Prescribed attractivity region selection for recurrent neural networks based on deep reinforcement learning

Recurrent neural networks' (RNNs') outputs are the same when network states converge to the same saturation region. Strong external inputs can cause the neural network to converge to a prescribed saturation region. Different from previous works, this paper employs deep reinforcement learning to obtain external inputs to make network states converge to the desired saturation region. Firstly, for five-dimensional neural networks, the deep Q learning (DQN) algorithm is used to compute the optimal external inputs that make the network state converge to the specified saturation region. When scaling to n-dimensional RNNs, the problem of dimensional disaster is encountered. Then, it proposes a batch computation of the external inputs to cope with the curse of dimensionality. At last, the proposed method is validated by numerical examples, and compared with existing methods, it shows that less conservative external inputs conditions can be obtained.