Glossary
Definitions of terms used in AdBandit
Multi-Armed Bandit
A sequential decision problem that chooses among multiple options (arms) to maximize reward. It requires balancing exploration and exploitation.
UCB1 (Upper Confidence Bound 1)
An algorithm that scores each arm using average reward and uncertainty, balancing exploration and exploitation automatically.
Reward Function
A weighted score of impressions (α), clicks (β), and conversions (γ), where α + β + γ = 1.0.
Arm
A selectable creative option. In AdBandit, each arm is defined by combinations like angle, tone, and format.
Pruning
Removing low-performing arms after enough trials to focus resources on promising candidates.
Event
One cycle of generate → publish → observe. Metrics are fetched and arm statistics are updated.
Content Factory
Module that generates ad copy/images from arm parameters using prompt templates and generative AI.
Delivery Agent
Module that posts generated content to social platforms while considering API constraints and rate limits.
Observation
Collecting post metrics (impressions, clicks, conversions), computing reward, and updating arm stats.
Mutant Arm
A slightly modified new arm generated from a pruned arm to preserve exploration diversity.