Examples

BasicPrediction

The minimal hello-world for HypercubeESN. Demonstrates the complete pipeline on a sine wave: drive the reservoir, collect states, train the HCNN readout, and evaluate.

This is the place to start if you want to understand how the pieces fit together.

What it shows:

• ESN construction from an ESNConfig
• Warmup (wash out initial transients) and Run (collect states)
• HCNN readout training with cosine LR schedule
• R² and NRMSE evaluation on held-out test set

---

SignalClassification

Name which of four industrial process modes is active (Cruise / Chatter / Ramp / Spin-up — sine, square, triangle, chirp under the hood) from reservoir state alone. dim=8 gives strong overall ID; residual error clusters on Cruise ↔ Spin-up, with a short lock-on delay after each mode switch. A live block-stream table reports accuracy, softmax confidence, and time-to-lock.

What it shows:

• Reservoir as a feature extractor for multi-class mode ID
• HCNN native 4-class classification (softmax + cross-entropy)
• Live stream monitor: conf, TTL, LOCKED / SWITCHING / SETTLING
• Confusion matrix + early-window lock-on + mean time-to-lock

---

StreamingAnomaly

Simulates industrial process monitoring. The reservoir learns normal process behavior during a priming phase, then monitors a live stream in 200-step windows. Three anomaly types are injected — noise spike, DC drift, and frequency shift — each for 3 windows, separated by normal operation to show both detection and recovery.

What it shows:

• Batch training on historical "normal" data (priming)
• Anomaly detection via prediction error exceeding a threshold (10x baseline RMSE)
• Three distinct anomaly signatures with different RMSE ratios
• Automatic recovery without retraining as anomalies end

---

MemoryCapacity

Measures the standard Jaeger (2001) linear memory capacity. No HCNN is involved — only the raw reservoir state is read.

What it shows:

• How much short-term linear memory the hypercube reservoir holds, and how it scales with spectral radius, history_depth, leak rate, and dim

---

NARMA

Primary open-loop validator: white input u(t) → reconstruct NARMA-N y(t). One fixed config across tanh-wrapped orders 30 / 50 / 70 (best 5 of 20, test NRMSE). Full write-up: NARMA/NARMA.md.

---

Lorenz

Primary closed-loop validator: Lorenz-63 free-run via input-bank self-feedback (predicted [x, y, z, x*z] re-injected as the next drive). VPT at θ = 0.25. Full write-up: Lorenz/README.md.

---

Building

Example targets build automatically alongside the main harness; build the Release tree with the bundled toolchain, then run any target from cmake-build-release:

cmake-build-release\BasicPrediction.exe
cmake-build-release\SignalClassification.exe
cmake-build-release\StreamingAnomaly.exe
cmake-build-release\MemoryCapacity.exe
cmake-build-release\NARMA.exe
cmake-build-release\Lorenz.exe

In CLion, select the target from the run configuration dropdown (top toolbar).