LSTM-certainty as early warning signal for critical transitions

Abstract

We trained a long-short-term-memory (LSTM)-neural network on time series generated with an agent-based model that was designed to differentiate the drivers of its dynamics into external and internal forces, with the internal ones stemming from neighbourhood interaction considered as ‘social’ influence. The trained LSTM proved capable of predicting changes in the dynamics of time series from systems prone to critical transitions. The probability of the assessment – i.e. the ‘certainty’ of the LSTM for its prediction – thus can be used to indicate qualitative changes in a system's behaviour. In many cases, these certainties announce imminent state changes earlier and also more clearly than the set of statistical methods, which is suggested for predicting critical transitions under the term early warning signals.