Enhancing Multi-Agent Reinforcement Learning Intrusion Detection Systems Using Random Forest Q-Value Estimation
Ricky Aurelius Nurtanto Diaz, I Ketut Gede Darma Putra, Made Sudarma, I Made Sukarsa, I Wayan Budi Sentana, Ni Luh Gede Pivin Suwirmayanti · Engineering Technology & Applied Science Research · 2025
Intrusion Detection Systems (IDSs) analyze network traffic and system activity to identify anomalies or suspicious attack patterns. Various artificial intelligence-based approaches have been explored, including Deep Learning (DL) and Multi-Agent Reinforcement Learning (MARL) to increase their accuracy. This study combines MARL with Random Forest (RF) for Q-value estimation and utilizes two agents, a Detector and a Classifier. The proposed method was evaluated on three public datasets, including UNSW-NB15, NSL-KDD, and UKM-IDS20. The experimental results showed that the Detector Agent achieved higher accuracy (99.95%) compared to the Classifier Agent (80.63%) for the UNSW-NB15 dataset. On the NSL-KDD dataset, both agents performed similarly, with the Detector Agent achieving 99.82% accuracy and the Classifier Agent 99.80%. In contrast, for the UKM-IDS20 dataset, the Classifier Agent slightly outperformed the Detector Agent, with accuracies of 99.98% and 99.94%, respectively. These findings demonstrate the effectiveness of MARL-based IDS and highlight variations in agent performance across different datasets.