# Two-Scalar Robust Residual Adaptation

- ID: 2609
- Canonical URL: https://synthcore.org/idea/2609/two-scalar-robust-residual-adaptation
- API JSON: https://synthcore.org/api/idea/2609.json
- API Markdown: https://synthcore.org/api/idea/2609.md
- Verification status: unverified
- Source: [arXiv:2608.20182](https://arxiv.org/abs/2608.20182)
- Category: regularization
- Solves: stability, accuracy
- ML areas: rnn, ssm, training-dynamics, optimizer
- Math tags: control-theory, dynamical-systems, optimization, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Add two scalar adaptive gains to a neural controller or learned dynamical model: one estimates the unknown norm of the ideal neural approximation weights, and the other estimates the combined approximation, friction, and disturbance envelope. Sigma modification prevents unbounded gain growth, while the robust residual correction uses only these scalar estimates, independent of the number of neural features.

## Mathematical statement

Let $s$ be a tracking or prediction-error sliding variable, $\phi(x)$ the neural feature vector, and $W^*$ the unknown ideal feature weights. Maintain scalar estimates $\hat\alpha\geq0$ and $\hat\beta\geq0$, where $\alpha\geq\|W^*\|$ and $\beta$ bounds the aggregate unmodeled-error magnitude. Use sigma-modified updates $\dot{\hat\alpha}=\gamma_\alpha(\|\phi\|\|s\|-\sigma_\alpha\hat\alpha)$ and $\dot{\hat\beta}=\gamma_\beta(\|s\|-\sigma_\beta\hat\beta)$, with learning rates $\gamma_\alpha,\gamma_\beta>0$ and leakage rates $\sigma_\alpha,\sigma_\beta>0$. Apply the robust residual $r=-(\hat\alpha\|\phi\|+\hat\beta)\operatorname{sat}(s/\varepsilon)$, where $\varepsilon$ is a boundary-layer width. The Lyapunov function $V=\frac12s^\top s+\frac{(\hat\alpha-\alpha)^2}{2\gamma_\alpha}+\frac{(\hat\beta-\beta)^2}{2\gamma_\beta}$ gives practical decrease outside a residual ball; sigma leakage keeps estimates bounded under noise and persistent excitation.

## Key formulas

- $$\dot{\hat\alpha}=\gamma_\alpha\big(\|\phi(x)\|\,\|s\|-\sigma_\alpha\hat\alpha\big),\qquad \dot{\hat\beta}=\gamma_\beta\big(\|s\|-\sigma_\beta\hat\beta\big).$$
- $$r(x,s)=-(\hat\alpha\|\phi(x)\|+\hat\beta)\operatorname{sat}\!\left(\frac{s}{\varepsilon}\right),\qquad \operatorname{sat}(z)=\frac{z}{\max(1,\|z\|)}.$$
- $$V=\frac12\|s\|^2+\frac{(\hat\alpha-\alpha)^2}{2\gamma_\alpha}+\frac{(\hat\beta-\beta)^2}{2\gamma_\beta}.$$
- $$\dot V\leq-c\|s\|^2-\frac{\sigma_\alpha}{2}(\hat\alpha-\alpha)^2-\frac{\sigma_\beta}{2}(\hat\beta-\beta)^2+d_\varepsilon,$$

## Implementation notes

1. Integration point: attach the mechanism to a neural ODE controller, learned inverse-dynamics model, or recurrent predictor whose nominal output is $u_\theta(x)$. Define $s$ as a one-step state residual, tracking error, or teacher-forced hidden-state error. Add the robust correction $r$ to the nominal output, stopping gradients through $\hat\alpha$ and $\hat\beta$ for a deployment-style online controller. 2. Pseudocode: compute $\phi(x)$ and $s$; update $\hat\alpha\leftarrow[\hat\alpha+\Delta t\gamma_\alpha(\|\phi\|\|s\|-\sigma_\alpha\hat\alpha)]_+$ and $\hat\beta\leftarrow[\hat\beta+\Delta t\gamma_\beta(\|s\|-\sigma_\beta\hat\beta)]_+$; form $r=-(\hat\alpha\|\phi\|+\hat\beta)\operatorname{sat}(s/\varepsilon)$; apply $u=u_\theta+r$ or train on the corrected residual. 3. Computed from the paper: scalar uncertainty decomposition, sigma modification, and Lyapunov error structure. Estimated empirically: uncertainty scales, disturbance envelope, learning rates, leakage rates, and boundary-layer width. 4. First experiment: compare an MLP neural ODE with and without scalar adaptation on a forced Duffing oscillator containing hidden friction and additive disturbances; compare also against full per-feature adaptive weights. Measure long-horizon rollout error, gain variance, and tracking error. Prediction: scalar gains remain bounded and converge to a noise-dependent plateau; increasing $\sigma_\alpha$ or $\sigma_\beta$ reduces gain variance but increases steady residual, while decreasing $\varepsilon$ lowers tracking error until chattering and measurement noise dominate. At fixed disturbance amplitude, steady error should scale approximately linearly with $\varepsilon$.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.