AMM
A budget-aware optimizer routing method that uses online probing to analyze the landscape characteristics of a black-box optimization problem in real time, and dynamically selects the best-matching optimization algorithm, such as quasi-Newton methods, trust region, coordinate descent, multi-start, and evolutionary strategies.
Timeline (last 90 days)
Under a known budget, AMM (Adaptive Multi Mode) uses online probing to judge in real time the curvature, plateaus, separability, kinks, basins, and fractal structure of the current landscape, then dynamically allocates budget to the best-matching optimization engine.
Each of AMM's probes comes with default values, a dimension-budget validity range, and a probing cost counted against the total budget, so probing itself is not free, and routing decisions must trade off information gain against budget consumption.
AMM routes to five classes of engines: quasi-Newton, trust region, coordinate, multi-start, and evolution strategies.
AMM conducted a pre-registered development loop on the bbob benchmark, and 32 of 36 predictions held in the subsequent 16 confirmations.
AMM was tested once on held-out bbob instances 101-110 (D=2-40, budgets 10*D to 1000*D) against up to seven competitors; across 600 function-dimension-budget cells, AMM had the highest objective share in 382 cells (236 outright, 146 tied).
Under the descriptive posterior rule, AMM credibly trailed on only 35 cells in the held-out bbob test, concentrated mainly in the low-budget regimes of 10*D and 30*D.
At each budget tier, AMM also led the control groups built from its own engines—including the single best engine, the best engine selected by dimension and budget, and greedy scheduling—with the margin narrower at 10*D and 30*D; it also maintained a lead over nevergrad's NgIohTuned.
The pre-registered runs showed that AMM's lead over IPOP-CMA-ES holds on affine-transformed and box-constrained bbob, and also holds on the combined results for rotated f2-f4.
AMM has one clear failure scenario: after rotating the separable f3 and f4, most of AMM's lead there disappears, suggesting that the probes' judgment of separability may be misleading after rotation.
On held-out MA-BBOB, AMM's final share at a budget of 2000*D matches IPOP-CMA-ES, but its anytime area is lower.
3 more timeline events
All Facts (13)
The effectiveness of any router is tied to the coverage of its probing assumptions, and the failure on rotated separable functions reminds users of this boundary condition.
50%UnverifiedEach of AMM's probes comes with default values, a dimension-budget validity range, and a probing cost counted against the total budget, so probing itself is not free, and routing decisions must trade off information gain against budget consumption.
50%UnverifiedAMM routes to five classes of engines: quasi-Newton, trust region, coordinate, multi-start, and evolution strategies.
50%UnverifiedAMM conducted a pre-registered development loop on the bbob benchmark, and 32 of 36 predictions held in the subsequent 16 confirmations.
50%UnverifiedAMM was tested once on held-out bbob instances 101-110 (D=2-40, budgets 10*D to 1000*D) against up to seven competitors; across 600 function-dimension-budget cells, AMM had the highest objective share in 382 cells (236 outright, 146 tied).
50%UnverifiedUnder the descriptive posterior rule, AMM credibly trailed on only 35 cells in the held-out bbob test, concentrated mainly in the low-budget regimes of 10*D and 30*D.
50%UnverifiedAt each budget tier, AMM also led the control groups built from its own engines—including the single best engine, the best engine selected by dimension and budget, and greedy scheduling—with the margin narrower at 10*D and 30*D; it also maintained a lead over nevergrad's NgIohTuned.
50%UnverifiedThe pre-registered runs showed that AMM's lead over IPOP-CMA-ES holds on affine-transformed and box-constrained bbob, and also holds on the combined results for rotated f2-f4.
50%UnverifiedAMM has one clear failure scenario: after rotating the separable f3 and f4, most of AMM's lead there disappears, suggesting that the probes' judgment of separability may be misleading after rotation.
50%UnverifiedOn held-out MA-BBOB, AMM's final share at a budget of 2000*D matches IPOP-CMA-ES, but its anytime area is lower.
50%UnverifiedUnder a known budget, AMM (Adaptive Multi Mode) uses online probing to judge in real time the curvature, plateaus, separability, kinks, basins, and fractal structure of the current landscape, then dynamically allocates budget to the best-matching optimization engine.
50%UnverifiedAMM's probing framework can be extended—adding new probes or adjusting validity ranges could improve routing quality on specific problem families.
50%UnverifiedAMM has been released as an open-source Python package, installable via pip install amm-optimizer, with code at github.com/BjBodner/amm-optimizer.
50%