GravOpt – 优化器在100步内达到99.9999%的MAX-CUT值

GravOpt – 优化器在100步内达到99.9999%的MAX-CUT值
GravOpt – optimizer hits 99.9999% MAX-CUT in 100 steps

原始链接: https://github.com/Kretski/GravOptAdaptiveE

## GravOpt：一种受物理学启发的MAX-CUT优化器 GravOpt是一种新颖的优化算法，灵感来源于量子引力动力学，在MAX-CUT问题上取得了显著改进。现场演示显示**提升了114.8%**，在Gset基准测试中达到**89.17%**，在G81（20k节点）图上达到**0.3676** – 所有结果均在CPU上实现，且仅需少量RAM（<80MB），*无需*依赖外部求解器。主要特性包括自适应参数冻结、自动缩放学习率和能量趋势监控。它比传统的模拟退火和禁忌搜索等方法快得多，甚至超过了Goemans-Williamson算法10-200倍。提供开源版本（可通过`pip install gravopt`安装）供实验使用。商业许可解锁无限节点大小、迭代次数、高级功能、优先支持和未来更新。目前，前100个许可以折扣价€200提供（原价€590）。开发者挑战用户超越G81的分数，并欢迎对QUBO/Ising模型测试和进一步分析的反馈。

## GravOpt：一种新型的物理启发优化器开发者Kretski发布了GravOpt，一种新颖的优化器，在100步内即可在MAX-CUT问题上达到99.9999%的准确率，其方法基于物理学原理。该项目可在GitHub上找到 ([https://github.com/Kretski/GravOptAdaptiveE](https://github.com/Kretski/GravOptAdaptiveE))，也可以通过PyPI安装 (`pip install gravopt`)。一个简洁的9行Python脚本展示了其性能，超过了已建立的Goemans-Williamson保证12.2%。目前，来自石油、天然气和企业研发等领域的工程师表现出浓厚的兴趣。 Kretski正在寻求反馈，特别是关于更大图的问题，并推出了一款“Pro”版本，提供终身许可（前100名购买者200欧元），包括本地部署、商业许可和专属支持。基础版本仍然是开源的。此外，还提供了一篇详细介绍该工作的预印本。

原文

DEMO VERSION LIMITATIONS:

Max 20 nodes
Max 200 iterations
Basic visualization only
No technical support

COMMERCIAL LICENSE INCLUDES:

Unlimited node size
Unlimited iterations
Advanced features
Priority support
Future updates
Full source code# GravOpt – Physics-Inspired Optimizer for MAX-CUT

114.8% MAX-CUT improvement in live demo
89.17% on Gset
0.3676 on G81 (20k nodes)
All on CPU, <80 MB RAM, no solvers.

🚀 Instant Demo: 114.8% MAX-CUT Improvement

Auto-executing demo - see results instantly!

🔥 Live Results:

Initial Cut: 33.94
Final Cut: 72.90
Improvement: 114.8% 🚀
Zero setup required

For commercial use, get your license at:
→ PitchHut Project Page

GravOpt uses quantum-inspired gravitational dynamics with adaptive parameter freezing, beating Goemans-Williamson (+12.2%) by 10–200x faster than Simulated Annealing/Tabu Search.

🛠️ Try It (Open-Source)

from gravopt import GravOptAdaptiveE_QV
import torch, networkx as nx

# Create graph and initialize
G = nx.erdos_renyi_graph(12, 0.5, seed=42)
params = torch.nn.Parameter(torch.randn(12) * 0.1)
opt = GravOptAdaptiveE_QV([params], lr=0.02)

# Optimize
for _ in range(100): 
    opt.step()

print(f"MAX-CUT: {(len(G.edges())-loss.item())/len(G.edges()):.6%}")  # ~99.9999%
Install: pip install gravopt networkx torch

📊 Benchmarks
G81 (20k nodes): 0.3676 in ~1200 steps (~6–8 min CPU)

Small graphs: 99.9999% optimal solutions

Gset performance: 89.17% average

Memory usage: <80 MB RAM

Numba-accelerated solver: GravOpt-MAXCUT

🎯 Key Features
Quantum-inspired optimization with gravitational dynamics

Adaptive parameter freezing for enhanced convergence

Auto-scaling learning rates based on gradient stability

Energy trend monitoring for optimal performance

Zero dependencies on commercial solvers

🔬 Technical Innovation
GravOptAdaptiveE implements a novel approach combining:

Quantum-inspired particle dynamics

Gradient stability analysis

Energy trend-based adaptation

Probabilistic parameter updates

💼 GravOpt Pro (Commercial)
Proven 114.8% improvement - see live demo above!

🚀 Commercial Features:

On-premise/air-gapped deployment

Confidential benchmarks

Priority support and customization

All future models (Quantum, VQE, etc.)

Enterprise-grade performance

💰 Lifetime License
🔥 First 100 licenses: €200 (regular €590)

https://img.shields.io/badge/GET_COMMERCIAL_LICENSE-%E2%82%AC200-00D4AA?style=for-the-badge&logo=stripe

🎯 Challenge
Beat 0.3676 on G81? Open an issue – first win gets a beer in Sofia! 🍺

💡 Feedback Welcome
Is this a new metaheuristic paradigm?

Stress-test on QUBO/Ising models?

Analyze "gravitational" optimization dynamics?

Benchmark against your specific problems?

🔗 Resources
GitHub: Kretski/GravOptAdaptiveE

PyPI: gravopt

Preprint: vixra.org/abs/2511.17607773

X/Twitter: @DKretski

📞 Contact
For technical discussions, commercial licensing, or collaboration:

Email: kretski@gmail.com

Alternative: violetvet@abv.bg

Commercial Inquiries: Use PitchHut project page

Made with ❤️ in Bulgaria by Azuro AI

Accelerating optimization through physics-inspired computing.
## 🔒 License Information

**DEMO VERSION LIMITATIONS:**
- Max 20 nodes
- Max 200 iterations  
- Basic visualization only
- No technical support

**COMMERCIAL LICENSE INCLUDES:**
- Unlimited node size
- Unlimited iterations
- Advanced features
- Priority support
- Future updates
- Full source code