A constrained bandit extends ordinary bandit decision-making with operational limits. It chooses actions while respecting budgets, exposure caps, fairness goals, or similar constraints, rather than optimizing a single metric without safeguards. section 5.6
Use constrained bandits when business objectives extend beyond immediate clicks:
- Shape rewards to reflect profit, customer lifetime value, margin protection, brand safety, or fairness instead of optimizing clicks alone. section 5.6
- Apply explicit constraints when actions must stay within budgets, exposure limits, or fairness requirements. section 5.6
- Recognize that delayed outcomes may require proxy rewards or buffered updates before the system can learn fairly. section 5.6
Marketing decisions
A marketing system might otherwise maximize clicks by giving away too many discounts or promoting unsuitable products. A constrained bandit instead balances response with profit, margin, brand safety, fairness, and limits on exposure or spending. section 5.6
