evolino-pocket-lab

Evolino Pocket

Evolino Pocket retests the hybrid idea behind Evolino with modern tooling: evolution shapes a nonlinear recurrent feature generator, then ridge regression solves its output layer exactly. The model predicts 12 future points of a chaotic delayed Mackey–Glass series.

A same-size random recurrent feature generator receives the identical optimal linear readout. Three independent mutation searches measure whether evolution adds value beyond the reservoir-style control.

This is an Evolino-inspired controlled study, not a full reproduction of the original LSTM-based algorithm.

Verified result

Across three independent mutation searches, the 264-parameter evolved recurrent models reached 0.1731 mean held-out RMSE over twelve direct forecast horizons. Same-size random recurrent controls, each given the identical closed-form ridge readout, reached 0.3818 RMSE. Evolution therefore reduced multihorizon error by 54.7%.

For the first saved seed, horizon-12 RMSE fell from 0.6078 to 0.1775. The one-step result was slightly worse for the evolved model, 0.0129 versus 0.0086, so the measured advantage is specifically in longer-horizon recurrent features.

uv run python projects/evolino-pocket/train.py
uv run pytest tests/test_evolino_pocket.py

Hosted showcase

This free static Space preserves the complete original Gradio source, trained artifacts, evaluation files, and local launch requirements. Hugging Face now requires PRO for CPU-backed Gradio hosting, so the public landing page is static while the checked-in app.py remains the authoritative runnable demo.