📊 Edge–Cloud Offloading using CNN-LSTM-Attention + RL

This project implements an intelligent edge vs cloud offloading system using:

• Deep Learning (CNN + LSTM + Attention)
• SLA-aware cost optimization
• Reinforcement Learning (RL)

The objective is to minimize total cost (SLA + latency) by deciding whether to execute tasks on the edge or cloud.

---

🚀 Project Pipeline

1. Data Processing

• Uses Telecom Italia Milano Grid Dataset
• Extracts:
  • square_id
  • timestamp
  • internet traffic
• Selects top M = 49 grid cells
• Normalizes traffic data
• Creates time windows of size N = 6

---

2. Model Architecture

Edge Model

• CNN + LSTM + Attention
• Outputs:
  • yL → Edge prediction
  • z → latent features

Cloud Model

• Fully connected network
• Uses latent representation z
• Outputs:
  • yR → Cloud prediction

---

3. Training Phases

Phase 1: Edge Training

Train only edge model using SLA loss

Phase 2: Cloud Warmup

Train cloud model using edge features

Phase 3: Joint Training

Train both models together

Loss Function: Loss = 0.25MSE(yL, y) + 0.25MSE(yR, y) + alpha * SLA_loss

---

4. RL-Based Offloading

State: [ (CL - CR), RTT ]

Actions: 0 → Edge
1 → Cloud

Reward:

• Penalizes SLA cost
• Penalizes cloud latency

---

5. Metrics

• Edge / Cloud percentage
• Total cost
• SLA cost
• Latency
• Benefit = CL - CR

---

📂 Project Structure

project/ │── train.py
│── models.py
│── rl_agent.py
│── utils.py
│── dataset/
│── results/

---

⚙️ Installation

1. Clone Repository

git clone
cd project

2. Create Virtual Environment

python -m venv venv

Activate:

• Linux/Mac: source venv/bin/activate
• Windows: venv\Scripts\activate

3. Install Dependencies

pip install torch numpy pandas matplotlib

---

📁 Dataset Setup

1. Create dataset folder: dataset/

2. Add Milano dataset files: sms-call-internet-mi-2013-11-01.txt
   ...
   sms-call-internet-mi-2013-11-30.txt

3. Ensure path in code: data_path = "./dataset"

---

▶️ How to Run

Run the full pipeline: python train.py

---

🖥️ GPU Support

The code automatically uses GPU if available: device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

Check GPU: python -c "import torch; print(torch.cuda.is_available())"

---

📊 Output

All outputs are saved in: results/

Generated files:

• plot1.png → RTT vs Edge/Cloud decisions
• plot_metrics(plot2).png → Metrics comparison
• plot3.png → Spatial attention
• plot4.png → Temporal attention
• offloading_decisions.txt
• edge_cost.txt
• cloud_cost.txt

---

📈 Key Observations

• Low RTT → Cloud preferred
• High RTT → Edge preferred
• RL learns adaptive offloading
• Attention highlights important regions and time steps

---

🔧 Important Parameters

M = 49
N = 6
alpha = 0.3
w = 0.3
lambda_rtt = 0.015
rl_epochs = 350

---

🧠 Future Work

• Real-time data integration
• Advanced RL algorithms (DQN, PPO)
• Multi-agent systems
• Dynamic network adaptation

---

👨‍💻 Author

122cs0058,122ad0031