Algorithms, Artificial Intelligence, and Simple Rule-Based Pricing

Abstract

Automated pricing strategies in e-commerce generally fall into two categories: simple rule-based protocols, such as undercutting the lowest price, and sophisticated artificial intelligence (AI) powered algorithms, particularly those based on reinforcement learning (RL). While the industry is shifting toward AI, we challenge the assumption that sophisticated algorithms invariably outperform simple rules. Through extensive simulations and theoretical analysis, we demonstrate that a seller often achieves higher profits by adopting a simple rule-based algorithm against an RL competitor than by adopting RL themselves. We attribute this counter-intuitive result to "simultaneous exploration'': in symmetric RL competition, concurrent learning creates a non-stationary environment that traps agents in a sub-optimal "intermediate equilibrium.'' In contrast, a fixed rule-based strategy stabilizes the environment, enabling the RL opponent to learn a fully collusive "always cooperate'' strategy. However, this outcome depends on market structure; when product differentiation is low, the interaction instead triggers aggressive "Edgeworth price cycles.'' We further show that simple rule-based algorithms can lift market prices in oligopolies with up to five sellers, where pure RL collusion typically fails. Finally, we derive the conditions under which asymmetric technology adoption (one RL, one Rule-based) emerges as the equilibrium in an algorithm-choosing game. Our findings provide practical guidance for retailers on technology selection and offer new insights for policymakers regarding the risks of asymmetric algorithmic collusion.

Keywords: Algorithmic pricing, competition, rule-based pricing, reinforcement learning