• SARSA
• massive MIMO radar
• adaptive parameters
• target detection
• beamforming
• constant false alarm rate
• Reinforcement Learning

In the present work, a reinforcement learning (RL) based algorithm to optimize the transmit beampattern for a co-located massive MIMO radar is presented. A massive MIMO radar is a radar with N=NTNR>>1, where NT and NR are the number of transmitting and receiving antennas respectively. Under the massive MIMO regime it is possible to develop a detector that guarantees certain detection performances under any practical disturbance model. The RL paradigm allows the system to improve its performances by learning the optimal behaviour via interacting with the surrounding unknown environment. An original contribution of this master thesis is the derivation of a fully adaptive scheme to select the optimal values of the hyperparameters involved in the RL algorithm from the available data. This last aspect is of fundamental importance to make the proposed detection algorithm independent of any ad-hoc and potentially sub-optimal manual tuning of the hyperparameters. The resulting algorithm is tested against the conventional orthogonal beamformer and an adaptive beamformer without RL showing better detection performances in simulations with harsh scenarios where strong clutter and/or targets with low SNR are present. Finally, some critical aspects of the algorithm are analysed and possible solutions are proposed for future work.