This presentation outlines a research paper (ID 488) on optimizing traffic routing in Mobile Edge Computing (MEC) networks using Deep Reinforcement Learning (DRL). The core problem addressed is the selection of optimal data transmission paths in dynamic network environments to minimize both end-to-end delay and overall bandwidth utilization. The proposed solution employs a Deep Q-Network (DQN) algorithm, framing the routing challenge as a sequential decision-making problem. The model incorporates a comprehensive state representation using both link features (bandwidth) and node features (storage, computation capacity) from a real-world dataset of network topologies in Milan, Italy. A custom reward function is engineered to guide the learning agent toward the desired optimization goals. Experimental results show that the DRL agent successfully learns an effective routing policy, achieving a stable and positive cumulative reward after 1,000 training episodes, demonstrating its superiority over traditional, static routing methods.
Deep Reinforcement Learning for Smart Network Traffic Routing
0:00:14 - Research Focus: The presentation introduces a method for managing network traffic on Mobile Edge Networks (MEC) using Deep Reinforcement Learning (DRL).
0:00:59 - The Core Problem: The main challenge is to find the optimal path for data transmission in dynamic network environments. This is not always the shortest path but one that best utilizes resources like bandwidth and storage.
0:01:51 - Key Contributions: The project's two primary goals are to develop a DRL approach that minimizes link bandwidth utilization and reduces the average end-to-end delay compared to traditional hop-count-based routing methods.
0:02:23 - Justification from Prior Work: The presenter reviews previous studies, noting their limitations, such as focusing on traditional networks or using algorithms (like Bellman-Ford and Dijkstra) that cannot learn or adapt to dynamic network conditions with multiple parameters. This highlights the need for a new approach.
0:03:34 - Dataset and Parameters: The model is trained and evaluated using a real-world dataset from Milan, Italy, featuring network topologies with 25 to 100 nodes. It considers both link parameters (bandwidth) and node parameters (storage, computation capacity, traffic demands).
0:04:51 - Proposed Architecture: The routing problem is modeled as a sequential decision-making task, making it a perfect fit for reinforcement learning. The specific solution uses a Deep Q-Network (DQN) algorithm with a five-layer neural network to learn the optimal routing policy.
0:07:51 - RL Framework Design: The AI agent's learning is defined by its state, action, and reward function.
State: A comprehensive vector combining normalized features of both network nodes (e.g., storage) and links (e.g., bandwidth).
Action: Deciding the next node to send traffic to from the current node.
Reward: A custom function designed to incentivize the agent for minimizing delay and bandwidth consumption while adding a bonus for successfully completing a path to the destination.
0:08:51 - Training and Positive Results: After training the model for 1,000 episodes on various network topologies, the cumulative reward consistently converged to a stable, positive value (e.g., 65-68). This indicates the agent successfully learned an effective and optimal routing strategy.