Economics Seminar Series (ESS) - 2023/2024

Interviene: Daniele Condorelli (Warwick University)

Title: Deep Learning to Play Games

Abstract: We train a deep neural network to play arbitrary normal-form games. At each iteration, the network takes a random bimatrix as input and outputs a mixed-strategy over the row-player actions. Then, it evaluates what it would have played as the column-player and it adjusts its parameters by a small amount in the direction that moves its predicted choices toward those that minimise its instantaneous regret. Computationally, we show the network converges to play a Nash equilibrium in every game. In particular, it always plays a stable equilibrium and in nearly 70% of cases the one arising from the Harsanyi-Selten tracing procedure.