Chaos-Evolve V2 x Nero-Quantizer Core
An Empirical Investigation into the Quantization Robustness of Evolutionarily Developed Neural Networks.
Built entirely from scratch in NumPy with zero external machine learning frameworks.
Executive Summary
Can a neural network optimized via non-gradient stochastic search (Neuroevolution) survive aggressive sub-8-bit quantization?
We simulated a continuous evolutionary ecosystem (Chaos-Evolve V2) where agents trained via crossover and mutation strategies learned obstacle-avoidance maneuvers. We then subjected the elite champion to a tight, asymmetric, block-wise 4-bit linear quantizer (Nero-Quantizer Core), compressing floating-point weights down to signed INT4 integer boundaries ($[-8, 7]$).
The Finding: The compressed 4-bit model did not experience performance degradation. Instead, it achieved a 13.22% fitness efficiency boost, registering a total Intelligence Retention Rate of 113.22%.
Empirical Metrics
| Benchmark Metric | FP32 Base Precision | INT4 Block-Wise Quantized |
|---|---|---|
| Elite Agent Fitness | 538.63 | 609.83 |
| Quantization Noise (MSE) | 0.000000 (Ref) | 0.002113 |
| Memory Reduction (Theoretical) | 1x (Reference) | ~8x Footprint Compression |
| Intelligence Retention Rate | 100% | 113.22% |
Core Architecture & Mathematical Engine
1. Block-Wise Asymmetric Quantization
Standard per-tensor quantization methods drop precision entirely when exposed to high-magnitude parameter variations (outliers). To prevent matrix variance corruption, we designed a localized block-wise mapping pipeline operating at a micro-block interval ($Block_Size = 4$).
During targeted outlier stress testing, baseline per-tensor quantization degraded significantly to an MSE of 0.250919. Our custom block-wise implementation contained the mathematical distortion locally, proving 118.8x more accurate with a minimal global MSE of 0.002113.
2. Theoretical Breakdown: Why Did INT4 Outperform FP32?
- Quantization as a Low-Pass Filter: Stochastic evolutionary paths can introduce behavioral jitter or high-frequency floating-point noise into parameter configurations during rapid mutations. The strict integer boundaries of the 4-bit engine (
roundandclipgates) smoothed out these micro-oscillations, regularizing the network and yielding a cleaner spatial navigation path. - Granular Scaling: Restricting the scale constraints to tiny block windows ensured high localized precision, stabilizing the agent's forward policy executions across unknown environment resets.
Getting Started
Installation
Ensure you have the core scientific computing layer installed:
pip install numpy