Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Genya Ishigaki

Second Advisor

Amith Kamath Belman

Third Advisor

Vuthea Chheang

Keywords

Deep Reinforcement Learning, Semantic Communication, Adaptive Transmission, Edge Offloading, Variational Autoencoder

Abstract

The emerging paradigm of semantic communication shifts the transmission objective from bit-perfect reconstruction to the preservation of task-relevant meaning. However, most existing frameworks rely on a single transmission strategy and there- fore do not remain optimal when channel bandwidth, noise, and on-device compute resources vary at runtime. This thesis presents an adaptive semantic communication framework in which a Deep Q-Network (DQN) agent selects among three transmission strategies: raw pixel transmission, edge-assisted semantic encoding via a full Stable- Diffusion variational autoencoder, and on-device semantic encoding via a distilled TinyVAE. The agent observes a five-dimensional state vector comprising CPU uti- lization, memory pressure, payload size, channel noise variance, and instantaneous bandwidth, and is trained using a multi-objective reward that penalizes both latency and perceptual reconstruction loss measured by Learned Perceptual Image Patch Similarity (LPIPS). The system is implemented as a distributed simulation of five containerized microservices with native bandwidth shaping through the Linux Traffic Control subsystem. Across three experimental scenarios, the agent learns the appro- priate transmission strategy from reward feedback alone, without manually tuned thresholds, and converges to a policy that outperforms fixed-strategy baselines. These findings show that learning-based transmission selection is a practical approach for adaptive edge orchestration under varying channel conditions.

Download

Available for download on Wednesday, May 26, 2027

Share

COinS