# Actionable-Information Optimizer

- ID: 1005
- Canonical URL: https://synthcore.org/idea/1005/actionable-information-optimizer
- API JSON: https://synthcore.org/api/idea/1005.json
- API Markdown: https://synthcore.org/api/idea/1005.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.16186](https://arxiv.org/abs/2607.16186)
- Category: dynamics
- Solves: stability, speedup, generalization
- ML areas: optimizer, training-dynamics, regularization
- Math tags: information-theory, stochastic-processes, control-theory, statistics
- Ratings: usefulness 7/10; difficulty 6/10; novelty 7/10

## Idea description

Insert a finite-resolution observation channel between minibatch statistics and the optimizer update, then distinguish information that predicts useful future loss reduction from information that is present in the gradient but has no control value. Use the actionable representation to select the update and suppress increasingly fine, noisy measurements that do not improve progress.

## Mathematical statement

Let S_t be the optimizer state and minibatch signal, Z_t = Q_delta(E_phi(S_t)) + noise be its finite-resolution observation, and A_t be the parameter update. Define acquired information as I_acq = I(Z_t; S_t). Define future utility as U_t = -(L(theta_(t+H)) - L(theta_t)), where H is a short horizon, and define useful information as I_use = I(Z_t; U_t). The unavailable component is I_un = I_acq - I_use. The controller chooses an update by maximizing predicted utility minus an information cost. The mechanism predicts that I_acq can grow as resolution improves while I_use and useful optimization progress plateau.

## Key formulas

- I_acq = I(Z_t; S_t) = I_use + I_un, where I_use = I(Z_t; U_t) and I_un = I(Z_t; S_t) - I(Z_t; U_t).
- U_t = -(L(theta_(t+H)) - L(theta_t)); theta_(t+1) = theta_t + A_t(Z_t); Z_t = Q_delta(E_phi(S_t)) + noise.
- A_t* = argmax_A { E[U_t | Z_t, A] - lambda I(Z_t; S_t) }, with lambda greater than zero.
- Expected useful progress is bounded by C times I_use, while I_acq may continue increasing as delta approaches zero.

## Implementation notes

1. Integration point: replace the raw-gradient input to SGD or Adam with a small controller. At step t construct S_t from the gradient g_t, optimizer second moment v_t, gradient norm, current loss, previous loss, and minibatch gradient variance. Encode S_t with a low-dimensional network E_phi, quantize or perturb it with resolution delta, and obtain Z_t. A policy head outputs a step-size multiplier, direction correction, or accept/reject decision. 2. Pseudocode: compute the base optimizer update a_base; compute z = quantize(E_phi(S_t), delta); enumerate candidate actions a; predict utility u_hat = f_psi(z, a); choose a_star = argmax_a(u_hat - lambda times kappa(z)); apply theta = theta + a_star. After H steps, label the observation with U_t = -(loss_(t+H) - loss_t), and train f_psi to predict this utility. Train the encoder using a variational information penalty or controlled quantization noise. 3. Computed from the mechanism: the finite-resolution observation channel and the acquired-versus-useful information decomposition. Estimated empirically: mutual-information proxies, utility prediction accuracy, update variance, and the largest stable learning rate. Use a variational or classifier-based information estimate initially rather than requiring exact mutual information. 4. First cheap experiment: compare SGD, Adam, and this controller on a two-layer MLP with MNIST and a small ResNet with CIFAR-10. Sweep delta over 8, 4, 2, 1, and 0.5 normalized gradient units, and sweep lambda. The predicted signature is that acquired information increases approximately as -d_eff times log(delta) as delta decreases, while useful information and final loss improvement plateau. Finer resolution beyond the plateau should increase update variance without improving convergence. Sweep learning rates to estimate the divergence boundary; the controller should preserve or increase the maximum stable learning rate relative to raw high-resolution gradients. Require the plateau and boundary shift to replicate across three seeds and two model sizes.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a finite-resolution actionable-information toy optimizer and numerically verified the proposed mechanism. Across three seeds, I_acq increased logarithmically with resolution (slope 0.9967 versus predicted 1.0, R²=0.999997), while the final fine-resolution I_use increment was only 0.000375 nats versus 0.6925 nats for I_acq, and progress changed by -0.00135. Thus the acquired-information/unavailable-information signature manifested clearly; quantization was not uniformly better than raw SGD, so this is not evidence of a general optimization win.

### Mechanism check

- Verdict: Built a finite-resolution actionable-information toy optimizer and numerically verified the proposed mechanism. Across three seeds, I_acq increased logarithmically with resolution (slope 0.9967 versus predicted 1.0, R²=0.999997), while the final fine-resolution I_use increment was only 0.000375 nats versus 0.6925 nats for I_acq, and progress changed by -0.00135. Thus the acquired-information/unavailable-information signature manifested clearly; quantization was not uniformly better than raw SGD, so this is not evidence of a general optimization win.
- Confidence: 8/10
- Limitations: Only a one-dimensional quadratic Monte Carlo channel was tested; no learned encoder, utility predictor, Adam/SGD neural-network training, MNIST/CIFAR, lambda information-cost sweep, stability boundary, wall-clock speed, or two-model-size replication was evaluated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [actionable_optimizer_toy.py](https://synthcore.org/code/413/actionable_optimizer_toy.py)
- [bench_actionable.py](https://synthcore.org/code/413/bench_actionable.py)
- [bench_report.json](https://synthcore.org/code/413/bench_report.json)
- [report.md](https://synthcore.org/code/413/report.md)
- [report_bench_2026-08-31T114502.md](https://synthcore.org/code/413/report_bench_2026-08-31T114502.md)
- [toy_results.json](https://synthcore.org/code/413/toy_results.json)
- [verification_summary.json](https://synthcore.org/code/413/verification_summary.json)
- [Download all files as ZIP](https://synthcore.org/download/413)

## Disclaimer

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