Adaptive Operator Learning Network (AOL-Net)

• AOL-Net is a family of adaptive architectures that condition operator computations on task-specific signals such as rewards, graph states, and spectral data.
• It employs diverse techniques including attention-based resolvent operators, NTK-guided reweighting in DeepONets, and graph-based selections for zero-shot adaptation.
• Empirical results show AOL-Net variants achieve faster convergence, reduced errors, and improved performance across reinforcement learning, inverse problems, crop yield prediction, and physics-informed PDEs.

Searching arXiv for the specified AOL-Net-related papers to ground the article in current records. I’m going to look up the listed arXiv records so the article can cite them precisely. Adaptive Operator Learning Network (AOL-Net) is a non-unified designation used in several arXiv works for architectures that adapt an operator-valued computation to changing rewards, graph states, spectral coefficients, spatial filters, scales, or loss balances. In its most explicit operator-theoretic form, AOL-Net denotes a learned resolvent operator $\mathcal O:\mathcal R\to\mathcal Q$ that maps a reward function $r:\mathcal S\times\mathcal A\to\mathbb R$ to its corresponding optimal or policy-evaluation $Q$-function, thereby enabling zero-shot reward transfer in offline reinforcement learning (Tang et al., 2022). The same label is also attached to NTK-reweighted DeepONets, graph-based operator selectors in ALNS, adaptive spectral inverse solvers, adaptive edge-operator modules for crop yield prediction, adaptive-scale physics-informed DeepONets, and Pareto-balanced replica-exchange operator learners (Wang et al., 2021, Johnn et al., 2023, Dong et al., 21 Mar 2026, Zhang et al., 8 Jul 2025, Mou et al., 18 Nov 2025, Lu et al., 31 Aug 2025). This suggests that AOL-Net is best understood not as a single canonical model class, but as a family of adaptive mechanisms applied to operator learning and operator selection.

1. Scope and nomenclature

The literature assigns the name AOL-Net to several technically distinct constructions. In reinforcement learning, it refers to a learned mapping from reward functions to value functions (Tang et al., 2022). In improved DeepONets, it denotes a combination of adaptive re-weighting and an architecture designed to mitigate magnitude bias and vanishing gradients (Wang et al., 2021). In ALNS, it denotes a Deep RL + GNN replacement for the classic Roulette-Wheel adaptive layer, where the adaptive object is the choice of destroy and repair operators (Johnn et al., 2023). In inverse problems, SC-Net is presented as an instance of AOL-Net because it learns a pointwise adaptive spectral filter (Dong et al., 21 Mar 2026). In crop-yield prediction, AOL-Net is a module that dynamically selects among Sobel, Scharr, and a learnable operator (Zhang et al., 8 Jul 2025). In physics-informed PDE learning, related usages include adaptive-scale trunk embeddings and multi-objective balancing of operator and physics losses (Mou et al., 18 Nov 2025, Lu et al., 31 Aug 2025).

Setting	AOL-Net realization	Adaptive mechanism
Offline RL	Resolvent operator $r\mapsto Q(\cdot;\,r)$	Reward-conditioned value operator
DeepONet training	Weighted branch–trunk network	Per-term NTK-guided weighting
ALNS	DQN + GNN policy	Destroy/repair operator selection
Inverse problems	SC-Net	Spectral filter $\phi_\theta(\sigma)$
Crop yield prediction	CNN branch with operator bank	Hard selection among edge operators
Physics-informed PDEs	Adaptive-scale or multi-objective operator learner	Coordinate rescaling or Pareto loss balancing

A common misconception is that AOL-Net designates one standardized architecture. The published usages do not support that interpretation. The consistent theme is adaptivity at the level of an operator, operator family, or operator-conditioned computation, but the mathematical object being adapted differs substantially across papers.

2. Resolvent-operator formulation in reinforcement learning

The most explicit formalization appears in Operator Deep Q-Learning, where AOL-Net is a learned resolvent operator

$\mathcal O:\mathcal R\longrightarrow\mathcal Q,$

taking an arbitrary reward function to its corresponding $Q$-function (Tang et al., 2022). For any fixed reward $r$, the optimal-Bellman operator is

$$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$

The target operator $\mathcal O$ satisfies

$r:\mathcal S\times\mathcal A\to\mathbb R$0

so that $r:\mathcal S\times\mathcal A\to\mathbb R$1 is the unique fixed point of the Bellman equation. In the policy-evaluation variant, the paper instead approximates a linear operator $r:\mathcal S\times\mathcal A\to\mathbb R$2 satisfying

$r:\mathcal S\times\mathcal A\to\mathbb R$3

A key implementation issue is how to input a function $r:\mathcal S\times\mathcal A\to\mathbb R$4 into a neural network. The method discretizes $r:\mathcal S\times\mathcal A\to\mathbb R$5 over a fixed reference set $r:\mathcal S\times\mathcal A\to\mathbb R$6 and uses the vector

$r:\mathcal S\times\mathcal A\to\mathbb R$7

as a reward embedding. This converts operator learning into a parametric neural approximation problem while preserving the interpretation that the network computes a map from reward functions to value functions.

Three architectural variants are described. The attention-based evaluation operator computes non-negative normalized weights

$r:\mathcal S\times\mathcal A\to\mathbb R$8

and then forms

$r:\mathcal S\times\mathcal A\to\mathbb R$9

By construction, the weights are non-negative and sum to $Q$0, guaranteeing linearity, monotonicity, and constant-shift invariance of $Q$1. The linear-decomposition variant replaces attention weights by inner products $Q$2, giving a factored form whose per-batch cost is reduced from $Q$3 to $Q$4. For control, the max-out operator instantiates $Q$5 evaluation blocks and takes a pointwise maximum, reflecting the identity $Q$6.

The operational significance of this formulation is zero-shot adaptation. Once the network is trained on a family of rewards, a novel reward $Q$7 is simply embedded and passed through the pretrained encoders; no further gradient steps or environment interactions are required. In the terminology adopted there, this learned mapping

$Q$8

is the AOL-Net.

3. Architectural realizations across domains

Beyond reinforcement learning, the label AOL-Net encompasses several distinct architectural motifs. In improved DeepONets, the core approximation remains the branch–trunk decomposition

$Q$9

but the network is augmented with encoders, layerwise cross-gating, and residual signal propagation. The associated adaptive component is not operator selection in the combinatorial sense; it is adaptive re-weighting of training terms, motivated by NTK analysis and magnitude bias (Wang et al., 2021).

In ALNS, the architecture is a 3-layer Graph Attention Network or GraphSAGE style message-passing backbone plus a small MLP head that outputs one $r\mapsto Q(\cdot;\,r)$0-value per operator (Johnn et al., 2023). The state includes a static graph $r\mapsto Q(\cdot;\,r)$1, node features, the current solution $r\mapsto Q(\cdot;\,r)$2, the removal list $r\mapsto Q(\cdot;\,r)$3, the phase $r\mapsto Q(\cdot;\,r)$4, and the remaining budget $r\mapsto Q(\cdot;\,r)$5. Node features are augmented with binary indicators for membership in $r\mapsto Q(\cdot;\,r)$6 and $r\mapsto Q(\cdot;\,r)$7, and mean-pooling yields a graph embedding $r\mapsto Q(\cdot;\,r)$8. The network therefore adapts operator choice to graph-structured context.

In SC-Net for inverse problems, the operator is represented in the spectral domain of a compact linear forward map $r\mapsto Q(\cdot;\,r)$9 with singular system $\phi_\theta(\sigma)$0 (Dong et al., 21 Mar 2026). Instead of a classical regularization filter, the reconstruction uses

$\phi_\theta(\sigma)$1

The adaptive object is the scalar filter $\phi_\theta(\sigma)$2, implemented by a shared MLP acting pointwise on spectral inputs. Because all modes share the same MLP, the architecture is described as resolution-independent and mesh-independent.

In DFYP for crop yield prediction, AOL-Net is a local spatial branch operating on an RCA-refined feature map $\phi_\theta(\sigma)$3 (Zhang et al., 8 Jul 2025). An operator bank $\phi_\theta(\sigma)$4 is defined, with $\phi_\theta(\sigma)$5 and

$\phi_\theta(\sigma)$6

where

$\phi_\theta(\sigma)$7

A hard gate selects exactly one operator based on a historical performance score, and the selected edge map is fused with the raw feature map through

$\phi_\theta(\sigma)$8

The adaptive element is therefore explicit operator choice within a convolutional operator bank.

PAS-Net introduces a different notion of adaptation: adaptive-scale embedding in the trunk input of a PI-DeepONet (Mou et al., 18 Nov 2025). Its local rescaling

$\phi_\theta(\sigma)$9

or equivalently $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$0 modifies the feature geometry seen by the trunk net. Morephy-Net uses yet another mechanism: a Physics-Informed DeepONet with an added Fourier convolution layer, where adaptation occurs through multi-objective treatment of $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$1, $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$2, and $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$3, together with replica-exchange SGLD and posterior sampling (Lu et al., 31 Aug 2025).

Taken together, these realizations show that the word “operator” in AOL-Net ranges from solution operators between Banach or Hilbert spaces to action operators in metaheuristics and discrete edge operators in vision modules.

4. Training objectives and optimization strategies

Training procedures vary with the meaning assigned to the operator. In the RL resolvent formulation, the network is trained by minimizing Bellman squared error over tuples $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$4 obtained by sampling $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$5 from an offline replay buffer and sampling $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$6 from a pre-specified reward family (Tang et al., 2022):

$\mathcal O:\mathcal R\longrightarrow\mathcal Q,$7

A slowly updated target network, gradient clipping, and $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$8 regularization may be used exactly as in standard DQN.

Improved DeepONets replace uniform weighting by adaptive per-term weights $\mathcal O:\mathcal R\longrightarrow\mathcal Q,$9 in a fully decoupled loss

$Q$0

The diagonal NTK entries

$Q$1

drive the update. The paper states that $Q$2 gives “NTK weights,” $Q$3 gives “moderate NTK weights,” and $Q$4 reverts to no weighting (Wang et al., 2021). The stated purpose is to equalize convergence rates or gradient magnitudes and prevent large-output examples from dominating training.

The ALNS variant is trained by DQN with experience replay and a target network (Johnn et al., 2023). For each sampled transition $Q$5, the target is

$Q$6

and the loss is MSE. Reported hyperparameters include Adam with learning rate $Q$7, batch size $Q$8, $Q$9, target updates every $r$0 gradient steps, hidden dimension $r$1, and a softmax integration temperature $r$2.

SC-Net is trained on synthetic inverse-problem data generated by drawing ground truths with prescribed Sobolev regularity, applying the forward operator, and adding noise (Dong et al., 21 Mar 2026). The loss is a Sobolev-weighted $r$3 reconstruction error,

$r$4

optimized by Adam with batch size $r$5–$r$6 and learning rate $r$7.

In DFYP, the AOL-Net branch, the ViT branch, the fusion weights $r$8, the operator interpolation parameter $r$9, and the CNN parameters are trained end-to-end under a global MSE objective (Zhang et al., 8 Jul 2025). The gating rule is hard-max over historical performance, $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$0 is initialized at $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$1 and clamped to $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$2, and no auxiliary diversity or smoothness regularizer is introduced specifically for AOL-Net.

PAS-Net and Morephy-Net operate in the physics-informed regime. PAS-Net combines data, PDE-residual, and boundary-condition terms in

$$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$3

with automatic differentiation used for residuals (Mou et al., 18 Nov 2025). Morephy-Net does not fix scalar weights in advance; it treats operator and physics losses as separate objectives in a multi-objective problem, uses a refined NSGA-III to obtain a Pareto front, then employs replica-exchange SGLD for global exploration and posterior sampling (Lu et al., 31 Aug 2025).

5. Reported empirical behavior

The RL instantiation is evaluated on reward transfer for offline policy evaluation and offline policy optimization (Tang et al., 2022). In Pendulum-Angle, HalfCheetah-Vel, and Ant-Dir, the attention-based operator is reported to converge faster and to lower OPE-error than successor-feature baselines, the linear-decomposition operator, and a vanilla two-stream operator net à la DeepONet. In Pendulum-Angle, the attention operator achieves OPE-MSE $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$4 on new test rewards within $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$5 gradient steps, whereas successor-feature remains above $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$6. For control, the max-out operator reaches normalized return $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$7 of the online optimum in all three domains, zero-shot.

The ALNS variant reports gains both in the standalone MDP and when integrated into ALNS (Johnn et al., 2023). Once the destroy pool size satisfies $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$8, the DQN+GNN AOL-Net outperforms LRW and RAN across all three instance types, with gains up to $$(\mathcal T^r Q)(s,a) = r(s,a)+\gamma\,\mathbb E_{s'\sim P(\cdot\mid s,a)}\bigl[\max_{a'}Q(s',a')\bigr].$$9 cumulative-reward margin on larger portfolios. In integrated ALNS, it yields the lowest average and best costs for portfolios $\mathcal O$0, improving over CRW by $\mathcal O$1–$\mathcal O$2 on average. A GNN trained on $\mathcal O$3 generalizes to $\mathcal O$4, and the advantage is largest at small destroy scales $\mathcal O$5.

SC-Net reports a convergence study on a 1D Fredholm integral equation with $\mathcal O$6 and $\mathcal O$7 (Dong et al., 21 Mar 2026). It achieves empirical slope $\mathcal O$8 on a log-log error-versus-noise plot, matching the claimed theoretical rate $\mathcal O$9, while Oracle Tikhonov saturates at slope $r:\mathcal S\times\mathcal A\to\mathbb R$00. Under $r:\mathcal S\times\mathcal A\to\mathbb R$01, the learned filter has $r:\mathcal S\times\mathcal A\to\mathbb R$02 for $r:\mathcal S\times\mathcal A\to\mathbb R$03, $r:\mathcal S\times\mathcal A\to\mathbb R$04 for $r:\mathcal S\times\mathcal A\to\mathbb R$05, and a sharp, differentiable cutoff sharper than Tikhonov’s $r:\mathcal S\times\mathcal A\to\mathbb R$06. In zero-shot super-resolution, training on $r:\mathcal S\times\mathcal A\to\mathbb R$07 and testing on $r:\mathcal S\times\mathcal A\to\mathbb R$08 yields average relative errors $r:\mathcal S\times\mathcal A\to\mathbb R$09, $r:\mathcal S\times\mathcal A\to\mathbb R$10, $r:\mathcal S\times\mathcal A\to\mathbb R$11, and $r:\mathcal S\times\mathcal A\to\mathbb R$12.

In crop yield prediction, the ablation on the MODIS soybean dataset reports the following sequence (Zhang et al., 8 Jul 2025): the baseline dual-branch fusion without AOL or RCA has $r:\mathcal S\times\mathcal A\to\mathbb R$13, $r:\mathcal S\times\mathcal A\to\mathbb R$14, $r:\mathcal S\times\mathcal A\to\mathbb R$15; fusion + AOL only has $r:\mathcal S\times\mathcal A\to\mathbb R$16, $r:\mathcal S\times\mathcal A\to\mathbb R$17, $r:\mathcal S\times\mathcal A\to\mathbb R$18; fusion + RCA only has $r:\mathcal S\times\mathcal A\to\mathbb R$19, $r:\mathcal S\times\mathcal A\to\mathbb R$20, $r:\mathcal S\times\mathcal A\to\mathbb R$21; and full DFYP has $r:\mathcal S\times\mathcal A\to\mathbb R$22, $r:\mathcal S\times\mathcal A\to\mathbb R$23, $r:\mathcal S\times\mathcal A\to\mathbb R$24. Relative improvements due to AOL alone are reported as $r:\mathcal S\times\mathcal A\to\mathbb R$25 ($r:\mathcal S\times\mathcal A\to\mathbb R$26 reduction), $r:\mathcal S\times\mathcal A\to\mathbb R$27 ($r:\mathcal S\times\mathcal A\to\mathbb R$28 reduction), and $r:\mathcal S\times\mathcal A\to\mathbb R$29. On Sentinel-2, integrating AOL yields RMSE reductions of approximately $r:\mathcal S\times\mathcal A\to\mathbb R$30–$r:\mathcal S\times\mathcal A\to\mathbb R$31 bu/acre and MAE reductions of $r:\mathcal S\times\mathcal A\to\mathbb R$32–$r:\mathcal S\times\mathcal A\to\mathbb R$33, while boosting $r:\mathcal S\times\mathcal A\to\mathbb R$34 by $r:\mathcal S\times\mathcal A\to\mathbb R$35–$r:\mathcal S\times\mathcal A\to\mathbb R$36.

Improved DeepONets report error reductions by factors of roughly $r:\mathcal S\times\mathcal A\to\mathbb R$37–$r:\mathcal S\times\mathcal A\to\mathbb R$38 across four operator-learning benchmarks (Wang et al., 2021). The anti-derivative ODE goes from baseline relative $r:\mathcal S\times\mathcal A\to\mathbb R$39 error $r:\mathcal S\times\mathcal A\to\mathbb R$40 to $r:\mathcal S\times\mathcal A\to\mathbb R$41; 1D advection from $r:\mathcal S\times\mathcal A\to\mathbb R$42 to $r:\mathcal S\times\mathcal A\to\mathbb R$43; 1D Burgers with $r:\mathcal S\times\mathcal A\to\mathbb R$44 from $r:\mathcal S\times\mathcal A\to\mathbb R$45 to $r:\mathcal S\times\mathcal A\to\mathbb R$46; and 2D Stokes flow from $r:\mathcal S\times\mathcal A\to\mathbb R$47–$r:\mathcal S\times\mathcal A\to\mathbb R$48 to $r:\mathcal S\times\mathcal A\to\mathbb R$49–$r:\mathcal S\times\mathcal A\to\mathbb R$50. PAS-Net reports lower errors than DeepONet and PI-DeepONet on viscous Burgers, diffusion-reaction, and 2D eikonal problems, with, for example, mean time-averaged relative $r:\mathcal S\times\mathcal A\to\mathbb R$51 error $r:\mathcal S\times\mathcal A\to\mathbb R$52 for Burgers versus $r:\mathcal S\times\mathcal A\to\mathbb R$53 for PI-DeepONet (Mou et al., 18 Nov 2025). Morephy-Net reports improved forward and inverse accuracy on Burgers and TFMDWE benchmarks, including Burgers inverse relative $r:\mathcal S\times\mathcal A\to\mathbb R$54 error $r:\mathcal S\times\mathcal A\to\mathbb R$55 versus $r:\mathcal S\times\mathcal A\to\mathbb R$56 for PI-DON and $r:\mathcal S\times\mathcal A\to\mathbb R$57 for PI-FDON, and TFMDWE inverse relative $r:\mathcal S\times\mathcal A\to\mathbb R$58 error $r:\mathcal S\times\mathcal A\to\mathbb R$59 versus $r:\mathcal S\times\mathcal A\to\mathbb R$60 and $r:\mathcal S\times\mathcal A\to\mathbb R$61 (Lu et al., 31 Aug 2025).

6. Theoretical properties and conceptual distinctions

Several AOL-Net variants are motivated by explicit structural priors. In the RL attention operator, non-negative normalized weights enforce linearity, monotonicity, and constant-shift invariance by design (Tang et al., 2022). In SC-Net, boundedness of $r:\mathcal S\times\mathcal A\to\mathbb R$62 yields a stability argument, universal approximation supports approximation of oracle filters on $r:\mathcal S\times\mathcal A\to\mathbb R$63, and the resulting reconstruction is described as discretization invariant because $r:\mathcal S\times\mathcal A\to\mathbb R$64 acts on continuous singular values rather than on a fixed mesh (Dong et al., 21 Mar 2026). In improved DeepONets, the central theoretical claim is that magnitude bias arises in the NTK regime because larger outputs induce larger gradient norms, and adaptive re-weighting corrects that bias (Wang et al., 2021). PAS-Net likewise uses NTK language, but its mechanism is geometric: the adaptive-scale channel adds a positive-semidefinite contribution to the NTK and increases the smallest eigenvalue, which accelerates gradient-flow convergence (Mou et al., 18 Nov 2025).

These theoretical claims are not interchangeable. The RL formulation is about a Bellman resolvent operator; SC-Net is about regularized inversion of compact operators; improved DeepONets and PAS-Net are about trainability of branch–trunk operator learners; Morephy-Net is about Pareto balancing, stochastic posterior exploration, and uncertainty quantification; and the ALNS and DFYP versions are closer to context-conditioned operator selection than to learning a continuous solution operator (Johnn et al., 2023, Zhang et al., 8 Jul 2025, Lu et al., 31 Aug 2025).

A second misconception is that “adaptive” always means the same thing. In the published usages, it may refer to ingesting a novel reward function without retraining, choosing among destroy or repair operators conditioned on the current graph state, learning a sharp spectral cutoff based on local signal-to-noise ratio, selecting among fixed and learnable edge filters, modifying trunk coordinates by local rescaling, or balancing data and physics losses on a Pareto front. The unifying interpretation is therefore narrower than the terminology might imply: AOL-Net names architectures in which the operative transformation is conditioned on task-dependent structure, but the form of that conditioning is domain-specific.

This suggests a broad research direction rather than a settled taxonomy. Across the cited works, the recurring objective is to preserve operator-level generalization under distributional shifts that are naturally expressed as changes in rewards, graph states, discretizations, resolutions, crop-year regimes, localized scales, or physics/data trade-offs.