Dynamic Self Stabalizing Loops

2 hours ago 1

A minimal simulation laboratory for exploring bistable workspace dynamics, self-referential models, and autotuning in neural-like systems.

miniBrain provides a compact, interactive environment to experiment with three related recurrent dynamics models:

Option A: Bistable units with a global workspace coupling.
Option B: Reflective hierarchical "why-loop" architecture.
Option C: Self-referential workspace that maintains and predicts a compressed self-model.

The lab includes background autotuning (meta-tuner NN), perturbations to maintain entropy, and instrumentation for complexity metrics (Shannon entropy, Lyapunov proxy, Lempel-Ziv complexity, mutual information).

Interactive GUI: Real-time heatmaps, phase coherence plots, rolling windows, and step counter.
Autotuning: Background meta-learning to bias parameters toward high complexity (entropy + coherence).
Headless Mode: CSV export and smoke tests for reproducibility.
Metrics: Entropy, Lyapunov proxy, LZ complexity, pairwise mutual information.
Perturbations: Irrational-time perturbations with drift to avoid resonances.
Documentation: LaTeX findings document in docs/findings.tex.

Python 3.10+
NumPy
Matplotlib
scikit-learn
PyTorch (optional, for meta-tuner; falls back to heuristic if missing)

Clone the repository:

git clone https://github.com/BrandonRaeder/miniBrain.git cd miniBrain
Create a virtual environment:

python3 -m venv .venv source .venv/bin/activate # On Windows: .venv\Scripts\activate
Install dependencies:

pip install numpy matplotlib scikit-learn torch

Run the full simulation with autotuning:

Displays 3x2 layout: heatmaps + R-phase plots for each model.
Autotune starts automatically; parameters update in real-time.
Rolling window (default 500 steps) on phase plots.

Quick validation without GUI:

python3 -c "import matplotlib; matplotlib.use('Agg'); import lab; lab.smoke_test_autotune(2.0)"

Runs autotune for 2 seconds, prints buffer size and samples.

Modify n_layers, dt, rolling_window in lab.py.
For longer runs, adjust ROLLOUT_STEPS and reward weights.

miniBrain/ ├── lab.py # Main simulation and GUI ├── models/ │ └── self.py # Self-model predictor and utilities ├── tools/ │ └── smoke_autotune.py # Lightweight headless test ├── docs/ │ └── findings.tex # LaTeX findings document ├── README.md # This file └── .venv/ # Virtual environment (not committed)