Furthermore, per-concept masking alone yields a 0.09 reward improvement and a 2.5× reduction in P99 delay. (ii) NUDGE significantly outperforms linear distillation. It achieves a reward within 0.003 of an MLP model while maintaining an 81× lower Ω. This confirms that the IF-THEN rules are sufficiently expressive for handovers when grounded in high-level concepts. (iii) Safe-decision retrieval accounts for nearly the entire tail-delay reduction. Omitting this module maintains a relatively high reward (0.805) but degrades the P99 delay to 265 ms, as the distilled IF-THEN rules occasionally trigger ping-pong handovers under extreme load conditions.

Semantica NRTb QoSc

c Provides QoS guarantees (✓) otherwise (✗).

hypergraphs to produce human-readable rules. However, these static conversions often lack the wireless context required for root-cause analysis. EXPLORA [6] addresses this gap by using attributed graphs to link agent actions to the input state space. To handle the temporal complexity of mobile traffic, AICHRONOLENS [7] integrates classical XAI with timeseries analysis. SIA [10] further introduces an influence score to audit forecast-augmented agents within sub-millisecond windows, and reveals issues such as temporal misalignment. SYMBXRL [5] and ComTree [11] explore FOL and large language model (LLM)-guided metrics to improve the readability of explanations. However, these methods remain reactive to distribution shifts. inRAN [27] addresses the non-stationarity issue by employing Kolmogorov-Arnold Networks (KANs) as interpretable surrogate models for online Bayesian learning at the cost of moderate computational overhead. Table 7 compared DeRAN against existing XAI methods. Knowledge Distillation for DRL. Knowledge distillation transfers behavior from a high-capacity teacher model into a compact student model. Policy distillation was first introduced for deep Q-networks [20]. Subsequent symbolic extensions, such as deep symbolic policies (DSP) [17] and NUDGE [4], further distill neural controllers into closed-form expressions and differentiable first-order logic (FOL) rules, respectively. Unlike prior work on classical control or game benchmarks, DeRAN brings symbolic policy distillation to the O-RAN control plane by grounding concepts in 3GPP/O-RAN specifications and enforcing QoS and Near-RT constraints through symbolic action shielding for safe deployment.

Related Work

DRL for Network Automation. DRL has become a key driver of network control and management in O-RAN. For example, ORANSlice [3] demonstrates DRL-based xApps for closed-loop resource slicing, while xSlice [24] and EExApp [12] incorporate GNNs and multi-actor PPO for topologyaware slicing and energy-efficient RU scheduling. In mobility and system optimization, prior work explores hierarchical multi-agent DRL for handover [14] and frameworks such as Mutant [16] and DTOTO [9] for congestion control and MEC offloading. Despite these advances, existing approaches rely on black-box neural networks with limited interpretability and high computational overhead, hindering efficient deployment on resource-constrained platforms. XAI for Networking. Research on network interpretability has evolved from basic structural conversion toward real-time symbolic reasoning. METIS [15] introduces the distillation of deep neural network (DNN) policies into decision trees and

Conclusion

In this paper, we presented DeRAN, a system for transforming opaque DRL-based control in O-RAN into interpretable, verifiable, and deployable policies through symbolic distillation. By introducing a spec-grounded conceptizer, DeRAN bridges low-level KPMs and high-level operator-defined concepts, enabling human-aligned representations of network state. Building on this abstraction, the symbolic distiller converts neural 12

policies into compact mathematical expressions and logical rules, while preserving fidelity to the original DRL behavior. To ensure safe deployment, the symbolic action shielding module enforces system-level constraints through projection and retrieval mechanisms. Overall, this work highlights a practical path toward trustworthy AI-driven network control by combining learning-based optimization with symbolic reasoning and domain knowledge.

References

[10] MohammadErfan Jabbari, Abhishek Duttagupta, Claudio Fiandrino, Leonardo Bonati, Salvatore D’Oro, Michele Polese, Marco Fiore, and Tommaso Melodia. Sia: Symbolic interpretability for anticipatory deep reinforcement learning in network control. arXiv preprint arXiv:2601.22044, 2026.

[1] 3GPP. Management and orchestration; 5G performance measurements. Technical Specification (TS) TS 28.552, 3rd Generation Partnership Project (3GPP), 2024. Version 19.1.0, Release 19.

[11] Lianchen Jia, Chaoyang Li, Ziqi Yuan, Jiahui Chen, Tianchi Huang, Jiangchuan Liu, and Lifeng Sun. Beyond interpretability: Exploring the comprehensibility of adaptive video streaming through large language models. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 12035–12044, 2025.

[2] 3GPP. NR; Radio Resource Control (RRC); Protocol specification. Technical Specification (TS) TS 38.331, 3rd Generation Partnership Project (3GPP), 2025. Version 18.5.1, Release 18. [3] Hai Cheng, Salvatore D’Oro, Rajeev Gangula, Sakthivel Velumani, Davide Villa, Leonardo Bonati, Michele Polese, Tommaso Melodia, Gabriel Arrobo, and Christian Maciocco. Oranslice: An open source 5g network slicing platform for o-ran. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 2297–2302, 2024.

[12] Jie Lu, Peihao Yan, and Huacheng Zeng. Eexapp: Gnnbased reinforcement learning for radio unit energy optimization in 5g o-ran. arXiv preprint arXiv:2602.09206, 2026. [13] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017.

[4] Quentin Delfosse, Hikaru Shindo, Devendra Dhami, and Kristian Kersting. Interpretable and explainable logical policies via neurally guided symbolic abstraction. Advances in Neural Information Processing Systems, 36:50838–50858, 2023.

[14] Irshad A Meer, Karl-Ludwig Besser, Mustafa Ozger, Dominic Schupke, H Vincent Poor, and Cicek Cavdar. Hierarchical multi-agent drl based dynamic cluster reconfiguration for uav mobility management. IEEE Transactions on Cognitive Communications and Networking, 2025.

[5] Abhishek Duttagupta, MohammadErfan Jabbari, Claudio Fiandrino, Marco Fiore, and Joerg Widmer. Symbxrl: symbolic explainable deep reinforcement learning for mobile networks. In IEEE INFOCOM 2025-IEEE Conference on Computer Communications, pages 1–10. IEEE, 2025.

[15] Zili Meng, Minhu Wang, Jiasong Bai, Mingwei Xu, Hongzi Mao, and Hongxin Hu. Interpreting deep learning-based networking systems. In Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication, pages 154–171, 2020.

[6] Claudio Fiandrino, Leonardo Bonati, Salvatore D’Oro, Michele Polese, Tommaso Melodia, and Joerg Widmer. Explora: Ai/ml explainability for the open ran. Proceedings of the ACM on Networking, 1(CoNEXT3):1–26, 2023.

[16] Lorenzo Pappone, Alessio Sacco, and Flavio Esposito. Mutant: Learning congestion control from existing protocols via online reinforcement learning. In 22nd USENIX Symposium on Networked Systems Design and Implementation (NSDI 25), pages 1507–1522, 2025.

[7] Claudio Fiandrino, Eloy Pérez Gómez, Pablo Fernández Pérez, Hossein Mohammadalizadeh, Marco Fiore, and Joerg Widmer. Aichronolens: advancing explainability for time series ai forecasting in mobile networks. In IEEE INFOCOM 2024-IEEE Conference on Computer Communications, pages 1521–1530. IEEE, 2024.

[17] Brenden K Petersen, Mikel Landajuela, T Nathan Mundhenk, Claudio P Santiago, Soo K Kim, and Joanne T Kim. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients. arXiv preprint arXiv:1912.04871, 2019.

[8] Joseph Y Halpern and Vicky Weissman. Using firstorder logic to reason about policies. ACM Transactions on Information and System Security (TISSEC), 11(4):1– 41, 2008.

[18] Michele Polese, Leonardo Bonati, Salvatore D’oro, Stefano Basagni, and Tommaso Melodia. Understanding o-ran: Architecture, interfaces, algorithms, security, and research challenges. IEEE Communications Surveys & Tutorials, 25(2):1376–1411, 2023.

[9] Zhihao Hu, Yaozong Yang, Wei Gu, Ying Chen, and Jiwei Huang. Drl-based trajectory optimization and task offloading in hierarchical aerial mec. IEEE Internet of Things Journal, 12(3):3410–3423, 2024.

[19] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the 14

predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016. [20] Andrei A Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy distillation. arXiv preprint arXiv:1511.06295, 2015. [21] SRS. oran-sc-ric. [Online]. Available: https:// github.com/srsran/oran-sc-ric, 2024. [22] SRS. srsRAN Project. [Online]. Available: https: //github.com/srsran/srsRAN_Project, 2025. [23] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319–3328. PMLR, 2017. [24] Peihao Yan, Jie Lu, Huacheng Zeng, and Y Thomas Hou. Near-real-time resource slicing for qos optimization in 5g o-ran using deep reinforcement learning. arXiv preprint arXiv:2509.14343, 2025. [25] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017. [26] Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, Kaixuan Wang, and Xudong Liu. A novel neural source code representation based on abstract syntax tree. In 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE), pages 783–794. IEEE, 2019. [27] Ming Zhao, Yuru Zhang, Qiang Liu, Ahan Kak, and Nakjung Choi. inran: Interpretable online bayesian learning for network automation in open radio access networks. arXiv preprint arXiv:2601.03219, 2026.

Record · ID 175158 · SHA-256 cefbad787e5ab278
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.