Open_Research

Open Research

Interference detection works reliably on linear coupling channels. The open research frontier is extending detection to nonlinear, cascaded, and non-stationary channels — and moving from binary detection to quantitative understanding.

Methods Evaluated

Over the course of development, several detection approaches were evaluated. The early methods were tested before the watermarking system matured — the original approach added noise to the sampler output rather than producing a clean sinusoidal signal by modulating the parameter directly. This means early failures are confounded with immature signaling. The methods have not been re-evaluated with the current watermarking.

ABA (Pause One Breeder)

Stop one optimizer and observe whether the other's behavior changes. Conceptually simple but operationally expensive — it requires pausing optimization campaigns, which defeats the purpose of continuous online detection. Each ABA cycle produces one data point. Compared to watermark-based detection, which produces a signal every trial without interrupting optimization, ABA is sluggish and impractical for production use.

Statistical Methods (Cross-Correlation, Granger Causality, Mutual Information, Transfer Entropy, Convergent Cross Mapping, HSIC)

A range of statistical and information-theoretic methods were evaluated for detecting coupling between optimizer time series. These are passive approaches — they don't require an injected signal. None produced reliable results on the greenhouse bench.

Important caveat: these were evaluated with early signaling that added noise to the sampler rather than modulating parameters with a clean sinusoidal. The current watermarking produces a well-defined spectral signature. Whether these passive methods would perform better with the improved signal — or whether they could detect coupling even without watermarking — remains an open question.

Lock-In Detection

Amplitude-phase detection at the known watermark frequency. Requires the optimizer to converge so that noise_std drops below the signal amplitude. Multi-objective Pareto optimization never converges — the optimizer keeps exploring indefinitely, so noise_std never stabilizes. Lock-in detection is effective for single-objective optimization where convergence occurs, but fundamentally incompatible with Pareto front exploration.

FFT Spectral Detection

The current method. Goertzel algorithm at known watermark frequencies with Hann windowing, noise floor estimation from neighboring bins, and permutation testing for significance. Works reliably on linear additive channels (microgrid bench, all coupling strengths 0.0-0.9). On deeply nonlinear channels, the signal is too heavily distorted by the time it reaches the objectives for FFT to extract it at 200-300 trial budgets.

Greenhouse Nonlinear Channels

The greenhouse bench represents the hardest detection case. The coupling signal traverses 6+ nonlinear stages before reaching the objectives, with an SNR of ~0.002. Solving this would demonstrate that interference detection works across the full complexity spectrum, not just on linear channels.

The coupling signal passes through:

Coupling delta (tiny perturbation)
Thermal inertia (×0.01 per tick)
Zone physics (mixing, diffusion)
Multiplicative growth model with dead zones (zero derivative in optimal range)
Phase-dependent sensitivity (1× → 3× jump between growth phases)
Irreversible damage (permanent attenuation past thresholds)
Sensor noise
Receiver's own exploration variance (500× larger than coupling signal)

Every method tested so far has not produced reliable results at 200-300 trials. The question is whether higher trial counts or different detection strategies can overcome the signal death.

For a visual comparison of signal propagation through the microgrid (linear) vs greenhouse (nonlinear) channels, see the Signal Death Diagram.

Intermediate State Detection

Measuring interference at raw sensor readings (zone temperatures, CO2 levels) before the growth model, where the signal is less distorted. The coupling signal may be detectable at intermediate points in the channel even when it's invisible at the final objectives.

This requires targets to expose raw sensor channels in addition to optimization objectives — a natural extension of the target contract. The observer would analyze these raw channels using the same spectral pipeline, but on data that hasn't passed through the growth model's dead zones.

Dedicated Analysis Phases

Holding the receiver optimizer still while the sender probes aggressively. This eliminates the 500× exploration noise that currently overwhelms the coupling signal. The trade-off is paused optimization on the receiver side during analysis.

A dedicated analysis phase could be triggered automatically when the observer needs more signal power — the receiver pauses its optimization, the sender injects stronger watermarks, and the observer collects a clean signal. After analysis completes, both optimizers resume.

CDMA Encoding for Scale

The current FDMA approach (prime-numbered periods) scales to ~20-30 breeders with a few hundred trials. For significantly larger deployments, spread-spectrum approaches like direct-sequence CDMA with orthogonal Walsh-Hadamard codes would offer better noise resilience.

CDMA requires chip sequences of length 50-100+ per parameter to achieve reliable orthogonality. With only 200-300 total trials, codes are too short to correlate reliably. This becomes viable with thousands of trials in production deployments — a natural transition as optimization campaigns grow longer.

Impulse-Based Detection

All methods tested so far send continuous signals (sines, codes) and rely on frequency content surviving the channel. Nonlinear transforms distort frequencies. Dead zones kill small perturbations. Non-stationary channels shift characteristics mid-signal.

An alternative: send discrete extreme-value impulses instead of continuous tones. Push parameters to the edge of (or beyond) the safe operating range for a single trial. The impulse does not rely on frequency content — it relies on timing and amplitude. You know when the impulse was sent. You look for a response in the receiver's objectives after that timestamp.

Detection method: split the receiver's objective values into post-impulse windows vs baseline. Run a rank-sum test (Mann-Whitney U) or Kolmogorov-Smirnov. No spectral analysis. No frequency preservation assumptions. No stationarity assumptions.

This is not a new invention — it is the standard approach in fields that deal with hostile channels:

Seismology: earthquake impulses through heterogeneous rock
Active sonar: acoustic pings through nonlinear ocean layers
Ground-penetrating radar: electromagnetic impulses through soil
Medical percussion: tapping to detect fluid in tissue
Network traceroute: ICMP packets through congested routers
Materials ultrasound: acoustic pulses to detect cracks in metal

In every case, continuous signals fail because the channel is hostile. Impulses succeed because they don't need the channel to preserve frequency content — only to propagate a perturbation.

Implications for the channel taxonomy:

Channel Type	Current Method	Impulse Method
Linear additive	FFT + permutation (works)	Impulse + distribution test (also works, overkill)
Nonlinear, measurable intermediate	Spectral on sensors (research)	Impulse + distribution test (promising)
Deeply nonlinear cascaded	No reliable method	Impulse + distribution test (untested, theoretically strong)
Non-stationary	Phase-aware methods (research)	Impulse + distribution test (timing-based, non-stationarity not a blocker)

The strategy becomes: unknown channel -> impulse probe first. If it's linear, downgrade to cheaper sine watermark. One method with a strength dial, instead of four channel types with four methods.

Implications for architecture: impulses are inherently disruptive. They don't mix with optimization. This forces the separation that the current code ducks: detection and optimization are different jobs. Options are a dedicated probe agent (exists only to send impulses) or a dedicated phase (existing breeder switches to impulse mode temporarily).

Not yet tested. The greenhouse bench is the first target.

Interference Intensity Measurement

Detection answers "is there interference?" The next question is "how much?" The spectral power at watermark frequencies scales with coupling strength — this relationship can be calibrated into an interference intensity metric.

An optimizer running against a power grid at 10% interference and one at 90% interference face fundamentally different problems. Intensity measurement gives operators the information to decide whether to act.

Interference Topology

With three or more optimizers, pairwise detection results form an interference graph — a directed, weighted graph where nodes are optimizers, edges are interference channels, and weights are intensity.

    ┌──────────┐  0.7  ┌──────────┐
    │ Breeder A │──────▶│ Breeder B │
    └──────────┘       └──────────┘
         │                  │
      0.2│               0.1│
         ▼                  ▼
    ┌──────────┐  0.0  ┌──────────┐
    │ Breeder C │──────▶│ Breeder D │
    └──────────┘       └──────────┘

This topology reveals clusters of tightly-coupled optimizers, isolated components, and bottleneck resources that concentrate interference. The detection doesn't just observe the problem — it maps the structure.

Detection Difficulty Variables

Coupling factor is not the only variable that determines whether detection succeeds. Multiple factors interact:

Variable	Effect on Detection	Range
Coupling strength	How much signal transfers	0.001 (barely visible) to 0.9 (trivial)
Signal-to-noise ratio	Receiver's own exploration drowns weak coupling	0.002 (greenhouse) to 1.0+ (microgrid)
Number of nonlinear transforms	Each stage distorts or attenuates the signal	1 (linear) to 6+ (cascaded)
Transform type	Additive (benign), multiplicative (scaling), dead zone (destructive)	Varies per channel
Dimensionality	More parameters and objectives increase noise floor	2-3 (simple) to 20+ (complex)
Stationarity	Coupling characteristics change over time	Stable to highly non-stationary
Watermark design	Frequency choice, amplitude, number of frequencies	Configurable
Receiver exploration variance	Wild exploration drowns weak coupling signals	1× to 500× signal strength

The space is combinatorial but not infinite — roughly 5-6 key variables with 2-3 levels each. Characterizing where a bench falls in this space is necessary for choosing the right detection method.

Channel Classification

An automated assessment that characterizes the coupling channel before choosing a detection method. Inject calibration signals, measure what comes out, classify:

Classification	Detection Strategy
Linear additive	FFT + permutation test (current)
Weakly nonlinear	Adapted spectral methods (research phase)
Strongly nonlinear	No reliable method yet
Non-stationary	Phase-aware methods (research phase)

Open design question: where does the classifier live? Part of the breeder (startup calibration), a separate probe unit, or part of the observer? Depends on whether channels are stable or shift over time — non-stationary channels suggest ongoing assessment is needed rather than one-time calibration.

Scaling Beyond Two Breeders

Validated at 6-breeder scale (microgrid bench, coupling_factor=0.9). Pairwise FFT + Rayleigh detection generalizes to multi-breeder environments — 20 pairwise tests with high specificity. Multi-frequency composite watermarks with unique prime period pairs successfully separate overlapping signals from 6 independent optimizers.

Remaining open questions: - Does interference cascade transitively (A->B->C)? - Does the presence of multiple interferers change detection reliability for individual pairs? - What trial budgets are needed to map N breeders for N >> 6? - Does pairwise detection hold at weaker coupling factors (below 0.5) with many simultaneous interferers?

Transitive Topology Discovery

Each pairwise detection reveals one edge in a coupling graph. With three or more breeders, accumulated detections form a partial topology — not assumed, not modeled, but discovered from the system's own dynamics.

The hypothesis: as more breeders explore and the observer detects more edges, the topology of how the system is internally connected emerges. This is proven for single edges on linear channels and demonstrated at 6-breeder scale (20 pairwise detections forming a partial interference graph). Whether it scales to reveal the full coupling structure of larger systems (N >> 6) is an open question.

This connects to interference topology (above) but goes further — not just mapping who interferes with whom, but discovering the hidden structure of the system through transitive signal propagation.

Meta-Optimization of Detection

The detection pipeline itself has tunable parameters: watermark frequency, amplitude, number of frequencies, trial budget, analysis window, permutation count. Choosing optimal values for these parameters given the channel characteristics is itself an optimization problem — the system optimizing its own sensing strategy.

This connects to the broader meta-optimization direction: godon optimizing its own parameters (observer sensitivity, guardrail thresholds, watermark configuration) using the same optimization infrastructure it applies to external systems.