Explore, Then Exploit
Do I know where the gains are likely to be, or am I already optimizing one path?
Situation
A problem has many possible places to intervene and no obvious starting point.
The idea
Early experiments should sometimes maximize diversity of learning rather than incremental performance. Once a promising region appears, deeper optimization becomes more useful.
Useful when
New productsambiguous technical problemsoptimizationexperimentation
Watch for
Exploration can become endless experimentation. Shift modes once enough structure appears.
Where this came from
Applied machine learning, Research