A REINFORCEMENT LEARNING APPROACH TO THE DINING PHILOSOPHERS PROBLEM
Konya Journal of Engineering Sciences · 04.06.2026 · 14 (2) · 613-627
DOI: 10.36306/konjes.1652625 ↗Kurz gefasst
Fachartikel in der Zeitschrift Konya Journal of Engineering Sciences, online erschienen am 04.06.2026. Verfasst von Cem Özkurt und 2 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben.
Abstract (Originalsprache)
The Dining Philosophers problem represents a fundamental challenge in concurrent programming, where resource allocation must balance efficiency with fairness to prevent deadlock and starvation. This study investigates how modern reinforcement learning algorithms compare against classical distributed coordination methods in solving this synchronization problem. We implemented and compared four RL approaches—proximal policy optimization (PPO), deep Q-Network (DQN), advantage actor-critic (A2C), and soft actor-critic (SAC)—along with three classical algorithms—Dijkstra's resource ordering, the centralized waiter protocol, and Chandy-Misra's distributed token-passing—in a multi-agent environment with varying philosopher counts (5, 40, 160). Our experimental results across 20000 training episodes reveal distinct performance characteristics. At a small scale (n=5), PPO achieved the highest cumulative reward (750) with fairness 0.90, matching Chandy-Misra's equity while exceeding its throughput. SAC achieved comparable performance (700 reward, 0.88 fairness) with superior convergence stability. At the medium scale (n=40), PPO maintained strong performance (300 reward, 0.85 fairness), whereas classical algorithms demonstrated consistent deterministic behavior. At large scale (n=160), all methods struggled: PPO and SAC degraded gracefully (-5000, -6000 rewards; 0.65 fairness), classical algorithms maintained predictable performance, whereas DQN failed catastrophically (-10,000+ reward). Multi-dimensional analysis across sample efficiency, scalability, and final performance revealed that classical algorithms provide immediate deployment readiness with zero training cost, whereas RL methods achieve higher performance ceilings after extensive training (1000-2000 episodes). These findings demonstrate that algorithm selection critically depends on deployment constraints: classical methods excel in static environments with deterministic requirements, whereas policy-gradient RL offers adaptive optimization for dynamic systems where coordination rules must evolve, though at a significant computational cost.
Zitieren
APA
Özkurt, C., Kala, A., & Onay, M. E. (2026). A REINFORCEMENT LEARNING APPROACH TO THE DINING PHILOSOPHERS PROBLEM. Konya Journal of Engineering Sciences, 14(2), 613-627. https://doi.org/10.36306/konjes.1652625
Vancouver
Özkurt C, Kala A, Onay ME. A REINFORCEMENT LEARNING APPROACH TO THE DINING PHILOSOPHERS PROBLEM. Konya Journal of Engineering Sciences. 2026;14(2):613-627. doi:10.36306/konjes.1652625
BibTeX
@article{zkurt2026,
author = {Özkurt, Cem and Kala, Ahmet and Onay, Muhammed Emin},
title = {A REINFORCEMENT LEARNING APPROACH TO THE DINING PHILOSOPHERS PROBLEM},
journal = {Konya Journal of Engineering Sciences},
year = {2026},
volume = {14},
number = {2},
pages = {613--627},
doi = {10.36306/konjes.1652625}
}
EUVIMED