【摘要】 本文在具有需求差异的两个市场中，以实验方法研究了人工智能Q学习算法的寡头重复价格竞争行为。基准实验结果发现，智能体通过训练能够学会惩罚策略，进而在两个市场实现歧视性定价合谋，且价格歧视水平和平均利润水平会随着智能体“理性”程度的提升而先提高后下降。随着参与竞争的智能体数量的增加，两市场的价格均会下降，价格歧视水平会趋向于0。进一步实验结果表明，当消费者存在不公平定价厌恶时，智能体通过学习能够“理解”这种消费心理，进而降低高需求市场价格、提高低需求市场价格，达到新的价格合谋。随着不公平定价厌恶对需求的影响力提高，两市场的价格差异不断下降并向0收敛。当消费者存在相对低价偏好时，智能体会利用这种消费心理，通过提高低需求市场的定价使高需求市场的相对价格变得更低，从而扩大高需求市场的需求量而提升总利润水平。随着相对低价偏好对需求的影响力提高，两市场的价格歧视程度会随之下降，并维持在较低水平，在此过程中低需求市场的收敛价格会超过高需求市场收敛价格。本文结论在对现有算法价格竞争文献进行拓展的同时，也为相关支持和规制政策的制定提供依据，同时显示当前学界对算法歧视性定价的担忧存在过度倾向。更多还原

【Abstract】 This paper studies experimentally the behavior of artificial intelligent agents powered by the Q-learning algorithm in a oligopoly model of repeated price competition in an environment consisting of two markets with different demands. We find that the Q-learning agents can learn punitive strategies through training, thereby achieving collusive price discrimination across the two markets. The price discrimination level and average profits first increase and then decrease as the agents’ rationality rises. As the number of competing agents increases,prices decline in both markets, and the price discrimination level tends toward zero. Additional experimental findings indicate that when consumers exhibit aversion to unfair pricing, agents can adapt to these consumer preferences, leading them to reduce prices in high-demand markets and raise prices in low-demand markets, thereby achieving a new price collusion. As the impact of unfair pricing aversion on demand intensifies, the price discrimination level continues to narrow and converges toward zero. When consumers exhibit a preference for relatively low prices, agents leverage it to raise price in low-demand markets, which in turn makes relative prices in high-demand markets lower. This pricing strategy expands demand in high-demand markets and increases the overall profit. As the impact of the preference for relatively low prices on demand intensifies, the price discrimination level decreases and remains at a low level, and the convergent price in the low-demand market exceeds that in the high-demand market. The conclusions of this study suggest that the concerns about algorithm discriminatory pricing expressed in existing literature are excessive. This paper not only extends the existing literature but also provides an important basis for the formulation of supportive and regulatory policies.更多还原