Interference Detection

When multiple optimizers operate on shared infrastructure — a power grid, a data center, a factory floor — their decisions can interfere with each other. Each optimizer acts independently, making locally optimal choices, but through the shared physical or computational medium, one optimizer's actions affect another's outcomes.

Neither optimizer knows this is happening.

godon's interference detection makes this invisible problem visible. It works by injecting a known mathematical signal (a watermark) into each optimizer's parameters, then observing whether that signal appears in another optimizer's objective values. If it does, the optimizers are interfering.

The Problem

Consider two independent optimization loops:

    ┌─────────────┐              ┌─────────────┐
    │  Optimizer A │              │  Optimizer B │
    │  (grid load) │              │  (grid cost) │
    └──────┬───────┘              └──────┬───────┘
           │                             │
           ▼                             ▼
    ┌──────────────────────────────────────────┐
    │           Shared Power Grid               │
    │    (the coupling channel)                 │
    └──────────────────────────────────────────┘

Optimizer A adjusts load distribution. Optimizer B adjusts cost parameters. Both affect the same grid. When A shifts load, it changes the conditions B is optimizing against. B's "optimal" decisions are now based on a landscape A has altered. Both optimizers degrade each other's performance, but neither can see the cause.

This isn't theoretical — it happens in any system where multiple optimization workloads share resources.

How Detection Works

The detection pipeline has three stages:

1. Watermark Injection

Each breeder (optimizer driver) injects a unique mathematical fingerprint into its trial parameters. The fingerprint is a multi-frequency signal — a sum of sinusoidal perturbations at specific periods, applied to selected parameters.

    Breeder A watermark: periods [29, 37]
    Breeder B watermark: periods [67, 71]

    (Non-overlapping periods — no shared frequencies between breeders)

The periods are drawn from a pool of prime numbers, with each breeder assigned a unique, non-overlapping set. This ensures that any spectral power at Breeder A's frequencies in Breeder B's observations can only come from interference — not from Breeder B's own watermark.

2. Observation and Alignment

The observer collects trial data from both breeders, aligning them by timestamp. After self-subtraction (removing each breeder's own watermark contribution from its objective values), the residual signal contains only external influences.

3. Spectral Detection

The observer performs an FFT (Fast Fourier Transform) on the aligned residual signal, measuring spectral power at the sender's known watermark frequencies. A permutation test (5000 random shuffles) determines whether the observed spectral power is statistically significant:

Null hypothesis: the residual signal has no structure at the sender's frequencies
Test statistic: combined spectral power at the sender's periods
p-value: fraction of random shuffles that produce equal or greater power
Detection threshold: p < 0.05

The permutation test is non-parametric — no assumptions about the noise distribution, the optimizer's dynamics, or the coupling channel's characteristics.

Validation Results

The detection was validated on a microgrid optimization scenario with two breeders optimizing four objectives (load balance, renewable utilization, grid stability, and energy cost). The coupling factor was varied from 0.0 (no coupling) to 0.9 (strong coupling).

Coupling	obj0	obj1	obj2	obj3 (energy cost)
0.0	clean (p=0.90)	clean (p=1.00)	clean (p=0.27)	clean (p=0.99)
0.1	detected (p=0.005)	clean (p=0.61)	detected (p=0.0002)	clean (p=0.73)
0.5	detected (p=0.005)	detected (p=0.034)	detected (p=0.004)	clean (p=0.74)
0.9	detected (p=0.0004)	detected (p=0.0002)	detected (p=0.0002)	clean (p=0.88)

Key observations:

Zero false positives: at coupling=0.0, no objectives are falsely flagged. The non-overlapping period assignment eliminates spectral contamination between breeders.
Monotonic detection: as coupling strength increases, more objectives show detectable interference, and detection confidence (lower p-values) increases.
Objective-specific sensitivity: energy cost (obj3) shows no interference at any coupling level — this objective is not affected by the coupling channel in the microgrid scenario.
Spectral power scales with coupling: the actual measured power at watermark frequencies increases proportionally with coupling strength, providing a path toward intensity measurement.

Design Principles

Non-overlapping frequency assignment: Each breeder uses unique prime periods from a shared pool. With 12 primes, up to 6 breeders can run simultaneously with 2 periods each, guaranteed zero frequency overlap.

No distributional assumptions: The permutation test compares the observed spectral power against random shuffles of the same data. It works regardless of the optimizer's dynamics, the noise characteristics, or the coupling channel's properties.

Scenario-agnostic: The watermark encoding (MultiFrequencyMultiParam) accepts arbitrary parameters and periods. The detection reads watermark metadata from trial records. No scenario-specific code in the detection pipeline.

Non-invasive: The watermark amplitudes are small perturbations within the optimizer's parameter space. They don't alter the optimization's convergence behavior — they're embedded in the existing search process.

Outlook

Interference detection is the sensing layer — the first step toward making multi-optimizer systems work reliably. The detection result (present/absent) opens the door to a much richer understanding of how optimizers interact through shared systems.

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, the question becomes: who interferes with whom, through what? The pairwise detection results between all breeders form an interference graph — a directed, weighted graph where nodes are optimizers, edges are interference channels, and weights are intensity. This topology reveals the structure of the system: clusters of tightly-coupled optimizers, isolated components, and bottleneck resources that concentrate interference.

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

In this example, A heavily interferes with B through shared infrastructure, while C and D are essentially independent. The topology tells you where to focus decoupling efforts.

Per-Parameter Tracing

Not all parameters contribute equally to interference. By watermarking individual parameters and analyzing which ones propagate through the coupling channel, the detection can trace interference to specific configuration seams — the exact points where one optimizer's decisions leak into another's domain. This enables targeted decoupling: instead of isolating entire optimizers, you can adjust specific parameters or add constraints at the seams.

Permeating Configuration

As interference understanding deepens, the configuration of the optimization system itself can be adapted. If the interference topology shows that two optimizers conflict through a specific resource, that resource's configuration can be partitioned, scheduled, or isolated. The detection doesn't just observe the problem — it informs the system's architecture. Optimization becomes aware of its own multi-agent context and reconfigures at the seams.

Temporal Dynamics

Interference isn't static. As optimizers converge, their interference patterns shift. A resource that was heavily contested early in optimization may become stable later. Temporal tracking of interference intensity reveals these dynamics, enabling adaptive strategies: coordinate during high-interference phases, run independently during low-interference phases.

Alternative Encoding Schemes

The current approach uses sinusoidal watermarks detected via FFT — but this is one point in a larger design space. Other encoding and detection schemes may offer advantages in different scenarios: spread-spectrum codes for noise resilience, chirp signals for time-varying channels, or wavelet-based encodings for non-stationary interference. The optimal scheme depends on the coupling channel's characteristics, the number of optimizers, and the available trial budget. Studying alternative encodings is an open research direction.

Cross-Scenario Generalization

The detection framework is scenario-agnostic — the watermark encoding and spectral analysis make no assumptions about what's being optimized. Validating across diverse scenarios (power grids, data centers, supply chains, chemical plants) establishes the generality of the approach and reveals which interference patterns are universal vs. domain-specific.

Interference Detection