LSTM model predicting outcome of strategic thinking task exhibits representations of level-k thinking

Abstract

Which neural mechanisms underlie strategic thinking in the human brain? Neuroeconomic research has not yet bridged the gap between theoretical models of higher-order reasoning and the precise mechanisms implemented in neural networks in the human brain. In this paper, I demonstrate that a recurrent neural network model can learn to perform strongly in the simple strategic game Rock-Paper-Scissors. In doing so, it develops implicit representations of strategically important variables (the levels $k$ of reasoning) which economists have postulated in theoretical models. These representations can be extracted from the hidden activations of the neural network. These findings hint at a connection between the mechanisms implicit in recurrent neural networks and models of strategic thinking in economic theory. Future empirical brain research can investigate whether these mechanisms correspond to mechanisms implicit in biological neural networks.