ML-Based Parameter Adaptation for OFDM-ISAC: A USRP Experimental Study

The Integrated Sensing and Communication (ISAC) system developed using GNU Radio and the Universal Software Radio Peripheral (USRP) B200 is practically realized in this paper. The proposed framework enables simultaneous wireless data transmission and environmental sensing within the same radio frequency range. To improve overall system performance, a number of machine learning (ML) models, such as Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), Support Vector Regression (SVR), and Random Forest (RF), were trained to optimize critical transceiver parameters, such as FFT size, payload modulation format, and pilot symbol allocation. For adaptive configuration during real-time operation, these optimized parameters were dynamically applied to GNU Radio flowgraphs. When compared to non-optimized values, the RFbased optimization had the greatest overall performance among the assessed methods, with significant gains in both sensing and communication. At a distance of 4.12 meters, the optimized arrangement improved communication power by up to 2.47 dBm and sensing power by 3.2 dBm. Additionally, the RF model's better reliability was proven by Bit Error Rate (BER) analysis across several machine learning techniques, which produced the lowest BER values under all test scenarios. These findings demonstrate the RF model as a reliable, effective optimization method for realtime adaptive ISAC systems and confirm the efficacy of machine-learning-driven ISAC optimization employing inexpensive SDR platforms.