Punishment update

This commit is contained in:
Efim Beshmenev
2026-07-11 11:11:28 +03:00
parent 5ba553a1d4
commit a8a8f50fbd
10299 changed files with 11355 additions and 4516 deletions
+10
View File
@@ -39,6 +39,11 @@ be starved indefinitely. A clean stop saves each
topology under `results\search\topology_N`; pressing the same button later resumes only
those checkpoints. `results\search\leaderboard.tsv` is the current ranking.
The GUI search-mode selector keeps this quality-first allocation as the default. Select
`Проработка худших` to reverse only the quality prior: depth work then favors topologies
with the worst current C/I result. Improvement reward, UCB exploration, staleness, and the
full refresh remain active, so the mode is a priority rather than an exclusive lock.
The CUDA hot path uses standard FP32 math. Persistent chains retain their RNG, cooling,
stagnation, and restart state between batches. Twenty-five percent of chains remain the
unchanged simulated-annealing control; disjoint cohorts add replica exchange, adaptive
@@ -51,6 +56,11 @@ the CPU in `double`; candidates with at most 12 total defects are reclassified w
custom `WideReal` double-double type (about 31 decimal digits). Every claimed `0/0` is
serialized and checked again at multiple tolerances down to `1e-13`.
The degeneracy guard is deliberately mild. Its default weight is `0.01`, and only the
worst of the determinant, edge-ratio, turn-angle, and extent barriers is counted fully;
the other three contribute five percent. This still rejects catastrophic collapse without
letting several correlated moderate warnings overpower an otherwise useful candidate.
The MAP-Elites archive separates geometry into determinant, edge-ratio, turn-angle, and
extent bins. All valid historical checkpoint generations are deduplicated and re-evaluated
on startup, so accumulated searches seed the archive instead of merely supplying one best