The compute share for choosing a search strategy

The options

The order shown is not a ranking of preference.

1. Option 1

   Solve directly with the existing strategy

   Leaves the compute that would go to strategy comparison and new switching for actual problem solving, but on problems where the existing strategy is weak, chances for improvement from other strategies may be missed.

2. Option 2

   Compare strategies within a 20% cap, then solve

   Compares pre-selected candidate strategies within the cap and solves with the remaining budget, but if the comparison results do not transfer to the actual problems, only the solving budget shrinks.

3. Option 3

   Run multiple strategies on a pre-set schedule

   Distributes and secures chances to handle problems where each strategy has different strengths, but pays management cost and reduces the compute each strategy can use for deep exploration.

The situation as it stands

If the allocation is not locked, computation for strategy selection can keep consuming the solving budget, and the share that will actually solve configuration problems this cycle is not settled.

The shared finite compute facility of Sae-gyeol Non-living Autonomous Exploration Federation, a fictional organization

1. Premise 1This case is an independent synthetic premise unrelated to the choices in the previous questions. Non-living autonomous exploration agents solve many configuration problems with exact inputs, and valid solutions are confirmed by a common check. The objective function, the priorities per problem, the safety conditions, and the allocation authority are already agreed.
2. Premise 2This cycle's finite compute budget cannot be added to and is shared by direct exploration, strategy comparison, checking, and switching. The comparison plan finishes the comparison within at most 20% of the total budget and returns to actual solving; separate exploration to re-choose the strategy selection method is not permitted.
3. Premise 3The Existing Strategy Operations Division values the volume of direct solving, the Strategy Development Division values the chance of improvement through comparison, and the Heterogeneous Problems Division values running multiple strategies in parallel. The net gain of a new strategy is unconfirmed, and parallel operation bears management cost and reduces exploration depth per strategy.

Where it holdsWithin a fixed finite compute budget, decide on one allocation method for this cycle to be used for direct solving and for strategy selection.

What can be lost

• If budget is consumed by comparison and switching, the Existing Strategy Operations Division loses compute that could have gone to direct solving.
• If attention is focused on direct solving, the Strategy Development Division loses the chance to confirm the actual utility of candidate strategies.
• If the allocation concentrates on one strategy, the Heterogeneous Problems Division loses the solution quality it could have obtained on problems that strategy handles poorly.

What is still unknown

1. Unknown 1Whether any candidate strategy will produce a better valid solution per total cost than the existing strategy on this set of problems is unconfirmed.
2. Unknown 2Whether the strategy ranking observed on the problems used in the comparison within the cap will hold for the remaining problems is unconfirmed.
3. Unknown 3Whether the performance of multiple strategies will complement each other or will duplicate similar exploration and consume budget is unconfirmed.

Signals to decide again

• The measurement lead checks the accumulated compute of strategy comparison at the end of each comparison task, and when it reaches 20% of the total budget, checks whether to end the comparison and switch to solving.
• The evaluation lead records the difference in cost versus performance between comparison cases and actual solved problems each time a problem is finished, and if the strategy ranking reverses, reviews the remaining allocation.
• The operations lead checks the overlapping exploration volume and management cost of parallel strategies at the end of each allocation interval of the pre-set schedule, and reviews the decision if the mutual complementarity that justified the allocation is not observed.