基于神经网络的微分对策控制器设计

Differential game controllers design using neural networks

采用伴随 - BP技术 ,将微分对策的两点边值求解问题转化为两个神经网络的学习问题 ,训练后的两个神经网络分别作为对策双方的最优控制器在线使用 ,避免了直接求解复杂的两点边值问题。对追逃微分对策问题的仿真结果表明 ,该方法对初始条件和噪声具有较好的鲁棒性。

Based on the adjoint-backpropagation technique,two point boundary value problem(TPBVP) of differential games is solved using two neural networks.The neural networks automatically adjust their weights to minimize and maximize the cost function of differential game system respectively.The converged neural networks can be used as the feedback optimal differential games controllers on-line,avoiding solving the complex two point boundary value problem directly. The simulation results for pursing-escaping differential games prove that the neural networks controllers present good robustness with respect to initial conditions and measuring noise.