RISnet: Scalable RIS Optimization and Networking

• RISnet is a family of neural network-based approaches that optimize the placement and phase configuration of reconfigurable intelligent surfaces across diverse wireless settings.
• RISnet methods leverage unsupervised and reinforcement-learning techniques with CNN-based feature extraction to achieve rapid, near real-time inference compared to iterative optimization.
• RISnet extensions address practical challenges such as partial CSI, mutual coupling, and network-level coordination, enhancing scalability and efficient RIS-assisted communications.

RISnet denotes several related constructs in the reconfigurable intelligent surface literature rather than a single universally fixed model. In the supplied corpus, the term first appears as the name of a CNN-based deep Q-network controller for joint RIS placement and phase-shift design in an uplink RIS-enhanced device-to-device underlay cellular network (Ji et al., 2020). In later work, RISNet more commonly denotes a dedicated, domain-knowledge-driven, scalable neural architecture for RIS phase optimization in multi-user downlink systems, typically trained in an unsupervised manner and coupled with analytical or hybrid BS precoding (Peng et al., 2022). Other papers use “RISnet” more broadly to describe a coordinated RIS-enabled network fabric spanning terrestrial, aerial, satellite, deep-space, or network-controlled deployment settings (Tekbıyık et al., 2020).

1. Terminological scope

Across the cited papers, RISnet is a paper-specific label whose meaning depends on the system model and optimization objective. The capitalization also varies between “RISnet” and “RISNet”. This suggests that the term functions less as a single standardized architecture name than as a family label for RIS-specific control or networking designs.

Paper	Meaning of RISnet	Setting
(Ji et al., 2020)	CNN-based deep Q-network controller	RIS-assisted D2D underlay cellular uplink
(Peng et al., 2022)	Dedicated scalable neural architecture for RIS optimization	RIS-assisted downlink broadcast channel
(Peng et al., 2023, Peng et al., 2024, Peng et al., 10 Aug 2025, Peng et al., 11 Aug 2025, Fang et al., 15 Apr 2026)	Extensions with partial CSI, mutual coupling, implicit channel estimation, NOMA, and RSMA	Large-scale multi-user downlink
(Tekbıyık et al., 2020, Zhao, 2024)	Broader coordinated RIS-enabled network fabric or deployment architecture	NTN/DSN and network-controlled cellular RIS systems

The narrowest and most technically explicit uses name a neural controller that outputs RIS configurations. The broader uses treat RISnet as an RIS-enabled communications fabric or regulation architecture. For arXiv readers, this distinction is important because the associated assumptions, objectives, and algorithmic primitives differ substantially.

2. RISnet as a reinforcement-learning controller for RIS-enhanced D2D

In “Reconfigurable Intelligent Surface Enhanced Device-to-Device Communications,” RISnet is the proposed CNN-based deep Q-network controller used to jointly optimize the placement and phase configuration of a reconfigurable intelligent surface in an uplink RIS-enhanced D2D network underlaying a cellular network (Ji et al., 2020). The network comprises $K$ cellular users transmitting to a base station and $I$ D2D pairs reusing the uplink resources of the cellular users, with the RIS containing $N$ passive reflecting elements. The optimization target is the total sum rate

$$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$

subject to minimum QoS/SINR requirements at the D2D receivers and the base station, phase constraints, and a variable RIS installation position.

The effective D2D and interference channels are modeled through direct and reflected components. With reflection matrix

$$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$

the effective D2D channel is written as a reflected term plus a LoS link, and the D2D and cellular SINRs are defined through the resulting cascaded channels and resource-reuse indicator $\rho_{k,i}$. The joint placement-and-phase problem is non-convex because of the mixed discrete and continuous action structure and the phase constraints.

RISnet casts that problem as deep reinforcement learning. The state is

$${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$

with cardinality

$|{\cal S}|=(2I+K+N+2).$

The action is

$a_t=[\Delta\boldsymbol\Theta;\boldsymbol s_{RIS}],$

where each $\Delta\theta_n\in\{-\delta,0,+\delta\}$ and $I$0, so the action space size is

$I$1

The reward is explicitly QoS-aware: if both QoS constraints are met, the reward is the sum rate $I$2; otherwise a failure reward is used so that the agent improves whichever side of the link violates the SINR requirement.

The neural approximation follows the standard DQN structure with evaluation and target networks, replay memory $I$3, the Bellman target

$I$4

and squared-error loss. What distinguishes RISnet from a plain DQN is the convolutional front end. The input state is arranged into a $I$5 matrix with user, RIS, and BS positions in the first row, RIS phase shifts in the second row, and

$I$6

A convolutional layer extracts features, followed by a flatten layer, a fully connected hidden layer with 256 ReLU units, and a linear output layer with $I$7 outputs. The first-layer parameter count is only

$I$8

rather than the much larger fully connected alternative.

The numerical study uses $I$9 in a $N$0 square, $N$1 candidate RIS locations, $N$2, $N$3 dB, $N$4 dB, noise power $N$5 dBm, path-loss parameter $N$6, discount factor $N$7, and an exploration factor that starts at $N$8 and increases by $N$9 per iteration until $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$0. Against conventional DQN, Q-learning, and random control, the proposed CNN-based DQN achieves the highest reward and faster convergence than conventional DQN, while Q-learning converges faster but to a lower reward; the random policy performs worst. The reported sum-rate curves also show that RIS-assisted schemes outperform no-RIS baselines, that increasing the number of RIS elements improves sum rate with saturation, and that the transmit-power dependence is non-monotone because stronger desired-signal terms coexist with stronger interference.

3. RISNet as a dedicated scalable neural architecture for RIS optimization

In “RISnet: a Dedicated Scalable Neural Network Architecture for Optimization of Reconfigurable Intelligent Surfaces,” RISNet is no longer a reinforcement-learning controller but a dedicated neural architecture for weighted sum-rate maximization in an RIS-assisted downlink broadcast channel (Peng et al., 2022). The signal model is

$$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$1

with $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$2 BS antennas, $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$3 single-antenna users, and $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$4 RIS elements. The optimization objective is the weighted sum-rate under BS transmit-power and unit-modulus RIS constraints. The paper rewrites the channel as

$$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$5

so that the direct link can be represented in an RIS-element-aligned form.

The central architectural principle is parameter sharing across homogeneous RIS antennas. For RIS element $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$6, the feature vector is

$$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$7

giving $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$8 real-valued features per element. RISNet uses these element-wise features together with local and global processing blocks, so that the number of trainable parameters is independent of $$R=\sum_i B_i \log_2(1+\gamma_i^D)+\sum_k B_k \log_2(1+\gamma_k^C),$$9. Two variants are defined. The permutation-variant version is sensitive to user ordering and is suited to asymmetric objectives such as weighted sum-rate. The permutation-invariant version processes users symmetrically and is appropriate when the objective is symmetric with respect to users.

Training is unsupervised and follows an alternating-optimization procedure. The BS precoder is computed by weighted minimum mean squared error, held fixed, and the RISNet parameters are updated by stochastic gradient ascent using Adam so as to improve the weighted sum-rate. The final output is a real-valued vector converted to

$$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$0

which guarantees unit-modulus diagonal phase shifts.

The scalability argument is quantified explicitly. The paper reports 10,001 trainable variables for the permutation-variant RISNet and 7,301 for the permutation-invariant RISNet, compared with 1,049,600 trainable variables for a single $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$1 fully connected layer. In the reported experiments with $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$2, $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$3, $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$4, Rayleigh fading, and TSNR values $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$5, $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$6, and $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$7, both RISNet variants significantly outperform random phase shifts and the BCD algorithm. At TSNR $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$8, the reported weighted sum-rate is $$\boldsymbol{\Theta} \triangleq \operatorname{diag}(A e^{j\theta_1},\dots,A e^{j\theta_N}),$$9 for permutation-variant RISNet, $\rho_{k,i}$0 for permutation-invariant RISNet, $\rho_{k,i}$1 for random phases, and $\rho_{k,i}$2 for BCD. The runtime comparison is similarly stark: RISNet testing on 1024 samples takes about 1 minute on a laptop, whereas BCD on the same 1024 samples requires more than 24 hours on a server.

4. Extensions: partial CSI, mutual coupling, NOMA, implicit estimation, and RSMA

Subsequent papers generalize RISnet along several orthogonal axes. “RISnet: A Scalable Approach for Reconfigurable Intelligent Surface Optimization with Partial CSI” keeps the SDMA weighted-sum-rate setting but assumes CSI is available for only 16 RIS elements out of a $\rho_{k,i}$3-element surface, i.e. only $\rho_{k,i}$4 of the elements (Peng et al., 2023). The network uses expansion layers so that partial CSI from anchor elements can be propagated to all RIS elements. In deterministic ray-tracing channels, partial CSI achieves performance similar to full CSI; in pure i.i.d. Gaussian channels, partial CSI fails because the 16-element CSI is not representative. The same paper reports that BCD needs more than 10 minutes for 2000 iterations on one channel realization, whereas a trained RISnet produces the configuration in a few milliseconds, and it also supports discrete phase shifts

$\rho_{k,i}$5

with only moderate performance loss.

“RISnet: A Domain-Knowledge Driven Neural Network Architecture for RIS Optimization with Mutual Coupling and Partial CSI” extends the channel model to

$\rho_{k,i}$6

so that mutual coupling between RIS elements is explicit rather than ignored (Peng et al., 2024). The paper defines a permutation-invariant RISnet for SDMA and a permutation-variant RISnet for NOMA, retains the partial-CSI anchor idea, and reports that considering mutual coupling in training significantly improves performance when the test RIS has mutual coupling. It also states that RISnet computes phase shifts in a few milliseconds, whereas BCD takes more than 20 minutes.

The NOMA line reformulates RISnet around quasi-degraded channels and analytical precoding. “Non-Orthogonal Multiple Access Assisted by Reconfigurable Intelligent Surface Using Unsupervised Machine Learning” uses RISnet to shape the channel so that the optimal NOMA precoding admits a closed-form expression, with the learning objective

$\rho_{k,i}$7

balancing quasi-degradation and transmit power (Siegismund-Poschmann et al., 2023). “RIS-Assisted NOMA with Partial CSI and Mutual Coupling: A Machine Learning Approach” then combines analytical optimal BS precoding, partial CSI, mutual coupling, and a scalable element-wise architecture; it reports control of more than 1000 RIS elements, uses only 16 RIS elements out of 1296 to provide CSI in the reported setup, and states that SDR fails to scale beyond 64 RIS elements while RISnet handles 1296 elements, with trained inference in milliseconds versus seconds for SDR (Peng et al., 11 Aug 2025).

A further generalization replaces explicit CSI by pilot-derived observations. “A Scalable Machine Learning Approach Enabled RIS Optimization with Implicit Channel Estimation” combines RISnet with a structured pilot-probing scheme in which the RIS cycles through preset configurations, and the network maps pilot differences such as

$\rho_{k,i}$8

and

$\rho_{k,i}$9

directly to RIS configurations without explicit channel estimation (Peng et al., 10 Aug 2025). On DeepMIMO channels with 1296 RIS elements, the reported runtimes are about 70 ms for RISnet, about 253 s for BCD, and about 0.23 s for a DRL baseline; the reported weighted sum-rate bars are about 2.30 bit/s/Hz for BCD, about 0.38 bit/s/Hz for Random, about 1.13 bit/s/Hz for DRL, about 3.11 bit/s/Hz for RISnet at pilot SNR $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$0, and about 3.55 bit/s/Hz for RISnet with no thermal noise.

The RSMA extension preserves the partial-CSI philosophy while changing the multiple-access layer. “Scalable Design for RIS-Assisted Multi-User Downlink System Empowered by RSMA under Partial CSI” couples RISnet with a low-complexity RSMA precoder and states that RISnet infers full CSI from partial observations, with 8 layers in the reported implementation, namely 6 dense layers and 2 expansion layers (Fang et al., 15 Apr 2026). In its layer-depth study, $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$1 yields WSR $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$2 bit/s/Hz and training time $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$3 min, compared with $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$4 bit/s/Hz at $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$5, $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$6 bit/s/Hz at $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$7, and $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$8 bit/s/Hz at $${\cal S}=[\boldsymbol S_t,\boldsymbol S_r,\boldsymbol S_u,\boldsymbol s_{RIS},\boldsymbol s_{BS};\boldsymbol \Theta],$$9. The paper further reports that partial-CSI RISnet approaches the full-CSI baseline in deterministic ray-tracing channels, while RSMA significantly mitigates performance loss when channel uncertainty increases.

5. Network-level and non-terrestrial interpretations

Not all uses of RISnet denote a single neural architecture. In “Reconfigurable Intelligent Surfaces in Action for Non-Terrestrial Networks,” the term is used in a broader system sense: “a network of smart surfaces embedded into the vertical communications stack” connecting terrestrial networks, HAPS, satellites, relay spacecraft, and deep-space nodes (Tekbıyık et al., 2020). The emphasis here is on low-power, low-weight, low-profile, and low-cost hardware under SWaP constraints, together with passive beam steering, multi-hop relays, and channel-estimation robustness through graph attention networks. The paper also highlights environmental impairments: for a Starlink-like intra-plane inter-satellite link at 23 GHz with satellite spacing of 945.4 km, moving from weak scintillation to the transition region requires roughly 14 dB more transmit power to maintain the same error rate, and under misalignment fading around 250 dB additional power may be required to maintain BER around $|{\cal S}|=(2I+K+N+2).$0 when the beam center offset is about 1 m.

A similarly broad use appears in “RIS Assisted Wireless Networks: Collaborative Regulation, Deployment Mode and Field Testing,” which treats an RIS-assisted wireless network as a networked system organized around regulation architecture rather than only phase optimization (Zhao, 2024). The paper distinguishes Network Controlled Mode and Standalone Mode, formulates multi-RIS collaboration through

$|{\cal S}|=(2I+K+N+2).$1

and discusses phase alignment, multi-user pattern addition, blocking mechanisms, and multi-cell coordination. Its field-test prototype is a reflective RIS with center frequency 27 GHz, working band 26–28 GHz, UPA structure $|{\cal S}|=(2I+K+N+2).$2, beam scanning range $|{\cal S}|=(2I+K+N+2).$3, beam width $|{\cal S}|=(2I+K+N+2).$4–$|{\cal S}|=(2I+K+N+2).$5, and aperture 660 mm. The reported field tests in an indoor L-shaped corridor and an indoor open office area show that Network Controlled Mode consistently gives higher RSRP and higher downlink rate than Standalone Mode. In these papers, RISnet is therefore closer to a coordinated RIS-enabled networking paradigm than to a specific learnable phase-prediction block.

6. Constraints, misconceptions, and practical outlook

A recurrent misconception is that RISnet names a single canonical model. The literature in this corpus does not support that reading. Depending on the paper, RISnet may mean a CNN-based DQN for joint placement and phase control, a scalable unsupervised RIS-specific neural architecture for WSR maximization, a partial-CSI or mutual-coupling extension, a quasi-degraded NOMA optimizer, an implicit-channel-estimation pipeline, or a broader coordinated RIS-enabled network fabric (Ji et al., 2020, Peng et al., 2022, Tekbıyık et al., 2020).

A second misconception is that RISnet methods uniformly assume full CSI. Several papers are explicitly motivated by the opposite claim: full CSI acquisition is a major bottleneck for large RISs. Partial CSI from anchor elements, pilot-derived implicit channel estimation, and expansion layers are therefore not minor implementation details but central design principles (Peng et al., 2023, Peng et al., 10 Aug 2025). At the same time, the reported results show that these strategies are channel-model dependent: partial CSI performs almost as well as full CSI in deterministic ray-tracing channels, degrades under ray-tracing plus i.i.d. perturbation, and fails in purely i.i.d. scattering-dominated settings (Peng et al., 2024).

A third misconception is that fast inference resolves the practical feasibility of RIS-assisted networking. The algorithmic papers do show that trained RISnet models replace expensive iterative optimization by near-instant or millisecond-scale forward passes. However, cellular deployment papers stress unresolved issues in signaling overhead, interference, spectral containment, deployment flexibility, hardware response time, and regulation. “RIS in Cellular Networks -- Challenges and Issues” argues that RIS was not continued as a study item in 3GPP Releases 18 and 19, compares RIS with the standardized network-controlled repeater, and reports that a 1000-element RIS halfway between BS and UE gives about 26 dB lower gain than direct LoS, while NCR is generally more robust and often better overall in the presented coverage studies (Åström et al., 2024).

The resulting picture is technically coherent even though the terminology is heterogeneous. In the narrow algorithmic sense, RISnet refers to domain-knowledge-driven neural controllers that exploit RIS homogeneity, local/global feature aggregation, unsupervised objective-driven training, and hybrid coupling with analytical precoding. In the broader networking sense, it denotes a coordinated RIS-enabled communications fabric whose effectiveness depends not only on optimization quality but also on channel structure, control-plane design, hardware nonidealities, and deployment economics.