Computational Performance of Deep Reinforcement Learning to Find Nash Equilibria

Abstract We test the performance of deep deterministic policy gradient—a reinforcement learning algorithm, able to handle continuous state and action spaces—to find Nash equilibria in a setting where firms compete offer prices through uniform price auction. These algorithms are typically considered “model-free” although large set parameters is utilized by algorithm. may include rates, memory buffers, space dimensioning, normalizations, or noise decay purpose this work systematically effect these parameter configurations on convergence analytically derived Bertrand equilibrium. choices that can reach rates up 99%. show algorithm also converges more complex settings with multiple players different cost structures. Its reliable make method useful tool studying strategic behavior even settings.