IRIS: Intelligent Random Access with Integrated Shared Learning for Decentralized LEO Direct-to-Satellite Uplinks

IRIS: Intelligent Random Access with Integrated Shared Learning for Decentralized LEO Direct-to-Satellite Uplinks Abdullah Al Noman∗ , Gaurav Singh∗ , Rameez Ahmed† , and Debashri Roy∗ ∗ The University of Texas at Arlington, Arlington, TX, USA † MathWorks Inc., USA

Abstract—Grant-free uplink access in low-Earth-orbit (LEO) non-terrestrial networks (NTNs) poses a coordination challenge when multiple UEs share a satellite channel without real-time centralized scheduling under long propagation delays and timevarying channels. Conventional random access does not exploit peer context, leading to collisions and inefficient channel use as contention increases. To address this problem, we propose IRIS, a multi-agent reinforcement learning framework for decentralized grant-free direct-to-satellite (D2S) uplink access, with A2A-IRIS as its sidelink-assisted Agent-to-Agent (A2A) realization. Each UE executes a shared binary transmission policy using local traffic/channel information augmented by compact peer states exchanged over UE-to-UE sidelink, while Vanilla-IRIS serves as a local-only ablation. Evaluations in rural and suburban LEO D2S deployments use TLE-based trajectories propagated using the Simplified General Perturbations 4 (SGP4), with a Walker-Delta constellation as fallback. A2A-IRIS achieves 2.1579 and 5.4478 Mbps, corresponding to 99.9% and 96.1% of the respective scenario-specific upper bounds. Relative to VanillaIRIS, throughput improves by 40.8% and 325.0%, with collision rates of 0.0000 and 0.0433, while one- and two-epoch delayed centralized schedulers are outperformed by approximately 51– 130%. A2A-IRIS also improves fairness, service coverage, and transmission efficiency with moderate Deep Q-Network (DQN) inference overhead. Active-UE load scalability results further show sustained performance over the evaluated load range. These results demonstrate the effectiveness of sidelink-assisted A2A coordination for decentralized LEO D2S uplink access under propagation delay and delayed channel feedback. Index Terms—Low-Earth-orbit satellite networks, direct-tosatellite, uplink random access, MARL.

I. I NTRODUCTION The growing integration of low-Earth-orbit (LEO) satellite systems with terrestrial cellular networks is expanding the role of non-terrestrial networks (NTNs) in beyond-5G and sixthgeneration (6G) communications [1]. In particular, directto-device and direct-to-satellite connectivity enables ordinary user equipment (UE) devices to communicate with satellite infrastructure without requiring a terrestrial base station in the service-link path, thereby extending coverage to areas where terrestrial radio access is unavailable, sparse, or temporarily disrupted [2]. Among the different NTN architectures, LEO systems are especially advantageous because their lower orbital altitude reduces propagation delay relative to higher-orbit satellite systems while supporting wide-area coverage. Challenges in LEO D2S Uplink Access. Despite these

D2S Links A2A Coordination Scheduling Request Scheduling Decision

LEO Satellite

Centralized Scheduler Trained IRIS Agent Lower scheduling delay Reduced control overhead Better scalability Collision-aware local coordination

High control overhead Scheduling delay Poor scalability under dense access

550 Km

arXiv:2611.07201v1 [cs.NI] 4 Nov 2026

Emails: abdullahal.noman@uta.edu, gaurav.singh@uta.edu, rrasheed@mathworks.com, debashri.roy@uta.edu

UE1

UE2

. . . UEn

UE1

UE2 A2A

(a) Satellite-Based Centralized Scheduling Decision

. . . UEn A2A

(b) Proposed Vanilla/A2A-IRIS D2S Uplink Scheduling

Fig. 1. Conceptual comparison between centralized LEO D2S uplink scheduling and the proposed IRIS decentralized Agentic AI framework. All UEs act as autonomous agents executing a shared policy using locally available information. The sidelink-assisted configuration additionally exploits Agentto-Agent (A2A) peer context, whereas the ‘Vanilla’ (local-only) configuration operates without A2A peer-state exchange over the sidelink.

advantages, efficient uplink access remains challenging. As illustrated in Fig. 1, a large number of geographically distributed UEs may contend for a shared satellite uplink, and poorly coordinated simultaneous transmissions can result in collisions, wasted transmission energy, increased waiting time, and persistent queue buildup. These effects become more pronounced under bursty traffic and finite UE buffers, where repeated unsuccessful access attempts directly reduce the useful payload delivered over the satellite link. The access problem is further complicated by the dynamics of the LEO channel. Rapid satellite motion causes time-varying visibility, elevation-dependent path loss, propagation delay, and substantial Doppler shifts, all of which must be handled by the air interface [1], [3], [4]. For representative LEO deployments, 3GPP studies report service-link delays on the order of several milliseconds together with Doppler shifts reaching tens of kilohertz at S-band frequencies [4]. Although centralized scheduling remains possible in NTN

systems, these propagation and channel dynamics increase the staleness of channel-state information (CSI) and the controlloop overhead required for rapid per-epoch decisions based on fresh network-wide channel information. The increasing complexity and scale of these dynamic telecommunications environments demand intelligent automation, where Agentic AI has emerged as a pivotal paradigm enabling autonomous agents to perceive, reason, act, and continuously learn to optimize network configurations. This motivates access mechanisms that transition decision-making directly to the UE edge, relying on distributed intelligence rather than strictly centralized control loops. Limitations of Existing NTN Access Mechanisms. 3GPP has introduced several NR-NTN mechanisms for satellite operation. Release 17 includes NTN-specific timing management, GNSS-assisted Doppler pre-compensation, mobility support, and HARQ adaptations [1], [5], [6]. These mechanisms address synchronization, mobility, and PHY/MAC operation over long and time-varying satellite links, but do not directly determine how multiple continuously backlogged UEs should autonomously coordinate repeated access to a shared uplink contention domain. Prior work has addressed complementary NTN randomaccess issues, including two-stage detection for LEO NTN IoT access [7] and reinforcement learning (RL) and multiagent reinforcement learning (MARL)-based NTN resource allocation [8]. More broadly, Agentic AI extends multiagent learning toward self-organizing and adaptive network operation. However, decentralized Agentic AI for handheld UEs making binary access decisions over a single shared D2S contention domain under partial observability remains insufficiently explored. Proposed IRIS Framework. To address this problem, we propose IRIS, an Agentic AI-driven random-access framework based on parameter-shared MARL for decentralized LEO D2S uplinks. As illustrated in Fig. 1, conventional satellitebased scheduling relies on request/decision exchanges over the D2S path, incurring control overhead and delay. IRIS instead moves access decisions to the UE side, where each trained agent independently selects its uplink action. A2AIRIS further exploits sidelink-based Agent-to-Agent (A2A) coordination, allowing compact peer information to assist contention resolution without per-epoch satellite scheduling. All UEs share a DQN with binary action space {NT, T}, where NT and T denote no transmission and transmission, respectively. Parameter sharing keeps the policy parameterization independent of the UE population. A2A-IRIS is the proposed sidelink-assisted realization, while Vanilla-IRIS is its local-only ablation. Vanilla-IRIS observes normalized queue occupancy, effective SINR, waiting time, previous access outcome, and a UE-specific coordination token. A2A-IRIS additionally uses a fixed-dimensional histogram of delayed peer queue, channel, and waiting states, together with current relative contention and predicted-payload ranks. Peer exchange leverages the existing 3GPP PC5 sidelink interface [9], rather than introducing a new coordination

interface. Although conceptually related to Google’s generalpurpose A2A agent-interoperability paradigm [10], IRIS uses a lightweight, domain-specific A2A mechanism for D2S random access. Peer information does not deterministically select a transmitter; each UE independently evaluates the shared DQN and chooses NT or T. Thus, the Vanilla-IRIS comparison isolates the benefit of peer context under the same action space, reward, PHY, traffic, and learning framework. During training, load-aware exploration limits excessive random transmissions, event-preserving replay retains informative transmission events, and action-aware minibatch sampling reduces domination by NT transitions. A cooperative reward combines UE-specific service utility with a common access-outcome term to discourage collisions, idle epochs with backlog, and singleton PHY failures while promoting urgent service. During inference, the DQN is frozen and each UE executes greedily without replay insertion, gradient updates, or target-network synchronization. Our Contributions. The main contributions are: (1) Parameter-Shared Decentralized Agentic AI for D2S Random Access: We develop IRIS, an Agentic AI framework in which homogeneous UEs act as autonomous agents that share a common DQN and independently select between NT and T over a single shared D2S uplink contention domain. Parameter sharing avoids per-UE networks and enables decentralized online execution from local observations. (2) Agent-to-Agent (A2A) Sidelink-Assisted Observation Configurations: We propose A2A-IRIS as the sidelinkassisted realization and Vanilla-IRIS as its local-only ablation. Both use local queue, effective-SINR, waiting, previousoutcome, and contention-token information, while A2A-IRIS additionally incorporates fixed-dimensional peer context. Delayed peer states are compressed into a normalized histogram and combined with a current contention-priority rank, keeping the DQN input dimension independent of the UE population. (3) Access-Aware Deep-RL Training and Cooperative Reward: We design training for the imbalanced binary-access problem using load-aware exploration, event-preserving replay, and action-balanced mini-batches to control random attempts, retain informative TX experiences, and limit domination by silence transitions. The reward combines delivered payload, queue/waiting urgency, and a cooperative accessoutcome term to discourage collisions, unnecessary idling, PHY failures, and persistent UE capture. (4) Population- and Geometry-Driven LEO D2S Evaluation: We evaluate IRIS using population-density-driven UE populations, pedestrian traffic and mobility, and TLEbased trajectories [11] propagated with SGP4 [12]. The evaluation further incorporates a 3GPP-inspired LEO Sband channel with elevation-dependent propagation, Doppler compensation, MCS adaptation, BLER, and HARQ under suburban and rural conditions. A2A-IRIS is compared with Vanilla-IRIS, a centralized upper bound, current and one-/twoepoch delayed channel-quality-based centralized schedulers, p⋆ -ALOHA, and binary exponential backoff. Performance is evaluated using throughput, access outcomes, fairness, RF

energy and transmission resource costs, DQN inference complexity, and active-UE load scalability. II. R ELATED W ORK AND M OTIVATION NTN and LEO Satellite Architectures. Large LEO constellations have accelerated the integration of satellite access with terrestrial cellular systems and strengthened the role of NTNs in beyond-5G/6G networks [13]. Lin et al. [1] discussed 3GPP NTN standardization challenges in synchronization, mobility, timing advance, and random access. Recent studies have also explored multi-layer D2C architectures, including spectrumintelligent multi-tier LEO association and UAV-assisted spaceair-ground relaying for improving resource utilization and link reliability [14], [15]. These studies highlight that LEO access must accommodate time-varying geometry, mobility, and control-plane dynamics that differ substantially from static terrestrial RANs. Random Access and Synchronization in NTN. LEO random access is affected by differential propagation delay, Doppler, and geographically distributed terminals sharing satellite coverage. Chougrani et al. [16] studied NB-IoT NTN preamble detection and uplink synchronization under these conditions. Recent work has also addressed grant-free massive access. Shen et al. [17] proposed massive-MIMO OTFS access with joint activity detection and channel estimation; Ying et al. [18] considered cooperative multi-satellite quasi-synchronous access; Shen et al. [19] extended cooperative OTFS access to joint device identification, channel estimation, and symbol detection; and Zhang et al. [20] investigated grant-free NOMA with joint user identification, channel estimation, and data detection. These approaches primarily improve waveformand receiver-level access reliability, whereas our focus is decentralized UE-side control of whether to contend for the shared D2S uplink. Collision Detection and Access Robustness. Jeong et al. [21] proposed early preamble-collision detection for LEO NTN IoT access to accelerate collision recovery, while two-stage preamble detection has also been studied for robust LEO NTN IoT access [7]. Such methods improve collision detection and recovery after access begins; in contrast, our objective is to reduce simultaneous transmissions through decentralized UEside decisions. Learning-Based Random Access and Resource Control. Learning-based access control has been explored for adaptive random access. Jiang et al. [22] used DRL to tune access parameters under varying IoT traffic, while Lee et al. [23] proposed eRACH for decentralized LEO contention without interagent communication. Yu et al. [24] considered DRL-based NOMA slotted ALOHA with repetition and SIC, whereas Oh et al. [25] studied decentralized MARL with limited peer exchange in terrestrial settings. Queue-aware distributed access, exemplified by Q-CSMA, and CTDE-based MARL, exemplified by QMIX, provide complementary approaches to decentralized coordination [26], [27]. In contrast, A2A-IRIS uses a shared binary access policy with fixed-dimensional peer-assisted observations for LEO D2S uplink contention.

Research Gap and Motivation. Despite prior work on NTN architectures, synchronization, random-access robustness, collision handling, and learning-based control, decentralized UE-side contention for sustained LEO D2S uplink traffic remains underexplored. Existing studies largely emphasize PHY-layer access, receiver-side processing, centralized coordination, richer resource allocation, or terrestrial decentralized access settings whose signaling and carrier-sensing assumptions do not directly match LEO D2S operation. IRIS addresses this gap through parameter-shared MARL with a common DQN for binary {NT, T} access, evaluated under local-only and sidelink-assisted observations to isolate the value of compact peer-assisted coordination. III. S YSTEM M ODEL AND P ROBLEM F ORMULATION A. LEO D2S Network and Population-Driven Traffic Model We consider a LEO D2S uplink with a set U = {1, . . . , U } indexing the active UEs, which communicate directly with a serving LEO satellite over a dedicated FDD uplink carrier. Satellite positions and velocities are obtained from two-line elements (TLEs) propagated using SGP4, with a Walker-Delta constellation retained as a fallback when TLE-based geometry is unavailable. At decision epoch t, UE i considers satellites satisfying θi,s (t) ≥ θmin ,

(1)

where θi,s (t) is the elevation angle of satellite s as observed by UE i, and θmin is the minimum visibility elevation. UE i then associates with the candidate providing the largest desired-link reference received power. Thus, satellite association evolves with orbital geometry and the large-scale link budget. The satellite receive beam is directed toward the modeled service region. The active UE population is coupled to the deployment environment. For a square service region of side length Lsvc and population density ρpop , 
U = max 1, round ρpop L2svc ηcand ηact , (2) where ηcand is the population-to-candidate-handheld mapping and ηact is the activity factor. Hence, the contention load follows from the deployment density and service area rather than being independently selected. Scenario-specific values are given in Section V-A. UEs follow pedestrian mobility with independently drawn headings, which contributes to the relative satellite–UE velocity and instantaneous Doppler shift. Each active UE is offered traffic at bit-arrival rate λ, with the number of arriving bits per epoch modeled as Ai (t) ∼ Poisson (λ∆t) , (3) where ∆t is the MARL control epoch rather than an NR physical slot. Each UE maintains a finite queue of capacity Bmax . At the beginning of epoch t, newly generated uplink traffic Ai (t) is added to UE i’s existing backlog Bi (t) before channel observation and access selection. Hence, the post-arrival, preservice queue is

ei (t) = min {Bi (t) + Ai (t), Bmax } . B (4) If Di (t) payload bits are successfully delivered during epoch t, the queue evolves as n o ei (t) − Di (t), 0 . Bi (t + 1) = max B (5) Thus, queue evolution jointly reflects traffic arrivals, decentralized access decisions, contention outcomes, and the time-varying D2S link; silence, collisions, and PHY failures increase queue pressure, whereas successful singleton transmission reduces the accumulated backlog. B. Contention-Based Grant-Free Uplink Access Model All backlogged UEs contend over a single shared grant-free uplink domain. At each decision epoch t, UE i selects ai (t) ∈ Ai = {NT, T}, (6) where NT and T denote no transmission and transmission, respectively; empty-buffer UEs always select NT. Let o X n NT (t) = 1 ai (t) = T, Bei (t) > 0 (7) i∈U

denote the number of transmitting UEs. If NT (t) = 0, the epoch is idle; if NT (t) = 1, the singleton UE is evaluated by the D2S PHY and uses the complete modeled UL allocation of 24 PRBs; and if NT (t) > 1, a MAC-layer collision occurs with no recovered payload. Hence, contention failures are distinguished from singleton PHY failures caused by visibility, insufficient effective SINR, or stochastic decoding. The contention domain is modeled globally; hence, simultaneous transmissions are treated as a collision even if the involved UEs associate with different serving satellites. Let χi (t) ∈ {0, 1} denote the singleton decoding outcome and Ci (t) its PHY service opportunity. The delivered payload is Di (t) = 1{ai (t) = T, NT (t) = 1, χi (t) = 1} n o (8) ei (t), Ci (t) . × min B The common previous access outcome is   0, idle or unsuccessful singleton, (9) o(t) = 1, successful singleton,   2, collision. The policies do not enforce a deterministic single-winner rule; they learn to favor useful singleton access while reducing collisions and unnecessary idling. This single-contentiondomain model intentionally isolates decentralized access coordination and does not represent a complete NR scheduler with parallel PRB allocation, frequency-domain resource selection, or multi-user decoding. C. Channel and PHY Model The D2S PHY accounts for orbital geometry, free-space propagation, environment-dependent attenuation, satellite receivebeam gain, shadowing, small-scale fading, Doppler, residual carrier-frequency offset, inter-carrier interference (ICI), and thermal noise. For UE i, the large-scale propagation loss is Li (t) = LFS,i (t) + Ltropo,i (t) + Liono,i (t) + Arain,i (t), (10)

and the received power is Pr,i (t)[dBW] = EIRPi − Li (t) − Lcl,i (t)

(11) + Gr (ϕi (t)) + Gfad,i (t) + Gsh,i (t), where Lcl,i (t) denotes NLOS clutter loss, Gr (ϕi (t)) the satellite receive-beam gain, Gfad,i (t) the small-scale fading gain, and Gsh,i (t) the shadow-fading term. The receive beam follows a circular-aperture model and is directed toward the service-area center. Suburban and rural LOS probability, LOS/NLOS shadowing, and NLOS clutter loss are selected from the elevation-dependent S-band parameters in 3GPP TR 38.811 [28], using the nearest available reference elevation. LOS links follow an elevation-dependent Rician model, while NLOS links use Rayleigh fading, with an NTN-TDL-like multipath profile. Temporal fading correlation is modeled as ρf,i (t) = J0 (2π|fd,i (t)|∆t) , (12) where fd,i (t) is the geometry-dependent Doppler shift. Other visible satellites are not modeled as uplink interferers; simultaneous UE transmissions are handled by the MAC collision rule. Thus, singleton transmissions are noise limited before the residual Doppler-induced ICI penalty. We further assume GNSS-aided Doppler pre-compensation, with the residual frequency error modeled as p zi (t) ∼ N (0, 1), δfi (t) = αc δfi (t−1)+ 1 − αc2 σc zi (t), (13) where αc and σc control temporal correlation and residual uncertainty, respectively. The normalized residual offset is |δfi (t)| , ∆f with useful-signal and ICI fractions ϵi (t) =

(14)

Ssig,i (t) = sinc2 (ϵi (t)) ,

(15)

IICI,i (t) = 1 − Ssig,i (t).

(16)

Let γi (t) denote the nominal singleton-link SINR in linear scale. The effective SINR after the residual-ICI penalty is  −1 IICI,i (t) γieff (t) = γi−1 (t) + . (17) Ssig,i (t) The effective SINR is mapped to an MCS-dependent spectral efficiency ηiSE (t) and BLER. A bounded HARQ abstraction increases the effective decoding SINR across retransmission rounds until either the target BLER or maximum HARQ-round limit is reached. If Nr,i (t) denotes the resulting number of rounds, the singleton service opportunity is ηiSE (t)BUL ∆tηd , (18) Nr,i (t) where BUL is the occupied uplink bandwidth and ∆t = 0.5 s is a MARL control epoch rather than a physical NR slot, thereby aggregating multiple NR transmission opportunities. The factor ηd captures propagation-delay efficiency. HARQ is abstracted through the effective round count Nr,i (t), which reduces the service in (18). Hence, the model captures retransmission-induced throughput loss without explicitly simulating slot-level ACK/NACK timing or HARQ Ci (t) =

feedback latency. The final decoding outcome χi (t) is drawn from the residual BLER. D. Cooperative Partially Observable MARL Formulation The decentralized access problem is modeled as a cooperative partially observable stochastic game. At decision epoch t, the latent global state is n oU ei (t), Wi (t), γieff (t) X(t) = B ∪ {o(t − 1), G(t)} , (19) i=1

where Wi (t) is the service waiting age and G(t) captures the time-varying orbital and propagation geometry. Each UE observes only its local state si (t) and selects an action independently, yielding a(t) = [a1 (t), . . . , aU (t)].

(20)

The transition X(t + 1) ∼ P(· | X(t), a(t)) (21) is driven by traffic arrivals, queue service, satellite/UE motion, fading, residual Doppler error, and contention outcomes. All UEs share a common DQN parameterization. Let πθ (ai |si ) denote the shared decentralized policy. The objective is # "∞ U X X 1 t ri (t) , (22) θ ⋆ = arg max E γRL θ U i=1 t=0 where γRL ∈ (0, 1) is the RL discount factor; reward averaging avoids trivial scaling with the UE population. Each UE has partial information and interacts with others through the shared collision domain. In A2A-IRIS, compact coordination information is exchanged over an idealized collision-free UE-to-UE sidelink before the D2S access decision. Reliable delivery within the ∆t = 0.5 s control epoch is assumed, while sidelink PHY/MAC resource allocation, packet loss, interference, and retransmissions are abstracted. Delayed peer states use the preceding epoch, whereas current contention and payload-priority features are available before the present {NT, T} decision. IV. P ROPOSED IRIS F RAMEWORK As illustrated in Fig. 2, we develop a parameter-shared deepRL framework for decentralized LEO D2S random access. A2A-IRIS is the sidelink-assisted realization, while VanillaIRIS serves as its local-only ablation. Both use the same binary action space {NT, T}, reward, collision model, and training procedure, differing only in the availability of peerassisted coordination. All UEs share the action-value function Qθ : S × A → R, (23) with each UE selecting its action from its own observation. A. Local Observation and Sidelink-Assisted Peer Context Both configurations share the five-dimensional local observation  T o(t − 1) , ξi (t) , (24) xi (t) = B̄i (t), γ̄i (t), W̄i (t), 2

where the normalized decision-time queue occupancy and effective-SINR features are ! ei (t) B B̄i (t) = clip , 0, 1 , (25) Bmax and

  eff γi (t) − γmin , 0, 1 , (26) γ̄i (t) = clip γmax − γmin respectively. The waiting age is mapped smoothly to the unit interval as   Wi (t) W̄i (t) = 1 − exp − , (27) τW where τW controls the saturation rate. The final local feature is a stochastic contention token. UE i draws ui (t) ∼ U(0, 1), and forms log(1 + U ui (t)) . (28) log(1 + Umax ) where Umax is the maximum supported UE population used for normalization. The token does not directly determine access; it provides a private symmetry-breaking variable for UEs with otherwise similar local observations. ξi (t) =

1) Compact Delayed Peer-State Representation For sidelink-assisted coordination, each UE quantizes its buffer, channel, and waiting states into three levels, bin bi (t), cbin i (t), wi (t) ∈ {1, 2, 3}.

(29)

These states are packed into a 27-level code,

bin pei (t) = 9(bi (t) − 1) + 3 cbin i (t) − 1 + wi (t) − 1 + 1, Peer-state information is delayed by one decision epoch, so the observation at t uses messages generated at t−1. Let Vi (t− 1) denote the peers with valid delayed codes. The codes are compressed into a fixed-dimensional normalized histogram:

hi,k (t) =

 X  1{e pj (t − 1) = k}    j∈V (t−1) i

   0,

|Vi (t − 1)|

,

|Vi (t − 1)| > 0,

k = 1, . . . , 27,

|Vi (t − 1)| = 0,

(30)

with hi (t) = [hi,1 (t), . . . , hi,27 (t)]T .

(31)

Thus, the representation preserves the empirical distribution of peer states while remaining invariant to UE ordering and independent of U . 2) Current Contention and Payload Ranks In addition to the delayed peer histogram, A2A-IRIS uses two current relative-priority features. At the beginning of decision epoch t, each backlogged UE shares its contention token b i (t) through the ξi (t) and predicted queue-limited payload D b i (t) denotes the sidelink before the TX/NT decision. Here, D expected queue-limited payload computed from the current local PHY estimate, residual BLER, and available backlog. Define n o ej (t) > 0 . AB (t) = j : B (32)

Fig. 2. IRIS framework for decentralized LEO D2S uplink scheduling, showing observation construction, parameter-shared DQN access decisions, sharedchannel outcomes, centralized training with replay and target updates, and decentralized greedy inference using a frozen policy.

For i ∈ AB (t), the corresponding percentile ranks are P j∈AB (t) 1{ξj (t) ≤ ξi (t)} ξ ρi (t) = , |AB (t)| n o P b j (t) ≤ D b i (t) 1 D j∈AB (t) . ρD i (t) = |AB (t)|

(33)

(34)

For an empty-buffer UE, both ranks are set to zero. Unlike the one-epoch-delayed peer histogram, these ranks use coordination information available within the current control epoch. They provide relative priority only and do not impose a deterministic transmitter selection rule. The complete sidelink-assisted observation is h iT ξ T T D sSL . (35) i (t) = xi (t), hi (t), ρi (t), ρi (t) Its dimension therefore remains independent of the UE population. B. Parameter-Shared Deep Q-Learning For observation si (t), the shared DQN produces the two possible actions qi (t) = [Qθ (si (t), NT) , Qθ (si (t), T)] . (36) The network consists of a feed-forward input layer, two fully connected hidden layers with nonlinear activation, and a two-dimensional output layer. The layer widths and training hyperparameters are provided in Section V-A. For a replay-buffer transition tuple (s, a, r, s′ , d), where d ∈ {0, 1} is the terminal-state indicator, the target network Qθ− computes the one-step DQN target as y = r + γRL (1 − d) max Qθ− (s′ , a′ ). a′ ∈A The online network minimizes h i 2 L (θ) = E (Qθ (s, a) − y) .

(37) (38)

A separate target network is periodically synchronized with the online network to reduce instability caused by simultaneously changing bootstrapping targets. 1) Event-Preserving Replay The one-winner contention model naturally produces substantially more NT decisions than T decisions. Uniformly storing and sampling all local observations would therefore allow routine silence events to dominate the replay distribution. The replay procedure preserves every active-UE T transition and subsamples only redundant active-NT transitions when required by the per-epoch replay budget. Furthermore, whenever both action classes are present in replay memory, DQN minibatches are sampled approximately equally from T and NT transitions. The same replay rule is used for both the sidelink-assisted and local-only configurations. A nonterminal experience from epoch t is not inserted until the true decision state of epoch t + 1 has been generated. Hence, s(t + 1) contains the realized next-epoch arrivals, fresh channel observation, updated one-epoch-delayed peer information, and a new contention-token realization. The replay tuple therefore preserves the Bellman transition (s(t), a(t), r(t), s(t + 1)) (39) rather than constructing an artificial next state from information belonging to epoch t. C. Load-Aware Exploration Applying conventional independent ε-greedy exploration to every UE causes the amount of random access to increase with population size. In a large active population, this can create excessive exploration-induced collisions even when the learned policy itself is well coordinated. To avoid this behavior, ε(t) is interpreted as a epoch-level exploration intensity. Let

Nactive (t) = |AB (t)|.

(40)

Each active UE independently explores with probability   2ε(t) pexp (t) = min 1, . (41) Nactive (t) An exploring UE selects uniformly from {NT, T}. When the clipping in (41) is inactive, E [NTexp (t)] = Nactive (t)pexp (t) Pr{T | explore} 2ε(t) 1 = Nactive (t) (42) Nactive (t) 2 = ε(t). Thus, the expected number of random exploratory transmissions remains approximately independent of the UE population rather than growing linearly with the number of active terminals. D. Cooperative Anti-Capture Reward Design Avoiding collisions alone is insufficient, as repeatedly serving the same UEs can leave others backlogged, reducing goodput and fairness. The reward therefore combines individual and cooperative components: ri (t) = riind (t) + 1 {i ∈ AB (t)} ωteam Rteam (t),

(43)

where riind (t) is the UE-specific reward and Rteam (t) is a common access-outcome term applied to all backlogged UEs. The same reward design is used for both configurations. 1) Successful-Service Utility Suppose UE i⋆ is the unique transmitter and is successfully decoded. Its normalized delivered-goodput utility is   Di⋆ (t) gi⋆ (t) = log 1 + , Dtar = λ∆t, (44) Dtar where normalization by Dtar measures service relative to one epoch of offered traffic, while the logarithm limits sensitivity to large service events. The normalized pre-service backlog and waiting age are ! ei⋆ (t) B qi⋆ (t) = clip , 0, 1 , (45) Bmax and



 Wi⋆ (t) w (t) = 1 − exp − . (46) τW To discourage service capture, the successful UE is weighted by its relative backlog and waiting-age urgency among backlogged UEs:X n o 1 Bej (t) ≤ Bei⋆ (t) i⋆

ρB (i⋆ , t) = and

j∈AB (t)

|AB (t)| X

ρW (i⋆ , t) =

,

|AB (t)| The corresponding anti-capture urgency factor is p ui⋆ (t) = ρB (i⋆ , t)ρW (i⋆ , t).

.

where ωg , ωq , and ωw weight delivered goodput, queue pressure, and waiting age, respectively. 2) UE-Specific Access Reward Let S(t) denote a successfully decoded singleton transmission, and let T (t) denote the set of UEs selecting the transmit action at decision epoch t. The UE-specific reward is   Risucc (t), i = i⋆ , S(t), ⋆      −Rcol , i ∈ T (t), NT (t) > 1,      −Rfail , i = i⋆ , NT (t) = 1, ¬S(t), ind ri (t) = (51)  i ∈ AB (t), NT (t) = 0,  −Ridle ,   i̸=i⋆ , i∈AB (t),  −Rstarve , S(t),  Wi (t)>Wi⋆ (t),     0, otherwise. Thus, collisions, unsuccessful singleton attempts, and idle epochs with backlog are penalized. After a successful singleton transmission, a backlogged silent UE is penalized only if it has waited longer than the successful UE; otherwise, silence is neutral. 3) Cooperative Access-Outcome Reward The individual reward is complemented by a common team term:  succ  , S(t), Rteam     −Rcol , NT (t) > 1, team Rteam (t) = idle  −Rteam , NT (t) = 0 and |AB (t)| > 0,     −Rfail , N (t) = 1 and ¬S(t). T team (52) The same Rteam (t) is assigned to all backlogged UEs during centralized training, encouraging singleton access while preserving queue and waiting-aware credit assignment. This global access-outcome signal is used only for reward construction during training and is not required during decentralized inference; each UE still independently selects NT or T from its local observation.

(47)

E. Greedy Decentralized Inference After training, the DQN parameters are frozen and no replay insertion, Bellman update, or gradient calculation is performed. No team-reward or global access-outcome signal is required at inference. UE i selects

(48)

ai (t) = arg max Qθ (si (t), a) ,

(49)

while an empty-buffer UE is forced to choose NT. Although the DQN decision is greedy for a fixed observation, the contention token ξi (t) continues to be randomly

1 {Wj (t) ≤ Wi⋆ (t)}

j∈AB (t)

Thus, full success credit is favored when the transmitting UE is relatively urgent in both backlog and waiting age, limiting repeated service to recently served UEs with small fresh arrivals. The resulting success utility h is succ ⋆ Ri⋆ (t) = ui (t) ωg gi⋆ (t) + ωq qi⋆ (t) i (50) + ωw wi⋆ (t) ,

a∈Ai

(53)

regenerated. Consequently, the decentralized system retains stochastic symmetry-breaking during greedy inference. During inference, the local-only configuration uses only the local observation xi (t), whereas the sidelink-assisted configuration augments the local information with the delayed peer-state representation and coordination features defined in (35). No deterministic top-rank access rule is applied; each UE independently evaluates its learned action-value function using its available observation. V. E XPERIMENTS A. Experiment Settings The proposed MARL framework is evaluated using a population-driven LEO D2S uplink simulator incorporating orbital dynamics, location-dependent propagation, pedestrian UE mobility, finite-buffer traffic, and decentralized sharedchannel access, consistent with Section III. Satellite positions and velocities are obtained from CelesTrak TLEs and propagated using SGP4, with the catalog filtered to the considered 500–600 km LEO shell. A 72 × 22 Walker-Delta constellation at 550 km is retained as a fallback when TLE information is unavailable. The receive beam of the satellite is directed toward the service-region center. Two environments are considered: Suburban Erding, Germany, and Rural Mistelbach District, Austria. Their similar representative latitudes limit latitude-induced visibility differences while providing distinct population densities and propagation conditions. For the main evaluation, a 5 × 5 km2 service region is used. The active UE population follows (2), assuming one candidate handheld per resident and a 1.5% activity factor, yielding 61 Suburban and 22 Rural UEs. Thus, the contention load is determined by the deployment environment rather than selected independently. For the population-scalability study in Section VI-G, the service area and propagation environment are fixed while the number of simultaneously active UEs is varied to isolate the effect of increasing contention load. A separate train-and-test run is performed for each environment while retaining the same MARL architecture and access model. Both observation configurations have fixed input dimensions independent of U : five features for VanillaIRIS and 34 features for A2A-IRIS. Table I summarizes the location-dependent parameters, while Table II reports the common orbital, PHY, traffic, and learning settings. B. Propagation Environments The suburban and rural scenarios use the same channel model from Section III-C, differing only in environmentspecific parameters. The simulator follows the corresponding 3GPP TR 38.811 S-band Suburban/Rural configuration [28] for elevation-dependent LOS probability, LOS/NLOS shadow fading, and NLOS clutter loss, using the nearest reference elevation in {10◦ , 20◦ , . . . , 90◦ }. LOS links use elevation-dependent Rician fading with the corresponding NTN-TDL multipath profile, while NLOS links use Rayleigh fading with the NLOS profile. Free-space and

Table I: Location-Specific Scenario Parameters Parameter

Suburban

Reference location Representative latitude Representative longitude Population density Service-area population Active UE density Active UEs, U Shadow-correlation time Rain-rate parameter, R0.01 Rain variability, σln R Rician K offset Rician K elevation slope Minimum Rician K

Rural

Erding, Germany Mistelbach District, Austria 48.31◦ N 48.57◦ N 11.91◦ E 16.58◦ E 162.7 persons/km2 [29] 59.6 persons/km2 [30] ≈ 4068 ≈ 1490 2.44 UEs/km2 0.88 UEs/km2 61 22 20 s 40 s 80 mm/h 50 mm/h 1.26 1.10 5 dB 8 dB 0.20 0.25 3 dB 5 dB

Table II: Key Simulation and Learning Parameters. Orbital and Scenario

UE and Satellite RF

Parameter

Value

Parameter

Value

Primary geometry TLE group / altitude

TLE + SGP4 STARLINK / 500–600 km 1584

Duplex mode Carrier frequency, fc

FDD UL 2 GHz

UE transmit power

23 dBm

Walker, 72 × 22, 550 km, 53◦ 5 × 5 km2 1 UE/person

UE antenna gain

0 dBi

UE system loss UE EIRP

0 dB 23 dBm

1.5% 3 km/h 5◦

Satellite Rx gain Aperture diameter Satellite G/T

30 dBi 2m 1.1 dB/K

Max. propagated satellites Fallback geometry

Service area Candidate mapping, ηcand Activity factor, ηact UE mobility Min. elevation, θmin

Doppler and PHY

Traffic and Access

Parameter

Value

Parameter

Value

Doppler pre-compensation Residual ICI penalty

Enabled

Decision epoch, ∆t

0.5 s

Enabled

100 kbps 50 kbit/UE 4 Mbit Shared UL {NT, T} NT > 1

Velocity-error std. GNSS position error Frequency quantization Doppler-error correlation time Nominal carrier bandwidth Shared UL BW, BUL Subcarrier spacing, ∆f UL PRBs Max. HARQ attempts Target residual BLER Min. MCS SINR

0.01 m/s 3m 10 Hz 60 s

Offered UL traffic per UE, λ Traffic per epoch Buffer capacity, Bmax Contention domain Action space

10 MHz

Collision condition

Parameter

Value

Parameter

Value

Training episodes Hidden layers SL-assisted input dim. Replay capacity Minibatch size

10000 128, 128 34 105 256

Epochs per episode Local input dim. Learning rate Min. replay size Target update

500 5 10−3 2000 500 steps

8.64 MHz 30 kHz 24 4 0.10 −6 dB Training Parameters

atmospheric losses, rain attenuation, satellite receive-beam gain, thermal noise, and residual Doppler-induced ICI are also included. GNSS-aided Doppler pre-compensation is assumed; hence, the full geometry-induced Doppler is not treated as residual carrier offset. Instead, the remaining frequency error is modeled as a temporally correlated process accounting for

GNSS position uncertainty, satellite velocity uncertainty, and frequency quantization. The access model, action space, traffic process, decision interval, MCS/BLER mapping, and collision rule are identical across both scenarios. Thus, their performance differences arise from population-driven contention and environmentdependent propagation rather than algorithmic changes. C. Training and Evaluation Configuration The proposed MARL framework is instantiated in two configurations: Vanilla-IRIS and A2A-IRIS. Both use parametershared DQNs with the same architecture, action space, optimizer, reward, exploration, replay mechanism, and training budget. Vanilla-IRIS uses the features in Section IV-A, whereas A2A-IRIS additionally incorporates the one-epochdelayed peer-state histogram and current contention and payload-priority ranks. Load-aware ε-greedy exploration follows Section IV-C. Event-preserving replay follows Section IV-B, retaining all active-UE T transitions while subsampling redundant NT transitions. Nonterminal transitions are stored only after the realized next decision state is available. Steady-state performance is evaluated over 600 decision epochs using the training traffic model. A separate inferenceonly buffer-drain test initializes each UE with a 4-Mbit backlog, disables new arrivals, and runs the frozen DQNs greedily (ε = 0) for 15 epochs without replay or parameter updates. D. Upper Bound and Reference Methods The two IRIS configurations, A2A-IRIS and Vanilla-IRIS, are evaluated against an ideal upper bound (UB), three centralized channel-quality-based schedulers (D0, D1, and D2), and two conventional random-access schemes. A2A-IRIS and VanillaIRIS use the same parameter-shared MARL framework and differ only in sidelink-assisted peer context, isolating the benefit of peer-assisted coordination. • Upper-Bound Oracle (UB): The UB is a centralized current-CSI oracle that selects, in each epoch, the UE with the largest queue-limited deliverable payload, ) (
SE eff η γ (t) B ∆t η UL d i i UB ei (t), . (54) Ci (t) = min B Nr,i (t) ei (t) is the post-arrival UE backlog, η SE (γ eff (t)) is Here, B i i the SINR-dependent spectral efficiency, and Nr,i (t) is the required HARQ-round count; BUL , ∆t, and ηd denote the uplink bandwidth, epoch duration, and delay-efficiency factor, respectively. The oracle assumes collision-free access and no residual PHY failure, and its throughput is P maxi CiUB (t) , (55) TUB = t T ∆t where T is the number of evaluation epochs. • Centralized Channel-Quality-Based Schedulers (D0, D1, and D2): These collision-free baselines use centralized singleUE scheduling and differ only in channel-quality age. For

delay d ∈ {0, 1, 2}, the estimated service opportunity of backlogged UE i is
ηiSE γieff (t − d) BUL ∆t ηd (d) b Ci (t) = , (56)
br,i γ eff (t − d) N i with b (d) (t). i⋆d (t) = arg max C (57) i ei (t)>0 i:B

D0 uses current effective SINR, while D1 and D2 use oneand two-epoch-old effective SINR. Transmission is then evaluated using the current PHY state, including MCS, HARQ, BLER, and payload. Thus, D0 is a current-SINR centralized reference, whereas D1/D2 capture stale channel-quality loss. D0 remains distinct from the UB because it retains stochastic decoding and queue/HARQ effects. • p⋆ -ALOHA: Each backlogged UE independently transmits with probability 1 p⋆ = , (58) Nactive where Nactive is the number of non-empty UEs. This provides a load-aware uncoordinated reference [31], [32]. • BEB: Binary exponential backoff (BEB) [33] uses a contention window (CW) that determines the range from which each UE draws a random backoff counter, with CWmin = 4 and CWmax = 256. A UE transmits when its counter expires; collisions double the CW up to CWmax , while success resets it to CWmin . • Vanilla-IRIS (Ablation): Vanilla-IRIS uses the same twolayer DQN, binary action space, reward, and training budget as A2A-IRIS, but observes only five local features: normalized backlog, current effective SINR, waiting age, previous outcome, and the UE-specific coordination token. It excludes the peer-state histogram and relative coordination ranks available to A2A-IRIS. All schemes use the same satellite geometry, traffic and finite-buffer models, propagation assumptions, residualDoppler treatment, MCS/BLER mapping, and HARQ abstraction. D0–D2 use centralized collision-free selection, while p⋆ ALOHA, BEB, Vanilla-IRIS, and A2A-IRIS share the same contention-domain MAC model. E. Performance Metrics The evaluation uses complementary metrics covering access efficiency, reliability, fairness, resource cost, and computational overhead. • Effective Throughput/Goodput: The aggregate uplink PT PU goodput is ηsys = T 1∆t t=1 i=1 Di (t), where Di (t) is the successfully delivered payload, T is the number of evaluation epochs, and ∆t is the epoch duration. • Access Outcome Rates: The success rate is the fraction of epochs with a successfully decoded singleton transmission, while the collision rate is the fraction with multiple simultaneous transmissions. Idle and singleton PHY-failure rates are also reported. • Fairness: Inter-user fairness is measured using Jain’s index PU 2 ( P i=1 ηi ) [34], J = U U η2 , where ηi is the delivered throughput i=1

i

VI. E XPERIMENT R ESULTS We evaluate IRIS against the baselines in Section V-D using the setup of Section V-A. After training, the DQN policies are frozen and evaluated greedily (ϵ = 0), with no further adaptation. Suburban and Rural scenarios are evaluated separately to preserve the effects of their distinct UE densities and propagation conditions. A2A-IRIS adds compact peer context and current contention and payload-priority information, while all other learning, traffic, MAC, reward, and PHY settings remain identical, thereby isolating the benefit of sidelinkassisted A2A coordination. A. Throughput Comparison As shown in Fig. 3, A2A-IRIS achieves 2.1579 Mbps and 5.4478 Mbps in the Rural and Suburban scenarios, corre-

D1 D2

Throughput (Mbps)

BEB p*-ALOHA

UB

UB = 5.668 Mbps

6

96.1%

5 4 3

42.4%

UB = 2.161 Mbps

99.9%

2

70.9%

65.8%

66.0%

41.7%

22.6%

19.1%

41.8%

1

12.2%

24.5%

0

Rural

Suburban

Fig. 3. Effective throughput of the considered schemes in the Rural and Suburban scenarios. The dashed lines denote the scenario-specific upper bounds, while the percentages above the bars indicate the fraction of the corresponding UB achieved by each scheme. BEB

p*-ALOHA

0.869

1.0

Vanilla-IRIS

0.924

A2A-IRIS

0.8

0.267

0.238

0.300

0.277

0.4

0.256

0.6 0.308

F. Practicality of Real-World Evaluation Real-world evaluation of decentralized LEO D2S uplink access requires authorized access to an operational satellite link, compatible UEs, and coordination with the associated satellite and ground infrastructure. Such access is generally unavailable for controlled many-UE experiments, particularly when the underlying radio-access decisions must be modified or observed. Accordingly, IRIS is evaluated using a system-level simulation platform that combines TLE/SGP4 orbital dynamics, population-driven traffic, 3GPP TR 38.811-aligned Sband propagation [28], pedestrian UE mobility, GNSS-aided Doppler pre-compensation, residual ICI, MCS-dependent stochastic BLER, finite-buffer dynamics, and bounded HARQ. This enables controlled and reproducible evaluation under time-varying LEO geometry, contention, queueing, and PHY conditions. Operational LEO systems such as SpaceX Starlink and Amazon Leo rely on operator-controlled radio-access infrastructure, limiting access for external experimental modification. The simulation therefore captures the D2S propagation, mobility, queueing, contention, and scheduling characteristics relevant to the proposed access problem within the systemlevel abstraction of Section III.

A2A-IRIS Vanilla-IRIS

Successful-epoch rate

of UE i. Per-UE throughput and served-UE statistics are additionally used to assess access capture. • Energy and Transmission-Resource Cost: RF efficiency is measured in joules per delivered Mbit, together with transmission/HARQ round-equivalents per delivered Mbit. These capture the cost of failed attempts and retransmissions beyond aggregate goodput. • Computational Complexity: Policy overhead is characterized by trainable DQN parameters, forward-pass MAC operations per UE decision, and measured inference time. Because the DQN is shared across UEs, parameter count is independent of U , while aggregate neural inference cost scales linearly with U . Queue saturation, transmission attempts, and served-UE fraction are reported as diagnostic statistics where needed for interpretation.

0.2

0.0

Rural

Suburban

Fig. 4. Successful-epoch rate of the decentralized access schemes under the Rural and Suburban scenarios.

sponding to 99.9% and 96.1% of the respective UBs. In comparison, D1 and D2 achieve 65.8–66.0% of the UB in Rural and 41.7–42.4% in Suburban. A2A-IRIS improves throughput over D1 and D2 by 51.7% and 51.3% in Rural, and 126.7% and 130.4% in Suburban, respectively. Without peer-state information, Vanilla-IRIS achieves only 1.5326 Mbps in Rural and 1.2818 Mbps in Suburban, corresponding to A2A-IRIS gains of 40.8% and 325.0%. A2A-IRIS also substantially outperforms p⋆ -ALOHA and BEB, demonstrating the benefit of compact peer context for decentralized UE-side access. B. Access Outcomes and Coordination Behavior Figs. 4 and 5 further illustrate the access behavior of the decentralized schemes. A2A-IRIS achieves successful-epoch rates of 0.9244 and 0.8686 in the Rural and Suburban scenarios, respectively, while maintaining collision rates of 0.0000 and 0.0433. In comparison, Vanilla-IRIS exhibits substantially higher collision rates of 0.2735 and 0.1678, demonstrating the role of sidelink information in coordinating distributed transmission decisions. C. Fairness, Queueing, and Persistent UE Capture Fig. 6 compares the seed-wise Jain fairness of the evaluated schemes. A2A-IRIS achieves the highest fairness in both environments, reaching 0.9999 in Rural and 0.9954 in Suburban. Vanilla-IRIS obtains 0.9789 and 0.8338, while D1 and D2 achieve approximately 0.90–0.91 in Rural and 0.79 in Suburban. These results show that A2A-IRIS provides more balanced service across UEs, with a larger fairness advantage in the denser Suburban scenario.

p*-ALOHA

0.168

0.258

0.265

Vanilla-IRIS

D1

D2

0.995

0.979

D=0

D=1

0.788

A2A-IRIS Vanilla-IRIS

Vanilla-IRIS

(b) UE service coverage

0.687

0.6 0.4 0.236

0.2 0.000

Rural

Served-UE fraction

0.8

1.000

1.000

1.000

0.982

0.8 0.6 0.4 0.2

0.052

Suburban

0.0

Rural

p*-ALOHA

Suburban

Fig. 7. Queueing performance and UE service coverage of A2A-IRIS and Vanilla-IRIS in Rural and Suburban scenarios. Panel (a) shows the queuesaturation fraction, while panel (b) shows the served-UE fraction; lower saturation indicates reduced backlog accumulation, whereas higher coverage indicates broader UE service.

The complementary queueing and service-coverage results in Fig. 7 further illustrate this behavior. In Rural, A2A-IRIS reduces queue saturation from 0.236 to 0.000, while both schemes achieve a served-UE fraction of 1.000. In Suburban, the served-UE fraction increases from 0.982 to 1.000, while queue saturation decreases from 0.687 to 0.052. Overall, A2A-IRIS maintains broader UE coverage in Suburban while substantially reducing queue buildup in both environments. D. Sensitivity to Delayed CSI Fig. 8 shows the impact of CSI aging on the centralized scheduler. The D = 0 case is the instantaneous-CSI reference, while D = 1 and D = 2 represent stale-CSI operation. In the Rural scenario, throughput decreases from 1.7293 Mbps at D = 0 to 1.4222 Mbps and 1.4266 Mbps at D = 1 and D = 2, respectively. Similarly, in the Suburban scenario, it drops from 3.2979 Mbps to 2.4028 Mbps and 2.3643 Mbps. In

0.22

0.0

0.04

0.2 0.1

Suburban

Fig. 6. Jain’s fairness index of A2A-IRIS, Vanilla-IRIS, and the delayed centralized references under the Rural and Suburban scenarios. The reported values are obtained by computing Jain’s index for each evaluation seed and then averaging across seeds.

1.0

0.3

0.11

0.2

(a) Queue saturation

BEB

0.4

0.11

0.4

1.0

D1 D2

0.6 0.5

A2A-IRIS

D=2

Delay d (decision epochs) (b) Suburban

Fig. 8. Impact of CSI delay on the centralized scheduler in the Rural and Suburban scenarios. The D = 0 case denotes the instantaneous-CSI centralized reference, while D = 1 and D = 2 represent stale-CSI operation. Horizontal lines denote the UB and the decentralized access schemes.

0.789

0.6

Rural

0

D=2

Delay d (decision epochs) (a) Rural

RF energy (J/Mbit)

Jain's Fairness Index

1

0.906

0.8

0.0

Queue-saturation fraction

2

0.4

0.52

0.902

0.834

0.0

3

0.8

0.44

1.0

4

1.2

0.18

A2A-IRIS 1.000

D=1

5

1.6

Suburban

Fig. 5. Collision rate of the decentralized access schemes under the Rural and Suburban scenarios.

D=0

6

0.0

Rural

p *-ALOHA

2.0

0.13

0.0

BEB

Vanilla-IRIS

2.4

0.09

0.1

0.043

0.2

A2A-IRIS

UB

Rural

0.07

0.299

0.274

0.3

0.000

Collision rate

0.4

Throughput (Mbps)

0.5

Centralized CSI-based scheduler

0.07

BEB

0.11

Vanilla-IRIS

0.468

A2A-IRIS

Suburban

Fig. 9. RF transmission energy cost per successfully delivered Mbit under the Rural and Suburban scenarios. The higher cost of Vanilla-IRIS results from collision-induced wasted transmissions and subsequent reattempts of undelivered traffic.

contrast, A2A-IRIS achieves 2.1579 Mbps and 5.4478 Mbps in the Rural and Suburban scenarios, respectively, approaching the corresponding UB while outperforming the centralized delayed-CSI cases. E. Energy and Transmission-Resource Efficiency The efficiency results in Figs. 9 and 10 show that the throughput improvement of A2A-IRIS is not obtained through excessive transmission activity. A2A-IRIS requires only 0.091 and 0.037 J/Mbit in the Rural and Suburban scenarios, respectively, with corresponding transmission-resource costs of 0.913 and 0.372 transmission/HARQ round-equivalents per Mbit. In contrast, the substantially higher energy cost of Vanilla-IRIS is primarily caused by uncoordinated simultaneous transmissions. Collided transmissions consume RF energy without delivering payload, while the undelivered traffic remains queued and must be attempted again in subsequent epochs. Consequently, the combination of wasted collision energy and repeated transmission attempts increases both the energy and transmission-resource cost per successfully delivered Mbit. These results indicate that sidelink-assisted coordination simultaneously improves effective throughput and reduces unsuccessful transmission overhead. F. Computational Complexity Fig. 11 compares the inference complexity of A2A-IRIS and Vanilla-IRIS. Using the same two-layer DQN backbone, A2A-IRIS processes a richer sidelink-assisted observation and therefore requires 21,250 trainable parameters and 20,992 MACs per UE, compared with 17,538 parameters and 17,280

A2A-IRIS Vanilla-IRIS

D1 D2

BEB

A2A-IRIS

p*-ALOHA

5.22

0

Rural

0.68

0.68

1.10

1.83

2.22 1.10

1.11

0.37

1

1.33

2

20,992

17,538

17,538

15000 10000

17,280

17,280

15000 10000 5000

5000 0

20000

20,992

Rural

Suburban

0

Rural

Suburban

(c) Policy inference latency 40

Inference time (μs/UE)

(b) Forward-pass MACs

21,250

MACs per UE

Number of parameters

20000

Vanilla-IRIS

(a) DQN parameters

36.37 32.39

30

20 11.62 10.29

10

0

Rural

5 4 3 2 1 10

20

30

40

80

Number of active UEs

Fig. 10. Normalized transmission-resource cost, measured in transmission/HARQ round-equivalents per successfully delivered Mbit, under the Rural and Suburban scenarios. The larger cost of Vanilla-IRIS reflects the additional transmission attempts caused primarily by uncoordinated collisions.

21,250

D1

6

0

Suburban

A2A-IRIS

Throughput (Mbps)

4 3

Vanilla-IRIS

7

4.39

5

0.91

TX/HARQ rounds per Mbit

6

Suburban

Fig. 11. Computational complexity of A2A-IRIS and Vanilla-IRIS under the Rural and Suburban scenarios. The comparison includes the DQN parameter count, forward-pass MAC operations per UE, and measured policy inference latency per UE.

MACs for Vanilla-IRIS, corresponding to increases of approximately 21.2% and 21.5%, respectively. Despite this overhead, inference latency remains on the order of tens of microseconds per UE. A2A-IRIS additionally requires only lightweight histogram aggregation and rank computation, without iterative optimization or matrix inversion. Because the DQN input dimension is fixed, the per-UE forward-pass complexity remains O(1) with respect to the active UE population U , while the aggregate neural inference cost scales as O(U ). G. Sensitivity to Active UE Load Fig. 12 evaluates IRIS under increasing active-UE load while keeping the service area and propagation environment fixed. The active population is varied as U ∈ {10, 20, 30, 40, 80, 100}, while the nominal operating points remain U = 22 for Rural and U = 61 for Suburban. Thus, the largest tested load corresponds to approximately 4.5× and 1.6× the nominal Rural and Suburban populations, respectively. Across the evaluated range, A2A-IRIS achieves higher throughput than Vanilla-IRIS and D1, with throughput increasing as the contention population grows. Vanilla-IRIS and D1 exhibit weaker scaling due to the lack of peer awareness and stale channel information, respectively. These results highlight the potential of A2A-IRIS for populationdense deployments, where increasing active-UE populations create stronger contention over the shared uplink and make peer-assisted coordination increasingly beneficial. The sweep characterizes this behavior over the evaluated range rather than extrapolating beyond the tested populations.

100

10

20

30

40

80

100

Number of active UEs

Rural (b) Suburban Fig. 12. Fig. 12.(a) Active-UE load scalability of A2A-IRIS, Vanilla-IRIS, and D1 with fixed service area and propagation environment. The active UE population is evaluated at 10, 20, 30, 40, 80, and 100 UEs, while the nominal main-evaluation points remain 22 UEs for Rural and 61 UEs for Suburban.

H. Comparison with Existing NTN Random-Access Table III compares IRIS with representative NTN randomaccess approaches. Existing schemes mainly address contention, multi-user reception, activity detection, or cooperative reception, typically without explicit peer-state information at the transmitting UEs. In contrast, A2A-IRIS performs UE-side access decisions using compact sidelink-assisted peer context to reduce uncoordinated transmissions. The Vanilla-IRIS ablation isolates this contribution by retaining the same MARL framework without peer-assisted coordination. Thus, A2AIRIS is distinguished by integrating compact peer context with grant-free decentralized access control. I. Satellite Mobility and Handover Implications LEO orbital motion creates time-varying geometry that affects slant range, elevation, path loss, antenna gain, and Doppler. We capture these variations using CelesTrak TLE trajectories propagated with SGP4 when available, with a Walker constellation as fallback. For each UE, the serving satellite is selected among visible satellites using the highest reference received-power metric and may change as geometry evolves. Because the shared DQN is not satellite-specific, servingsatellite changes are handled through the evolving SINRrelated observations without policy reinitialization. The study therefore models dynamic link selection and PHY variation rather than full handover signaling. Handover overhead and beam-sweep effects are outside the scope. VII. C ONCLUSION This work presented IRIS, a parameter-shared MARL framework for decentralized grant-free LEO D2S uplink access, with A2A-IRIS as the sidelink-assisted realization and VanillaIRIS as its local-only ablation. A2A-IRIS achieves 2.1579 and 5.4478 Mbps in Rural and Suburban scenarios, corresponding to 99.9% and 96.1% of the scenario-specific UBs and gains of 40.8% and 325.0% over Vanilla-IRIS. It also outperforms one- and two-epoch delayed centralized schedulers by approximately 51–130%, while improving fairness, service coverage, and transmission efficiency with moderate computational overhead. Active-UE load experiments further show sustained throughput gains over the evaluated range. Overall, sidelinkassisted peer coordination improves decentralized LEO D2S contention resolution and resource efficiency. Future work will consider imperfect sidelink delivery and overhead, dynamic

Table III: Architectural Comparison of IRIS with Representative NTN Random-Access Approaches. Approach Lee et al. [23] Yu et al. [24] Shen et al. [17] Ying et al. [18] Vanilla-IRIS (Ablation) A2A-IRIS (Proposed)

Grant-Free Yes Yes Yes Yes Yes Yes

UE-Side Decision Yes Yes No No Yes Yes

Learning MARL DRL No No MARL (shared DQN) MARL (shared DQN)

handover and multi-satellite access, richer action spaces, and implementation-oriented validation. R EFERENCES [1] X. Lin, S. Rommer, S. Euler, E. A. Yavuz, and R. S. Karlsson, “5g from space: An overview of 3gpp non-terrestrial networks,” IEEE Communications Standards Magazine, vol. 5, no. 4, pp. 147–153, 2021. [2] H. B. Pasandi, J. A. Fraire, S. Ratnasamy, and H. Rivano, “A survey on direct-to-device satellite communications: Advances, challenges, and prospects,” in Proceedings of International Workshop on LEO Networking and Communication, 2024, p. 7–12. [3] Z. Xiao, J. Yang, T. Mao, C. Xu, R. Zhang, Z. Han, and X.-G. Xia, “Leo satellite access network (leo-san) toward 6g: Challenges and approaches,” IEEE Wireless Communications, vol. 31, no. 2, pp. 89– 96, 2024. [4] 3GPP, “Solutions for nr to support non-terrestrial networks (ntn),” 3GPP Technical Report TR 38.821, vol. 16, no. 0, Jan 2020, v16.0.0. [5] W. Liu, X. Hou, J. Wang, L. Chen, and S. Yoshioka, “Uplink time synchronization method and procedure in release-17 nr ntn,” in IEEE Vehicular Technology Conference, 2022, pp. 1–5. [6] 3GPP, “NR; NR and NG-RAN Overall description; Stage-2 (Release 17),” 3rd Generation Partnership Project (3GPP), Technical Specification (TS) 38.300, 2022. [7] J. Jeong and D. Hong, “Two-stage preamble detector for leo satellitebased ntn iot random access,” IEEE Transactions on Vehicular Technology, vol. 72, no. 11, pp. 14 443–14 455, 2023. [8] Y. Cho, W. Yang, D. Oh, and H.-S. Jo, “Multi-agent deep reinforcement learning for interference-aware channel allocation in non-terrestrial networks,” IEEE Communications Letters, vol. 27, no. 3, pp. 936–940, 2023. [9] Z. Ali, S. Lagén, L. Giupponi, and R. Rouil, “3gpp nr v2x mode 2: Overview, models and system-level evaluation,” IEEE Access, vol. 9, pp. 89 554–89 579, 2021. [10] R. Surapaneni, M. Jha, M. Vakoc, and T. Segal, “Announcing the agent2agent protocol (a2a),” https://developers.googleblog.com/en/a 2a- a- new- era- of- agent- interoperability/, 2025, Google Developers Blog, accessed Sep. 10, 2026. [11] T. S. Kelso, “CelesTrak,” https://celestrak.org/, 2026, NORAD two-line element (TLE) orbital data; accessed September 4, 2026. [12] D. A. Vallado, P. Crawford, R. Hujsak, and T. S. Kelso, “Revisiting Spacetrack Report #3,” in AIAA/AAS Astrodynamics Specialist Conference and Exhibit, Keystone, CO, USA, 2006, AIAA 2006-6753. [13] O. Kodheli, E. Lagunas, N. Maturo, S. K. Sharma, B. Shankar, J. F. M. Montoya, J. C. M. Duncan, D. Spano, S. Chatzinotas, S. Kisseleff, J. Querol, L. Lei, T. X. Vu, and G. Goussetis, “Satellite communications in the new space era: A survey and future challenges,” IEEE Communications Surveys & Tutorials, vol. 23, no. 1, pp. 70–109, 2021. [14] A. A. Noman, S. S. Reza, M. Yuksel, and D. Roy, “From orbit to device: Spectrum intelligence for multi-tier LEO-aided D2C IoT networks,” IEEE IoT Magazine, Sep 2026. [15] A. A. Noman, T. Tolga Sarı, S. Amatare, G. Secinti, and D. Roy, “Space-air-ground network for direct-to-cell communication,” in IEEE MILCOM, 2025, pp. 1–6. [16] H. Chougrani, S. Kisseleff, W. A. Martins, and S. Chatzinotas, “Nbiot random access for nonterrestrial networks: Preamble detection and uplink synchronization,” IEEE Internet of Things Journal, vol. 9, no. 16, pp. 14 913–14 927, 2022. [17] B. Shen, Y. Wu, J. An, C. Xing, L. Zhao, and W. Zhang, “Random access with massive mimo-otfs in leo satellite communications,” IEEE Journal on Selected Areas in Communications, vol. 40, no. 10, pp. 2865–2881, 2022.

Peer Context No No No Multi-satellite Local only Compact sidelink

Concurrent Access Handling Collision based NOMA/SIC Multi-user detection Cooperative reception Collision based Collision based

Primary Focus Decentralized RA LEO IoT RA Activity/channel estimation Quasi-synchronous RA Learned contention Peer-assisted contention

[18] K. Ying, Z. Gao, S. Chen, M. Zhou, D. Zheng, S. Chatzinotas, B. Ottersten, and H. V. Poor, “Quasi-synchronous random access for massive mimo-based leo satellite constellations,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 6, p. 1702–1722, 2023. [19] B. Shen, Y. Wu, S. Gong, H. Liu, B. Ottersten, and W. Zhang, “Massive mimo-otfs-based random access for cooperative leo satellite constellations,” IEEE Journal on Selected Areas in Communications, vol. 43, no. 1, p. 90–106, 2025. [20] C. Zhang, Y. Liu, J. Hu, and K. Yang, “Joint user identification, channel estimation, and data detection for grant-free noma in leo satellite communications,” IEEE Journal on Selected Areas in Communications, vol. 43, no. 1, pp. 107–121, 2025. [21] J. Jeong, H. Lee, and D. Hong, “Enhanced early preamble collision detector for leo satellite-based ntn iot random access,” IEEE Transactions on Vehicular Technology, vol. 74, no. 8, pp. 12 496–12 511, 2025. [22] N. Jiang, Y. Deng, A. Nallanathan, and J. Yuan, “A decoupled learning strategy for massive access optimization in cellular iot networks,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 3, pp. 668– 685, 2021. [23] J.-H. Lee, H. Seo, J. Park, M. Bennis, and Y.-C. Ko, “Learning emergent random access protocol for leo satellite networks,” IEEE Transactions on Wireless Communications, vol. 22, no. 1, pp. 257–269, 2023. [24] H. Yu, H. Zhao, Z. Fei, J. Wang, Z. Chen, and Y. Gong, “Deepreinforcement-learning-based noma-aided slotted aloha for leo satellite iot networks,” IEEE Internet of Things Journal, vol. 10, no. 20, pp. 17 772–17 784, 2023. [25] M. S. Oh, Z. Zhang, F. Hairi, A. Velasquez, and J. Liu, “Consensusbased decentralized multi-agent reinforcement learning for random access network optimization,” in Proceedings of ACM MobiHoc, ser. MobiHoc ’25, 2025, p. 241–250. [26] J. Ni, B. Tan, and R. Srikant, “Q-csma: Queue-length based csma/ca algorithms for achieving maximum throughput and low delay in wireless networks,” in 2010 Proceedings IEEE INFOCOM, 2010, pp. 1–5. [27] T. Rashid, M. Samvelyan, C. S. De Witt, G. Farquhar, J. Foerster, and S. Whiteson, “Monotonic value function factorisation for deep multiagent reinforcement learning,” J. Mach. Learn. Res., vol. 21, no. 1, 2020. [28] 3GPP, “Study on nr to support non-terrestrial networks,” 3GPP Technical Report TR 38.811, vol. 15, no. 4, Sep 2020, v15.4.0. [29] The Federal Returning Officer, Germany, “Structural constituency data: Erding,” https://www.bundeswahlleiterin.de/en/europawahlen/2024/str ukturdaten/bund-99/land-9/kreis-9177.html, 2024, accessed: 2026-0904. Population density: 162.7 inhabitants/km2 . [30] State of Lower Austria, “District statistics: Mistelbach,” https://www.no e.gv.at/noe/Zahlen-Fakten/Mistelbach.html, 2025, accessed: 2026-0904. Population: 76,986; area: 1,291.72 km2 ; population density: 59.6 inhabitants/km2 . [31] N. Abramson, “The aloha system: another alternative for computer communications,” in Proceedings of the Fall Joint Computer Conference. Association for Computing Machinery, 1970, p. 281–285. [32] R. T. B. Ma, V. Misra, and D. Rubenstein, “An analysis of generalized slotted-aloha protocols,” IEEE/ACM Transactions on Networking, vol. 17, no. 3, pp. 936–949, 2009. [33] G. Bianchi, “Performance analysis of the ieee 802.11 distributed coordination function,” IEEE Journal on Selected Areas in Communications, vol. 18, no. 3, pp. 535–547, 2000. [34] R. Jain, D. Chiu, and W. Hawe, A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer System. Eastern Research Laboratory, Digital Equipment Corporation, 1984.

Related documents

• Delay-Aware Large-Small Model Collaboration over LEO Satellite Networks#168284
• Monte Carlo Throughput Estimation in Dynamic LEO Satellite Networks#670306
• Monte Carlo Throughput Estimation in Dynamic LEO Satellite Networks#668241
• Adaptive Compressed Sensing Differential Privacy Federated Learning Based on Orbital Spatiotemporal Characteristics in Space-Air-Ground Networks.#1166
• Clockwork: Scheduling Compound AI Systems in Space Data Centers#1536724