Fig. 2: 3 state Markovian system: LSTM, ps-LSTM and analytical predictions.
From: Path sampling of recurrent neural networks by incorporating known physics
Here we show results of applying ps-LSTM to the 3 state Markovian system where we constrain 〈N〉. In (a), we provide the input transition kernel without constraints. In (b), we show the transition kernel obtained from ps-LSTM generated time-series via direct counting, where we achieve a 〈N〉 close to the target 〈N〉=0.13. The calculated values for 〈N〉 are shown in (c) for LSTM as the average of 100 predictions and for ps-LSTM as the average of 200 predictions. The error represents “error percentage” which is defined as the difference between ps-LSTM result and target value 〈N〉=0.13 divided by the target value.
