Exploitation chooses the action currently believed to produce the greatest reward. It turns existing estimates into immediate decisions: show the ad with the highest predicted click-through rate, offer the discount with the strongest conversion history, or promote the bundle with the highest upsell rate. section 5.2
Exploitation is inseparable from exploration. Acting greedily can lock a policy into a suboptimal choice, while exploring too much wastes opportunities to use reliable knowledge. Bandit and reinforcement-learning methods therefore balance current reward against information that may improve future decisions. section 5.2 section 5.7
The balance appears at different decision scales:
- Multi-armed bandits exploit the arm with the best estimated overall value. section 5.7
- Contextual bandits exploit actions suited to features such as demographics or time of day. section 5.7
- Monte Carlo Tree Search exploits promising branches while sampling alternatives, using UCB1 during selection. section 7.5
Choosing a campaign action
Suppose several discounts have been tested. Exploitation selects the discount currently estimated to maximize revenue, while continued limited exploration checks whether another discount performs better or protects margin under changing conditions. section 5.2 section 5.7







