Algoritmos para multi-armed bandits: teoria e aplicação à precificação dinâmica

Abstract

This work addresses the problem of sequential decision-making, focusing specifically on the multiarmed bandit (MAB) framework. In its classical formulation, the MAB problem involves an agent facing a row of slot machines (bandits), with a limited number of pulls (arms) available. The agent’s goal is to determine a sequence of actions that maximizes the total reward. The core challenge lies in balancing the trade-off between choosing the action that currently appears to yield the highest reward and exploring lesser-known alternatives (a dilemma known as exploration versus exploitation). In this study, we explore several algorithms designed to support decision-making within the multiarmed bandit setting. We also examine an application of this theory to the problem of dynamic pricing, i.e., determining optimal selling prices for products and services. In this context, the seller takes the role of the agent who aims to sell a product by selecting from a finite set of possible prices, without prior knowledge of demand or consumer behavior. The seller must therefore adopt a strategy that enables the identification of the optimal price over time.