On algorithmic collusion and reward-punishment schemes

A booming literature describes how artificial intelligence algorithms may autonomously learn to generate supra -competitive profits. The widespread interpretation of this phenomenon as "collusion"is based largely on the observation that one agent's unilateral price cuts are followed by several periods of low prices and profits for both agents, which is construed as the signature of a reward-punishment scheme. We observe that price hikes are also followed by aggressive price wars. Algorithms may also converge to outcomes that are worse than Nash and penalize deviations from it. While admissible in equilibrium, this behavior throws interesting light on the relationship between high algorithmic prices and the standard mechanisms behind collusion.