When a bandit launches a new option with little or no data, an optimistic prior gives that option the benefit of the doubt. Rather than letting established choices dominate immediately, the system treats the newcomer as promising enough to receive a fair chance. section 5.6
Use optimistic priors to address cold-start decisions such as:
- Launching a new advertisement or creative before its performance is known. section 5.6
- Serving a newly encountered customer type when similar historical observations are unavailable. section 5.6
- Preventing incumbent options from burying alternatives before they have accumulated evidence. section 5.6
A fair launch for a new creative
Suppose a contextual bandit already knows how several advertisements perform but has no observations for a newly launched creative. An optimistic prior initially assumes the newcomer may perform well enough to merit exposure. Its subsequent rewards can then replace that favorable starting belief with evidence. A warm start offers an alternative by borrowing knowledge from similar ads or users. section 5.6 section 5.7
