Deep Reinforcement Learning for Congested Multifunction Radar Task Scheduling
15th Mediterranean Conference on Embedded Computing, MECO 2026, Budva, Karadağ, 9 - 15 Haziran 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/meco70748.2026.11579076
- Basıldığı Şehir: Budva
- Basıldığı Ülke: Karadağ
- Anahtar Kelimeler: Deep Reinforcement Learning, Markov Decision Process, Multi-function Radar Systems, Radar Resource Management, Task Scheduling
- Hacettepe Üniversitesi Adresli: Evet
Özet
Task scheduling is a critical time-constrained decision problem affecting the performance of modern multifunction radar (MFR) systems, particularly under congested task conditions. While heuristic approaches such as Earliest Start Time (EST) are computationally efficient, their performance degrades significantly as task density increases. In this work, the radar task scheduling problem is formulated as a finite-horizon Markov Decision Process (MDP), and a systematic comparison of heuristic, value-based, and policy-gradient Deep Reinforcement Learning (DRL) methods is presented. We specifically evaluate the EST algorithm alongside Deep Q-Network (DQN), Advantage Actor-Critic (A2C), and Proximal Policy Optimization (PPO). The analysis covers scheduling cost, task drop rate, learning dynamics, and convergence behavior. Utilizing episodic rewards for DQN and dense stepwise rewards for policy-gradient methods under empirical hyperparameter tuning, the results demonstrate that A2C and PPO achieve substantially lower costs and drop rates than EST in congested scenarios, while maintaining robust performance in lightly loaded cases.