ROSE: Reinforcement Learning Optimised Spectrum Exploration for Intelligent Electronic Warfare Scanning

Abstract

This paper presents ROSE (Reinforcement Learning Optimised Spectrum Exploration), an intelligent spectrum scanning framework designed for wideband Electronic Warfare (EW) environments. The proposed system addresses the challenge of monitoring a wide radio-frequency spectrum when a receiver has limited instantaneous bandwidth. ROSE formulates spectrum scanning as a Partially Observable Markov Decision Process (POMDP) and employs a Deep Q-Network (DQN) to dynamically select frequency bands for scanning. The framework consists of nine interconnected modules covering RF environment simulation, receiver modeling, signal characterization, temporal pattern discovery, activity mapping, transition analysis, DQN-based scheduling, evaluation, and real-time visualization. The DQN uses a 65-feature observation vector and selects among 20 frequency-band scanning actions. Its performance is evaluated against Sequential, Random, and Priority-based scanning strategies across Mixed, Periodic, Intermittent, and Frequency-Agile RF scenarios. Experimental results demonstrate that the ROSE DQN achieves a probability of detection of 0.989 in the Mixed scenario, substantially outperforming conventional Sequential and Random strategies. However, the evaluation also identifies limitations in highly intermittent and frequency-agile environments. A detailed system audit reveals issues including an agile-emitter reset bug, the lack of transition-graph information in the DQN state representation, and a mismatch between the specified and simulated instantaneous bandwidth. These findings provide a basis for future improvements using transition-aware state representations and Dueling Double DQN architectures. The study demonstrates the potential of reinforcement learning for adaptive spectrum exploration and provides a simulation-based foundation for developing more responsive and efficient EW spectrum scanning systems.