We propose a joint source-channel coding (JSCC) scheme based on model-predictive control (MPC) for encoding an analog discrete-time source over a memoryless communication channel. The decoder is assumed to be fixed and given, and the MPC encoder is set to minimize the frequency-weighted mean-squared error (FWMSE) while satisfying a channel input constraint. We derive analytic expressions for the associated finite-horizon optimization MPC problem for an asymptotic (in time) average or sum channel input constraint. For the case of complex additive white Gaussian noise channel and an $M$-quadrature-amplitude modulation ($M$-QAM) decoder, these expressions can be written explicitly, which allows for an efficient numerical solution. For the case $M$=4 and considering a frequency weighting criterion that models the sensitivity of the human hearing, we show by simulations that our proposed scheme yields an FWMSE significantly smaller than that yielded by the corresponding non-JSCC scheme (a 2-bit uniform quantizer in tandem with a 4-QAM channel encoder and decoder).

我们提出了一种基于模型预测控制(MPC)的联合源信道编码(JSCC) 方案，用于通过无记忆通信信道对模拟离散时间源进行编码。假设解码器是固定的和给定的，并且 MPC 编码器被设置为在满足通道输入约束的同时最小化频率加权均方误差(FWMSE)。我们为渐近（在时间上）平均或总和通道输入约束的相关有限范围优化 MPC 问题推导出解析表达式。对于复杂的加性高斯白噪声通道和一个$米$-正交幅度调制（$米$-QAM) 解码器，这些表达式可以明确地写出来，这允许有效的数值解决方案。对于这种情况$米$=4 并考虑模拟人类听觉灵敏度的频率加权标准，我们通过模拟表明，我们提出的方案产生的 FWMSE 明显小于相应的非 JSCC 方案（一个 2 位均匀量化器与4-QAM 信道编码器和解码器）。