Local Budget Benefits, Population Commitment, and Solver Performance: A DGO Case Study with Controlled DE Transfer
Organizations: School of Traffic and Transportation, Lanzhou Jiaotong University
Abstract
An evaluated candidate can improve a local search step, alter the population used by later steps, and affect a solver's final result. We distinguish these three questions using Da Gou Gun Optimization (DGO), a population optimizer designed for this study. Its JQ100 configuration evaluates an interpolation candidate during a two-sided probe. We define local budget value, population commitment effects and complete-policy performance, and give an exact example in which a locally better commitment leads to a worse continuation. Short-horizon matched comparisons favor interpolation on eligible quadratic states. Paired continuations yield mixed effect estimates, with primary group intervals spanning zero. Complete-policy comparisons favor JQ100 over the tested classical references in aggregate; a matched-start comparison has favorable means but intervals spanning parity. On 24 boxed BBOB functions, target-truncated geometric-mean error ratios favor JQ100 over the tested classical policies and favor L-SHADE and CMA-ES over JQ100 under the stated initialization and stopping rules. Local evaluation gains, commitment effects and solver competitiveness therefore require separate comparisons.
Figures & tables
| Arm | Improved trials | Rejected trials | Terminal best |
|---|---|---|---|
| KEEP | , | , | .0008 |
| DROP | , | , | 0 |
| Step | Comparison and study | Unit and principal scope |
|---|---|---|
| Basic effectiveness | Classical references (JQC1); same start (PC1) | Complete run; eight profiles, 10D/20D, FE |
| Local budget value | One/two-point branches (J2); 24-FE branches (J3) | Eligible state with three common observations already paid |
| Population commitment | Accepted-state DGO/DE replay (J6/M3); preselected types (CT1) | State, with four futures (J6/M3) or two (CT1); different populations |
| Complete performance | Same-policy references (JQX1); public suite (BX1); restart (PC1) | Complete run; native stopping or separately specified restart |
| Study | Problems | Instances seeds | Runs | Policies |
|---|---|---|---|---|
| JQC1 | 8 profiles, 6 families | 960 | JQ100, DE5D, PSO, Random | |
| PC1 | Same 8 profiles | 1,296 | JQ100, DE40, DE5D, CMA-native/restart, 4 perturbations | |
| JQX1 | 5 custom conditions | 450 | JQ100, NQ, S0, L-SHADE, CMA-ES | |
| BX1 | 24 boxed BBOB functions | 1,728 | JQ100, DE5D, PSO, Random, L-SHADE, CMA-ES |
| Reference | Ratio | 95% interval | Hits JQ/ref | Criteria | |
|---|---|---|---|---|---|
| 10 | Random | 62/0 | Basic met | ||
| 10 | DE | 0.3727 | [0.3276, 0.4211] | 62/45 | Improvement met |
| 10 | PSO | 0.1256 | [0.0602, 0.2590] | 62/50 | Improvement met |
| 20 | Random | 58/0 | Basic met | ||
| 20 | DE | 0.00219 | [0.00157, 0.00352] | 58/0 | Improvement met |
| 20 | PSO | 0.1506 | [0.0518, 0.4719] | 58/42 | Improvement met |
| Comparison | 10D [interval] | 20D [interval] | Worst 10D/20D |
|---|---|---|---|
| JQ100/DE40 | 0.6154 [0.4476, 1.23] | 0.4446 [0.1172, 1.711] | / |
| JQ100/DE5D | 0.2727 [0.1243, 0.4049] | 0.002412 [0.001373, 0.006559] | 1.077 / 5.528 |
| DE40/DE5D | 0.4431 [0.1666, 0.8073] | 0.005426 [0.002241, 0.01392] | 1.942 / 5.819 |
| Family | 10D ratio | 10D hits | 20D ratio | 20D hits |
|---|---|---|---|---|
| Sphere | 1 | 9/9 | 1 | 9/9 |
| Ellipsoid | 0.08621 | 11/9 | 0.5549 | 9/9 |
| Rosenbrock | 0.6816 | 0/0 | 0.8739 | 0/0 |
| Rastrigin | 0.2383 | 0/0 | 0.2399 | 0/0 |
| Ackley | 1 | 9/9 | 1 | 9/9 |
| Doublewell | 3.882 | 8/9 | 0.06637 | 7/5 |
| Study and extra FE | Comparison | NQ base run | JQ base run |
|---|---|---|---|
| J2, 1 FE | Interpolation / new direction | 0.951 | 0.944 |
| J2, 2 FE | Interpolation + direction / two directions | 0.956 | 0.945 |
| J3, 24 FE | Fresh-pair reuse / interpolant first | 1.019 | 1.015 |
| J3, 24 FE | Fresh-pair reuse / no reuse | 0.996 | 0.980 |
| Study | Dimension | States/base runs | Mean log ratio [95% interval] |
|---|---|---|---|
| DGO replay (J6) | 10 | 12/12 | |
| DGO replay (J6) | 20 | 12/12 | |
| DE transfer (M3) | 10 | 8/8 | |
| DE transfer (M3) | 20 | 8/8 |
| Group | States | I | Strict I | R | N | E | All four |
|---|---|---|---|---|---|---|---|
| Quadratic | 64 | 57 | 57 | 13 | 8 | 14 | 1 |
| Controls | 64 | 43 | 43 | 20 | 13 | 20 | 1 |
| All states | 128 | 100 | 100 | 33 | 21 | 34 | 2 |
| Reference | 10D ratio [interval] | 20D ratio [interval] |
|---|---|---|
| DE5D | 0.3667 [0.278, 0.4742] | 0.1349 [0.1139, 0.1695] |
| PSO | 0.2464 [0.1868, 0.302] | 0.5877 [0.4989, 0.6965] |
| Random | 0.001489 [0.001135, 0.001866] | 0.001876 [0.00158, 0.002358] |
| L-SHADE | 1.653 [1.235, 2.204] | 1.897 [1.529, 2.531] |
| CMA-ES | 1.681 [1.169, 2.306] | 4.706 [3.401, 6.421] |
| Method | 10D hits | 20D hits |
|---|---|---|
| JQ100 | 81 / 41 / 25 / 12 | 56 / 26 / 12 / 6 |
| DE5D | 61 / 28 / 19 / 3 | 23 / 5 / 0 / 0 |
| PSO | 54 / 27 / 18 / 7 | 41 / 27 / 19 / 12 |
| Random | 2 / 0 / 0 / 0 | 0 / 0 / 0 / 0 |
| L-SHADE | 86 / 54 / 38 / 20 | 67 / 26 / 18 / 8 |
| CMA-ES | 90 / 64 / 56 / 50 | 82 / 52 / 48 / 43 |
| Comparison | 10D ratio | 20D ratio | Evaluation scope |
|---|---|---|---|
| PC1: CMA-restart/native | 0.370 | 0.317 | 2,160,000 / 994,172 total FE across both dimensions |
| BX1: JQ100/CMA-ES at matched counts | 5.893 | 10.648 | Pair-specific counts: 860–10,000 FE at 10D; 2,640–20,000 at 20D |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Study | Comparison and sampled units | Scope of interpretation |
|---|---|---|
| JQC1 | At 10D and 20D: eight profiles in six families, three fresh instances and five seeds per profile, four policies; 960 runs. | Unchanged JQ100 versus fixed DE, PSO and Random; common paid budget, shared 40-point initialization prefix, six-family weighting. |
| J1-QS | 10D; eight conditions; three new instance blocks and three seeds per condition. Five policies; 360 held-out runs. | Complete policies with their own later trajectories. |
| J2-OC | 96 fresh base runs, 5,760 paid-endpoint states, 5,194 eligible; eight repetitions. J versus N costs one FE; JN versus NN costs two FE. | Local branches do not update the base run; average within state before equal-weight hierarchical log aggregation. |
| J3-DV | 1,440 fixed-rule selected fresh states; four paired repetitions of three 24-FE policies. The 2-, 4- and 8-FE prefixes were prespecified. | Local policies do not change the full population or share discoveries. |
| J5-PATH | Two selected J1 seeds on one rotated condition- ellipsoid instance; first accepted point at round 2, FE 121; continuation to 10,000 FE. | Original-stream KEEP recovery required; illustrative cases, not new held-out samples. |
| J6 | 48 selected states from 48 base runs; four paired continuations per state, 16 fresh-direction and 16 random-interior one-point references. | Same-candidate local and commitment estimates; futures are nested within states. |
| Group | Selected | Mean log ratio [95% interval] | Joint | Hits K, D |
|---|---|---|---|---|
| 10D quadratic | 12/12 | 3/12 | 25/48, 24/48 | |
| 10D all | 24/24 | 5/24 | 25/96, 24/96 | |
| 20D quadratic | 12/12 | 3/12 | 9/48, 10/48 | |
| 20D all | 24/24 | 3/24 | 9/96, 10/96 |
| Group | Selected | Mean log ratio [95% interval] | Joint | Stricter |
|---|---|---|---|---|
| 10D quadratic | 8/8 | 4/8 | 1/8 | |
| 10D all | 13/16 | 4/16 | 1/16 | |
| 20D quadratic | 8/8 | 4/8 | 2/8 | |
| 20D all | 16/16 | 4/16 | 2/16 |
| D | Reference | Error ratio | Hits | Worst ratio | Criteria |
|---|---|---|---|---|---|
| 10 | NQ | 0.204 | 10/9 | 5.259 | Met |
| 10 | S0 | 104.5 | 10/11 | Not met | |
| 10 | L-SHADE | 3.673 | 10/11 | 64.68 | Not met |
| 10 | CMA-ES | 10/18 | Not met | ||
| 20 | NQ | 1.536 | 6/7 | 58.41 | Not met |
| 20 | S0 | 0.3757 | 6/3 | 13.84 | Not met |
| Quantity | Mean | 95% interval |
|---|---|---|
| Interpolation I | -0.04023 | [-0.16865, 0.09088] |
| Random interior R | -0.01512 | [-0.08134, 0.04642] |
| New direction N | -0.00667 | [-0.06962, 0.05181] |
| Better endpoint E | 0.01610 | [-0.03702, 0.06556] |
| I with fixed radii | 0.03336 | [-0.11766, 0.19132] |
| I minus R | -0.02511 | [-0.15232, 0.10693] |
| Numerator / denominator | 10D | 20D | Worst 10D | Worst 20D |
|---|---|---|---|---|
| JQ100 / DE40 | 0.61545 | 0.44457 | ||
| JQ100 / DE5D | 0.27272 | 0.002412 | 1.077 | 5.528 |
| DE40 / DE5D | 0.44313 | 0.005426 | 1.942 | 5.819 |
| CMA-restart / native | 0.36981 | 0.31661 | 1 | 1 |
| / JQ100 | 0.90224 | 0.79874 | 337.63 | 4.789 |
| / JQ100 | 0.85404 | 0.96652 | 10.791 |
| Reference | Raw W/T/L | Worst primary pair | |
|---|---|---|---|
| 10 | DE5D | 130/0/14 | 180.302 |
| 10 | PSO | 95/0/49 | 7.847 |
| 10 | Random | 144/0/0 | 0.671 |
| 10 | L-SHADE | 53/2/89 | 191.343 |
| 10 | CMA-ES | 45/1/98 | 610.384 |
| 20 | DE5D | 136/0/8 | 9.347 |
| Reference | Target clip | ||||
|---|---|---|---|---|---|
| 10 | S0 | 548.776 | 104.470 | 3.304 | 0.405 |
| 10 | CMA-ES | 41424.8 | 2.165 | ||
| 20 | S0 | 0.376 | 0.376 | 0.376 | 0.661 |
| 20 | CMA-ES | 618615 | 33.380 |
Metaheuristic Optimization
Metaheuristic optimization develops flexible search strategies for finding strong solutions to difficult problems when exact methods are too costly.
See all 122 papers