AI-Powered-Smart-Routing-Predicting-and-Avoiding-Network-Bottlenecks

---

📘 Overview

Traditional routing algorithms such as Dijkstra’s or A* are optimal for static graphs but fail under dynamic congestion and spatio-temporal bottlenecks.
This project introduces an AI-powered routing system that learns to avoid congestion dynamically through intelligent bottleneck detection, prediction, and adaptive routing.

The framework leverages:

• 🧩 GCN/GAT for topological bottleneck detection (spatial domain)
• ⏱ LSTM for time-series bottleneck forecasting (temporal domain)
• 🎯 Curriculum-based D-DQN for adaptive routing decisions
• ⚙️ Integration with classical SSSP (Dijkstra) to maintain stability and interpretability

---

📊 Dataset Information

Dataset Name: 5G Traffic Datasets
Author: Daegyeom Kim
Platform: Kaggle
Year: 2023

Description

A large-scale real-world 5G traffic dataset captured directly from mobile terminals using PCAPdroid and G-NetTrack Pro without specialized hardware.
It includes 328 hours of timestamped 5G packet data (CSV format) recorded from major video streaming, live streaming, video conferencing, metaverse, and gaming applications — ideal for network analysis, GNN modeling, and deep learning-based routing.

Key Highlights

• Captured on Samsung Galaxy A90 5G (Qualcomm Snapdragon X50 modem)
• Traffic types: Netflix, YouTube Live, Zoom, Roblox, MS Teams, etc.
• Each record includes packet headers, source/destination IPs, timestamps, and payload info
• Suitable for GANs, LSTMs, and spatio-temporal ML models

Structure

• packets.csv — Timestamped packet-level 5G traffic data
• applications.csv — Application-wise traffic statistics
• meta_info.json — Device and capture configuration details

📥 To Use:

1. Download from Kaggle → 5G Traffic Datasets
2. Extract to the /data/5g_traffic_datasets/ directory of this repository
3. Update dataset path in your configuration file if required

---

🧩 Key Features

Feature	Description
🧠 Hybrid AI Routing	Combines classical SSSP with AI for real-time decision making
🕸 Graph Neural Networks	GCN & GAT for spatial bottleneck localization
⏱ Temporal Prediction	LSTM-based spatio-temporal bottleneck forecasting
🎮 Reinforcement Learning	D-DQN with dueling architecture & curriculum learning
📈 Massive Performance Gain	7×–53× improvements in success, throughput, and latency
⚡ 5G Network Simulation	Tested on realistic 5G traffic topologies

---

🏗️ System Architecture

         ┌──────────────────────────────┐
         │     Network Topology G(V,E)  │
         └──────────────┬───────────────┘
                        │
        ┌───────────────┴───────────────┐
        │         Bottleneck Detection  │
        │           (Spatial)           │
        │                               │
        │    ┌──────────────┐           │
        │    │   GCN / GAT  │ → Spatial │
        │    └──────────────┘           │
        └──────────────┬────────────────┘
                       │
        ┌──────────────┴───────────────┐
        │        Prediction (Temporal) │
        │                              │
        │    ┌──────────────┐          │
        │    │    LSTM      │ → Temporal│
        │    └──────────────┘          │
        └──────────────┬────────────────┘
                       │
        ┌──────────────┴───────────────┐
        │      Routing Decision        │
        │                              │
        │  ┌────────────────────────┐  │
        │  │  Curriculum D-DQN      │ → Adaptive │
        │  └────────────────────────┘  │
        └──────────────────────────────┘

---

⚙️ Implementation Details

Component	Details
Dataset	5G traffic dataset (10–200 nodes/snapshot)
Graph Features	12 node features (degree, load, latency, etc.)
Loss Function	Focal loss (for imbalance), Cross-Entropy
Optimizer	AdamW with cosine annealing
Learning Rate	1e-4 → 1e-6 (decay 0.9995)
Batch Size	128
Episodes	5,000 with 200 warmup
Frameworks	PyTorch, PyTorch Geometric, NetworkX

---

📊 Experimental Results

Algorithm	Success (%)	Latency (ms)	Throughput (paths/s)	Bottleneck Hits	Risk-Free (%)
GCN-DDQN	🥇 99.5	0.04	3,908	0.60	45.7
GAT-DDQN	98.0	0.27	4,422	0.53	50.0
Dijkstra	98.5	0.07	8,595	0.67	41.1
A*	98.0	0.05	10,697	0.71	35.7
DRL-GNN [2]	12.5	21.15	73.2	0.80	36.0
GAT-RL [3]	18.5	5.95	106.1	0.73	35.1

Highlights:

• 🧩 99% spatial detection (GCN)
• 🕒 100% temporal prediction (GAT-LSTM)
• ⚡ 60× throughput vs DRL-GNN
• 🔄 22–529× latency reduction

---

🧩 Methodological Workflow

1. Graph Representation:
   • Network → Directed Graph (Nodes = Routers, Edges = Links)
2. Spatial Detection (GCN/GAT):
   • Detect topological bottlenecks using learned embeddings
3. Temporal Forecasting (LSTM):
   • Predict future congestion probabilities (1–2 steps ahead)
4. Adaptive Routing (D-DQN):
   • Optimize paths using curriculum-guided double Q-learning
5. Risk-Aware Path Update:
   • Integrate spatial and temporal risk penalties for decision-making

---

🧪 Evaluation Metrics

Category	Metrics
Routing	Success rate, path cost, latency, throughput
Risk	Bottleneck hits, spatial/temporal risk scores
Computation	FLOPs, runtime, decision latency

---

🧮 Reward and Learning Strategy

Reward Function Components

• ✅ Reaching destination
• 🚫 Bottleneck avoidance
• ⚡ Path efficiency
• 🔄 Step penalty
• 🎯 Risk minimization (spatial & temporal)

---