Reinforcement learning-based complex-valued space-time MIMO 2D-LSTM nonlinear equalizer for photonics-assisted THz indoor optical wireless access networks

Photonics-assisted terahertz (THz) communication provides a solution for indoor optical wireless access networks (OWANs) that require high-speed and low-latency wireless connections. The application of polarization division multiplexing (PDM) technology not only increases the transmission capacity of the system but also improves its robustness to physical blockages in indoor scenes. However, it also faces PDM crosstalk and nonlinear damage caused by optoelectronic devices. This paper proposes a complex-valued space-time multiple-input multiple-output two-dimensional long short-term memory neural network (ST-MIMO 2D-LSTM) equalizer that can simultaneously process signals with different polarization directions of the PDM. On the basis of our previous work, we propose a new, to the best of our knowledge, state update strategy for the space-time automatic search algorithm (ST-ASA) based on the deep deterministic policy gradient algorithm. By simultaneously updating the space-time parameters, the optimal internal structural parameters of the neural network equalizer are automatically searched. The experiment demonstrated the 2 m wireless transmission of a 40 GBaud PDM-QPSK signal in a 290 GHz photonics-assisted THz system. By using ST-ASA to automatically search for the optimal internal structural parameters of various equalizers, we verify the effectiveness of the new state update strategy and its superiority in practical applications. Subsequently, we compare and verify that our proposed ST-MIMO 2D-LSTM equalizer can achieve better bit error rate (BER) performance with lower computational complexity. When the optical power into the uni-traveling-carrier photodiode is 0.8 and 2.8 dBm, the BER of the received signals in the horizontal and vertical polarization directions can reach ${3.73} \times {{10}^{- 3}}$ and ${3.68} \times {{10}^{- 3}}$, respectively, which are lower than the 7% hard decision forward error correction threshold of ${3.8} \times {{10}^{- 3}}$, and the computational complexity is on average 26% lower than 1D-DNN, ST-MIMO 2D-DNN, and 1D-LSTM. The experimental results show that the ST-MIMO 2D-LSTM equalizer combined with ST-ASA using the new state update strategy can meet the communication requirements of a large capacity in indoor OWANs and has great potential in supporting user mobility and promoting the implementation of indoor machine to machine.