Roles instead of strategies: a design that was measured and removed
AHISTER used to give successor branches alternating search strategies, each with its own objective and gene pool. Three A/B comparisons showed no advantage over one objective with different branch roles, so the strategies were removed. A negative result, recorded.
The first design of AHISTER split the search into two strategies. An exploration strategy optimized quality. An exploitation strategy ran a separate non-dominated-sorting objective under an error bound and tried to compress what exploration had found. Successor branches alternated strictly between the two, and each strategy kept its own gene pool so that diversity was measured per strategy.
It was a reasonable design, and the earlier text on this site described it with some confidence. It also carried a lot of machinery: two objectives, two pools, an error bound handed from predecessor to successor, and a rule for who may claim which structural family.
The measurement
The question was simple: does the second strategy pay for itself? Three A/B comparisons against a version with a single objective put the split between 0.66 and 1.0 times parity per evaluation. At best it was as good as not having it.
What replaced it
Every branch now optimizes the same quality. What differs is the role a branch plays:
- Fresh exploration. A new population, grown in isolation.
- Pair successor. Two stagnated branches of similar maturity are paired; their elites form a tournament branch and the rest an advanced branch.
- Restart. A branch seeded from the best material found anywhere, as recorded in the Hall of Fame.
The asymmetric split that gives AHISTER its name survives unchanged. What went away is the idea that the two halves need different objectives to stay different. Different starting material turned out to be enough.
Why write it down
Because a design that is described and then quietly changed is worse than one that was never described. And because the result generalizes a little: in a system where every candidate is already refined to a local optimum before it competes, much of what a second objective is meant to provide is already provided by the refinement.