Enterprise AI Analysis: The Illusion of Collusion

The paper categorizes bandit algorithms into three types based on their randomness:

• Persistently Random Algorithms: Always maintain a non-zero probability of selecting any action (e.g., epsilon-greedy without decay). These algorithms lead to competition in the limit.
• Greedy-in-the-Limit Algorithms: Start with random sampling but converge to a degenerate distribution, playing only the observed optimal action in the long run (e.g., explore-then-commit, epsilon-greedy with decaying epsilon). These sometimes lead to collusion.
• Deterministic Algorithms: Follow a strictly deterministic approach for action selection, prescribing exactly one action with probability 1 (e.g., Upper Confidence Bound - UCB). These always lead to collusion.

The research has several implications for understanding and regulating algorithmic collusion:

• Naive Collusion: Firms using independent "textbook" experimentation algorithms can converge to supracompetitive prices without explicit intent or coordination.
• Beyond Competitor Prices: Limiting algorithms from conditioning on competitor prices may not prevent collusion, as naive collusion can emerge even without such information.
• Symmetry & Collusion: Symmetric deterministic algorithms always collude, suggesting that similar algorithms in a market can lead to supracompetitive prices.
• Algorithm Choice Matters: The specific choice and parameterization of learning algorithms dictate whether collusion arises, emphasizing the need for nuanced regulatory scrutiny.
• Trial-and-Error Effects: More extensive exploration can sometimes *reduce* collusion likelihood, a counter-intuitive finding against common assumptions.