Neural networks playing 'matching pennies' with each other: Reproducibility of game dynamics

Abstract

Reflection is an essential feature of consciousness and possibly the single most important one. This fact allows us to simplify the objective of the concept of 'neural correlates of consciousness' and to focus investigations on reflection itself. Reflexive games are the concentrated and pure embodiment of reflection manifestation without the addition of other higher cognitive functions. In this paper, we use the game 'matching pennies' ("Odd-Even") in order to trace the strategies and possible patterns of recurrent neural network operation. Experimental results show the splitting of all considered game patterns into two groups. A significant difference was observed in these groups of patterns, indicating a qualitative difference in game dynamics apparently due to the qualitatively different dynamic patterns of neuron excitations of the networks. A similar splitting of all players into two groups was found by other authors for human players, which differ in terms of the reflection availability. By this, we can assume that one of the causes of the splitting is that the presence of reflection in a particular group of recurrent neural networks dramatically changes the game meta-strategy.