# Uniform-History Reset Optimizer

- ID: 1702
- Canonical URL: https://synthcore.org/idea/1702/uniform-history-reset-optimizer
- API JSON: https://synthcore.org/api/idea/1702.json
- API Markdown: https://synthcore.org/api/idea/1702.md
- Verification status: unverified
- Source: [arXiv:2608.02297](https://arxiv.org/abs/2608.02297)
- Category: dynamics
- Solves: stability, generalization
- ML areas: optimizer, training-dynamics, regularization
- Math tags: dynamical-systems, stochastic-processes, control-theory, optimization
- Ratings: usefulness 6/10; difficulty 4/10; novelty 7/10

## Idea description

Augment gradient descent with stochastic relocations to uniformly sampled historical parameter vectors. In expectation, the optimizer receives a non-Markovian correction toward the running average of all previous iterates, which can suppress runaway directions and revisit earlier basins instead of remaining trapped in a sharp or unstable region.

## Mathematical statement

Let \(\theta(t)\in\mathbb{R}^d\) be network parameters, \(L(\theta)\) the loss, \(r\geq0\) the reset rate, and \(g(\theta)=\nabla L(\theta)\). The continuous mean-field analogue of uniform-history resetting is \(\dot\theta(t)=-g(\theta(t))+r(\bar\theta(t)-\theta(t))\), where \(\bar\theta(t)=t^{-1}\int_0^t\theta(s)\,ds\). The event process uses Poisson reset probability \(r\Delta t\) per step and, on reset, replaces \(\theta_k\) by \(\theta_J\) with \(J\) uniformly sampled from \(\{0,\ldots,k\}\). For a linearized unstable mode \(x\) with \(\dot x=ax\), \(a>0\), the resetting equation is \(\dot x=ax+r(t^{-1}\int_0^t x(s)\,ds-x)\). Because the historical average of an exponentially growing mode is asymptotically \(x/(at)\), the leading growth exponent is approximately \(a-r\), predicting stabilization near \(r_c\approx a\). Stable modes receive persistent historical forcing and need not relax purely exponentially.

## Key formulas

- $$\bar{\theta}(t)=\frac{1}{t}\int_0^t\theta(s)\,ds,\qquad \dot{\theta}(t)=-\nabla L(\theta(t))+r\bigl(\bar{\theta}(t)-\theta(t)\bigr).$$
- $$\theta_{k+1/2}=\theta_k-\eta\,\widehat{\nabla L}(\theta_k),\qquad \theta_{k+1}=\begin{cases}\theta_J,&U_k<r\Delta t,\ J\sim\operatorname{Uniform}\{0,\ldots,k\},\\\theta_{k+1/2},&U_k\ge r\Delta t,\end{cases}$$
- $$\dot{x}=a x+r\left(\frac{1}{t}\int_0^t x(s)\,ds-x\right),\qquad r_c\approx a.$$$$
- $$\frac{1}{t}\int_0^t e^{a s}\,ds=\frac{e^{a t}-1}{a t}\sim\frac{e^{a t}}{a t}\quad (a>0).$$

## Implementation notes

Integrate the mechanism into SGD or Adam after the ordinary gradient update and before the next minibatch. Maintain a parameter snapshot buffer containing every \(M\)-th iterate, or use reservoir sampling when storing all iterates is too expensive. At step \(k\), compute the normal update \(\theta'_{k+1}=\theta_k-\eta g_k\). Draw \(u\sim U[0,1]\); if \(u<r\Delta t\), choose an index \(J\) uniformly from the buffer and set \(\theta_{k+1}=\theta_J\); otherwise set \(\theta_{k+1}=\theta'_{k+1}\). Reset optimizer momentum together with the parameters, or test both momentum-preserving and momentum-resetting variants. A deterministic low-variance variant maintains an online historical mean \(\bar\theta_k\) and applies \(\theta_{k+1}=\theta'_{k+1}+\eta r(\bar\theta_k-\theta'_{k+1})\). The paper supplies the uniform-history kernel and its non-Markovian averaging mechanism; the gradient update, discrete-time implementation, buffer thinning, and optimizer-state policy are adaptations. First use the quadratic loss \(L(x,y)=\frac12(\lambda x^2-a y^2)\), with \(a>0\), and compare ordinary SGD with reset rates \(r/a\in\{0.5,0.8,1.0,1.2,2.0\}\). Measure the long-time growth of \(|y_k|\) and fit the stability boundary; the prediction is a transition near \(r_c\approx a\), with measured \(r_c/a\) within 20% after accounting for discrete step size. Then train a two-layer MLP on MNIST using SGD, testing several reset rates and a buffer of at least 100 snapshots. The predicted signatures are reduced parameter-norm divergence above the quadratic threshold, slower approximately algebraic relaxation rather than a single exponential, and measurable returns to earlier parameter basins. Compare accuracy, loss variance, reset frequency, and sensitivity to buffer size.

## Disclaimer

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