# Complete Interval Abstraction Training

- ID: 1907
- Canonical URL: https://synthcore.org/idea/1907/complete-interval-abstraction-training
- API JSON: https://synthcore.org/api/idea/1907.json
- API Markdown: https://synthcore.org/api/idea/1907.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.06689](https://arxiv.org/abs/2608.06689)
- Category: dynamics
- Solves: stability, generalization, accuracy
- ML areas: rl, rnn, world-model, regularization
- Math tags: control-theory, dynamical-systems, numerical-analysis, linear-algebra
- Ratings: usefulness 8/10; difficulty 6/10; novelty 7/10

## Idea description

Constrain a neural policy or recurrent dynamics model to be order-preserving, then construct upper and lower abstract transitions by evaluating monotone maps at opposite corners of each state-action cell. Train with a loss that rewards the upper abstraction for reaching safe target cells and the lower abstraction for avoiding unsafe cells, while reporting the undecided gap as a quantitative certificate.

## Mathematical statement

Let x be the state, u the action, and Fθ(x,u) a learned one-step dynamics map. The componentwise order is x ≤ y when xi ≤ yi for every coordinate. Assume Fθ is monotone: x ≤ y and u ≤ v imply Fθ(x,u) ≤ Fθ(y,v). For a cell C=[l,r] and action cell A=[a,b], monotonicity gives Fθ(l,a) ≤ Fθ(x,u) ≤ Fθ(r,b) for every (x,u) in C × A. Quantize with resolution h using Qh(z)i = h floor(zi/h). The lower successor is T−(C,A)=Qh(Fθ(l,a)), and the upper successor is T+(C,A)=Qh(Fθ(r,b)+h1), where 1 is the all-ones vector. Reachability sets satisfy R−k ⊆ Rk ⊆ R+k. If Fθ is L-Lipschitz, the one-step enclosure width is bounded by approximately L times the cell diameter plus 2h, before adding model-residual inflation.

## Key formulas

- x ≤ y iff xi ≤ yi for every i, and x ≤ y, u ≤ v implies Fθ(x,u) ≤ Fθ(y,v).
- Fθ(l,a) ≤ Fθ(x,u) ≤ Fθ(r,b) for every (x,u) in [l,r] × [a,b].
- T−(C,A) = Qh(Fθ(l,a)), and T+(C,A) = Qh(Fθ(r,b) + h1), where Qh(z)i = h floor(zi/h).
- R−k ⊆ Rk ⊆ R+k, and width(T+) is approximately bounded by L diam(C × A) + 2h.

## Implementation notes

1. Integration point: apply the method to a small deterministic RL environment, a learned world model, or an RNN dynamics predictor. Parameterize scalar outputs of Fθ with a monotone MLP using nonnegative weights, such as softplus-constrained weights. If actions have mixed effects, use separate positive and negative monotone channels and propagate their known signs. Partition the relevant state-action domain into uniform boxes of width h. 2. Pseudocode: for each cell C=[l,r] and action cell A=[a,b], compute yminus=Fθ(l,a) and yplus=Fθ(r,b); set sminus=Qh(yminus−δ) and splus=Qh(yplus+δ+h1), where δ is a coordinatewise validation residual bound. Build lower and upper transition graphs. Run reachability from initial cells for both graphs. Add λ times the upper-only volume, |Rplus minus Rminus|, to the dynamics or policy loss. 3. Computed quantities are corner evaluations, quantization, and graph reachability. The monotonicity property is guaranteed by the architecture; δ and the effective Lipschitz constant are estimated on held-out transitions. 4. First experiment: train a two-dimensional monotone MLP model on xnext=tanh(Ax+Bu) with nonnegative A and B. Compare nominal planning, ordinary interval propagation, and the paired abstraction for h=0.2, 0.1, and 0.05. The predicted signature is linear gap reduction with h until the residual floor O(δ): halving h should approximately halve the upper-lower gap before saturation. The lower abstraction should never certify a sampled unsafe trajectory as safe, and the upper abstraction should contain every held-out sampled trajectory after residual inflation.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built a monotone 2D dynamics interval-abstraction MVP with opposite-corner evaluation, quantization, residual inflation, and reachability-gap measurements. The mechanism manifested: monotonicity had zero sampled violations, no-residual samples were fully enclosed, the Lipschitz-plus-quantization bound held in all 12 fixed-cell cases, and containment reached 1.0 when delta equaled the injected residual. Resolution reduced mean width from 0.3061 at h=0.20 to 0.1398 at h=0.025, though scaling was not uniformly linear due to quantization floors.

### Mechanism check

- Verdict: Built a monotone 2D dynamics interval-abstraction MVP with opposite-corner evaluation, quantization, residual inflation, and reachability-gap measurements. The mechanism manifested: monotonicity had zero sampled violations, no-residual samples were fully enclosed, the Lipschitz-plus-quantization bound held in all 12 fixed-cell cases, and containment reached 1.0 when delta equaled the injected residual. Resolution reduced mean width from 0.3061 at h=0.20 to 0.1398 at h=0.025, though scaling was not uniformly linear due to quantization floors.
- Confidence: 8/10
- Limitations: This is deterministic toy verification rather than full neural policy or world-model training. Reachability is a small interval-area proxy, soundness is sampled rather than exhaustive, and no real RL task, learned residual estimator, FLOP comparison, or GPU experiment was tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.81695
- Paired wins: 3/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/706/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/706/bench_report.json)
- [interval_abstraction.py](https://synthcore.org/code/706/interval_abstraction.py)
- [report.md](https://synthcore.org/code/706/report.md)
- [report_bench_2026-09-02T225731.md](https://synthcore.org/code/706/report_bench_2026-09-02T225731.md)
- [results.json](https://synthcore.org/code/706/results.json)
- [Download all files as ZIP](https://synthcore.org/download/706)

## Disclaimer

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