The escalating demand for faster, reliable, and energy-efficient wireless communications has steered researchers towards millimetre-wave (mm Wave) frequencies, offering immense bandwidth and high data rates. To adapt to the increasing complexity of such networks, machine learning (ML)-assisted techniques are used for efficient adaptation without complete parameter dependence knowledge. ML-assisted adaptive techniques are applied to an OFDM-CSIM system over amm Wave channel, utilising index modulation and compressed sensing for improved spectral efficiency, energy efficiency, and system design freedom. A DNN-based classifier is proposed, enhancing throughput and outperforming traditional adaptive modulations. A novel multi-layer Sparse Bayesian learning algorithm estimates channel state information with lower complexity, providing more accurate estimation and better performance than conventional methods. Then, the ML-assisted techniques are extended to joint radar and communication systems, using radar-derived side information to adjust communication beams, reducing training overhead and complexity for channel estimation. The system employs a uniform rectangular planar array with adaptive adjustment of antenna elements and array configurations via deep neural network and convolutional neural network classifiers. The simulation results show that the proposed method can achieve a satisfactory data rate that approaches the upper bound obtained by the exhaustive search scheme as well as guaranteeing the required sensing performance. In contrast to previous joint radar and communication system designs that separate these functions through different sub-antenna arrays, a more efficient approach integrating both sensing and communication tasks within a single system, called dual functional radar-communication, is introduced. An ML-assisted beamforming design for ultra-dense device-to-device mm Wave networks uses a convolutional long short-term memory-integrated graph neural network (CL-GNN) to learn historical channel characteristics and predict the beamforming matrix. Our findings show that this design meets the required sensing performance and achieves a near-optimal sum rate. The adaptable CL-GNN can be generalised for networks of varying sizes and densities.