In repeated zero-sum games,instead of constantly playing an equilibrium strategy of the stage game,learning to exploit the opponent given historical interactions could typically obtain a higher utility.However,when pl...In repeated zero-sum games,instead of constantly playing an equilibrium strategy of the stage game,learning to exploit the opponent given historical interactions could typically obtain a higher utility.However,when playing against a fully adaptive opponent,one would have dificulty identifying the opponent's adaptive dynamics and further exploiting its potential weakness.In this paper,we study the problem of optimizing against the adaptive opponent who uses no-regret learning.No-regret learning is a classic and widely-used branch of adaptive learning algorithms.We propose a general framework for online modeling no-regret opponents and exploiting their weakness.With this framework,one could approximate the opponent's no-regret learning dynamics and then develop a response plan to obtain a significant profit based on the inferences of the opponent's strategies.We employ two system identification architectures,including the recurrent neural network(RNN)and the nonlinear autoregressive exogenous model,and adopt an efficient greedy response plan within the framework.Theoretically,we prove the approximation capability of our RNN architecture at approximating specific no-regret dynamics.Empirically,we demonstrate that during interactions at a low level of non-stationarity,our architectures could approximate the dynamics with a low error,and the derived policies could exploit the no-regret opponent to obtain a decent utility.展开更多
基金the Science and Technology Innovation 2030-"New Generation Artificial Intelligence"Major Project(No.2018AAA0100901)。
文摘In repeated zero-sum games,instead of constantly playing an equilibrium strategy of the stage game,learning to exploit the opponent given historical interactions could typically obtain a higher utility.However,when playing against a fully adaptive opponent,one would have dificulty identifying the opponent's adaptive dynamics and further exploiting its potential weakness.In this paper,we study the problem of optimizing against the adaptive opponent who uses no-regret learning.No-regret learning is a classic and widely-used branch of adaptive learning algorithms.We propose a general framework for online modeling no-regret opponents and exploiting their weakness.With this framework,one could approximate the opponent's no-regret learning dynamics and then develop a response plan to obtain a significant profit based on the inferences of the opponent's strategies.We employ two system identification architectures,including the recurrent neural network(RNN)and the nonlinear autoregressive exogenous model,and adopt an efficient greedy response plan within the framework.Theoretically,we prove the approximation capability of our RNN architecture at approximating specific no-regret dynamics.Empirically,we demonstrate that during interactions at a low level of non-stationarity,our architectures could approximate the dynamics with a low error,and the derived policies could exploit the no-regret opponent to obtain a decent utility.
