Stochastic NG-RC: Next-Gen Reservoir Computing

• Stochastic NG-RC is a framework that extends next-generation reservoir computing to nonlinear, high-dimensional stochastic systems via controlled Itô SDEs.
• It leverages a reservoir of delayed state features and noise inputs, trained using ridge regression for efficient one-step prediction and adaptive control.
• Empirical evaluations demonstrate robust performance in multiscale dynamics, including applications in seizure suppression using real EEG data.

Stochastic next-generation reservoir computing (S-NG-RC) is a control and modeling framework that extends the next-generation reservoir computing (NG-RC) paradigm to nonlinear, high-dimensional stochastic dynamical systems. S-NG-RC integrates the computational efficiency of NG-RC with explicit stochastic analysis, enabling robust, event-triggered adaptive control and data-driven system identification for both simulated and real-world, multiscale processes with significant noise and uncertainty (Cheng et al., 14 May 2025).

1. Mathematical Structure and Stochastic Modeling

S-NG-RC is built upon controlled Itô stochastic differential equations (SDEs) of the general form:

$$dX_t = [f(X_t) + u_1(X_t)]\,dt + [g(X_t) + u_2(X_t)]\,dW_t,$$

where $X_t \in \mathbb{R}^n$ denotes the system state, $f:\mathbb{R}^n \to \mathbb{R}^n$ the drift, $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$ the diffusion, $u_1$ and $u_2$ state feedback controls, and $W_t$ is an $m$-dimensional Brownian motion. For much of the exposition, this simplifies to a constant-diffusion model with additive drift control:

$$dX_t = [f(X_t) + u(X_t)]\,dt + \sigma\,dW_t, \qquad X_0 = x_0,$$

which is discretized by Euler–Maruyama:

$$X_{i+1} = X_i + [f(X_i) + u_i]\,\Delta t + \sigma \sqrt{\Delta t}\,\xi_i,$$

with $X_t \in \mathbb{R}^n$0 i.i.d.

The core computational unit is the feature (reservoir) vector at time $X_t \in \mathbb{R}^n$1:

$X_t \in \mathbb{R}^n$2

which aggregates current and delayed state, and selected nonlinear monomials (typically up to third order). Additional features include control input $X_t \in \mathbb{R}^n$3 and noise $X_t \in \mathbb{R}^n$4, where, for additive noise, $X_t \in \mathbb{R}^n$5, and for multiplicative noise, $X_t \in \mathbb{R}^n$6. The S-NG-RC one-step predictor is a linear readout:

$X_t \in \mathbb{R}^n$7

with $X_t \in \mathbb{R}^n$8 and $X_t \in \mathbb{R}^n$9 collecting the readout weights.

2. Training, Adaptive Control, and Stability Guarantees

Learning the reservoir readout proceeds via ridge regression, using $f:\mathbb{R}^n \to \mathbb{R}^n$0 sample pairs:

$f:\mathbb{R}^n \to \mathbb{R}^n$1

with closed-form solution:

$f:\mathbb{R}^n \to \mathbb{R}^n$2

where $f:\mathbb{R}^n \to \mathbb{R}^n$3 is the Tikhonov regularization parameter.

For adaptive control, with $f:\mathbb{R}^n \to \mathbb{R}^n$4 as the desired next state, define the tracking error $f:\mathbb{R}^n \to \mathbb{R}^n$5. The linear error-dynamics template

$f:\mathbb{R}^n \to \mathbb{R}^n$6

(with $f:\mathbb{R}^n \to \mathbb{R}^n$7 for spectral radius $f:\mathbb{R}^n \to \mathbb{R}^n$8) prescribes exponential error decay. Solving for the feedback input yields:

$f:\mathbb{R}^n \to \mathbb{R}^n$9

The asymptotic stability of this control law is theoretically ensured using an extended stochastic LaSalle theorem: under existence of a Lyapunov function $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$0, radially unbounded and smooth, with the generator $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$1 of the controlled SDE satisfying $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$2, $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$3, and $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$4 continuous, and bounded $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$5-th moments of $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$6, it follows almost surely that

$g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$7

with $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$8 identifying the zero-error invariant set (Cheng et al., 14 May 2025).

3. Algorithmic Implementation

The S-NG-RC workflow can be decomposed into the following stages:

1. Data Preprocessing
   • Collect open-loop trajectories $g:\mathbb{R}^n \to \mathbb{R}^{n \times m}$9 using random probe inputs $u_1$0; record corresponding noise $u_1$1.
   • Assemble features $u_1$2, $u_1$3, $u_1$4.
2. Reservoir Initialization
   • Select polynomial orders/delays for monomials in $u_1$5; set regularization $u_1$6.
3. Readout Training
   • Form sample matrices $u_1$7 and $u_1$8 as above.
   • Compute $u_1$9.
4. Closed-Loop Control
   • For $u_2$0, observe $u_2$1; construct $u_2$2, estimate or sample $u_2$3.
   • Compute $u_2$4.
   • On event-trigger ($u_2$5 threshold), update $u_2$6 via the feedback law; otherwise, set $u_2$7.
   • Apply $u_2$8 to the true system and increment index.
5. Control Iteration
   • Repeat the control loop until the time horizon is reached.

No backpropagation or online optimization is required; only a single ridge regression solve and linear controller updates at runtime.

4. Empirical Performance on Stochastic Van-der-Pol Dynamics

S-NG-RC demonstrates robust adaptive control on the multi-scale, noise-driven Van-der-Pol oscillator, described by:

$u_2$9

Testing encompassed both additive and multiplicative noise, with $W_t$0, and $W_t$1.

• Low noise ($W_t$2): 1–2 step convergence to target, RMSE $W_t$3.
• High noise ($W_t$4): classical NG-RC diverges; S-NG-RC stable, RMSE = 0.3632.
• Multiplicative noise ($W_t$5): RMSE = 0.2359, with persistent, controlled oscillations in the fast coordinate $W_t$6.

Robustness is summarized in the following RMSE heatmap (averaged over 5 runs):

$W_t$7	0.1	0.5	1.0	2.0
1.0	0.165	0.223	0.310	0.504
0.5	0.179	0.275	0.363	0.690
0.1	0.233	0.482	0.748	1.105

A plausible implication is that S-NG-RC achieves stability and error convergence across three temporal scales and a broad noise intensity range, outperforming traditional NG-RC in high-noise regimes.

5. Data-Driven Applications: Epileptic EEG Control

S-NG-RC has been deployed for closed-loop modulation of pathological dynamics reconstructed from real-world epileptic EEG recordings:

• Governing-law identification:
  • Single EEG channel ($W_t$8), normalized and down-sampled ($W_t$9 s).
  • Drift ($m$0) and diffusion ($m$1) terms fitted from empirical data using a Kramers–Moyal expansion with a polynomial basis $m$2, regularized by LASSO.
  • Sparse solution: $m$3, $m$4 with drift RMSE $m$5 and diffusion RMSE $m$6.
• Seizure suppression control:
  • Target: transition seizure activity to resemble resting-state dynamics using the first 500 resting samples as reference and the next 500 seizure samples as control interval.
  • Perturbed training data generated via injected random $m$7 in the learned SDE.
  • S-NG-RC (regularization $m$8): one-step prediction RMSE = 0.1331 for perturbed data.
  • Closed-loop control over 100 seizure samples: RMSE = 0.0752. Kernel density estimation shows effective amplitude regulation, shifting network states toward the resting distribution.

6. Scalability, Robustness, and Current Limitations

S-NG-RC achieves high computational scalability through a single ridge regression solve for training and per-step linear updates, with per-step cost $m$9 for $$dX_t = [f(X_t) + u(X_t)]\,dt + \sigma\,dW_t, \qquad X_0 = x_0,$$0 hundreds of features. No iterative, gradient-based optimization is required.

Robustness arises from the explicit inclusion of noise features $$dX_t = [f(X_t) + u(X_t)]\,dt + \sigma\,dW_t, \qquad X_0 = x_0,$$1 in the reservoir, facilitating the learning of state-noise interactions and maintaining stability with both additive and multiplicative noise across multiple time scales.

Current limitations include:

• Governing-law errors: low-dimensional SDEs may not capture full network or non-Gaussian noise present in real data (e.g., EEG).
• Model bias: the random perturbation design for control law training may introduce biases.
• Error accumulation: long-term iteration may lead to compounding prediction errors.

Potential extensions encompass:

• Enriching reservoir features with non-Gaussian ($$dX_t = [f(X_t) + u(X_t)]\,dt + \sigma\,dW_t, \qquad X_0 = x_0,$$2-stable) noise models.
• Automating basis-function selection with stochastic stability criteria.
• Joint amplitude-frequency regulation using time-frequency embeddings.
• Optimizing event-trigger design for neuro-modulation safety margins (Cheng et al., 14 May 2025).