Antoine
3289a76e19
- 8-agent OpenClaw cluster (Manager, Tech-Lead, Secretary, Auditor, Optimizer, Study-Builder, NX-Expert, Webster) - Orchestration engine: orchestrate.py (sync delegation + handoffs) - Workflow engine: YAML-defined multi-step pipelines - Agent workspaces: SOUL.md, AGENTS.md, MEMORY.md per agent - Shared skills: delegate, orchestrate, atomizer-protocols - Capability registry (AGENTS_REGISTRY.json) - Cluster management: cluster.sh, systemd template - All secrets replaced with env var references
40 lines
2.2 KiB
Markdown
40 lines
2.2 KiB
Markdown
# LAC Critical Lessons — NEVER FORGET
|
|
|
|
These are hard-won insights from past optimization sessions. Violating any of these will cause failures.
|
|
|
|
## NX Safety (CRITICAL)
|
|
- **NEVER kill ugraf.exe directly** → use `NXSessionManager.close_nx_if_allowed()`
|
|
- **PowerShell for NX journals** → NEVER use `cmd /c`
|
|
- **Always load `*_i.prt` before `UpdateFemodel()`** → mesh won't update without the idealized part
|
|
- **File chain must be intact:** `.sim → .fem → *_i.prt → .prt` (ALL must be present)
|
|
|
|
## Optimization (CRITICAL)
|
|
- **CMA-ES doesn't evaluate x0 first** → always call `enqueue_trial(x0)` to evaluate baseline
|
|
- **Surrogate + L-BFGS = DANGEROUS** → gradient descent finds fake optima on surrogate surface
|
|
- **NEVER rewrite `run_optimization.py` from scratch** → ALWAYS copy a working template (V15 NSGA-II is gold standard)
|
|
- **Relative WFE math:** use `extract_relative()` (node-by-node subtraction) → NOT `abs(RMS_a - RMS_b)` (wrong math!)
|
|
|
|
## File Management (IMPORTANT)
|
|
- **Trial folders:** `trial_NNNN/` — zero-padded, never reused, never overwritten
|
|
- **Always copy working studies** — never modify originals
|
|
- **Output paths must be relative** — no absolute Windows/Linux paths (Syncthing-compatible)
|
|
- **Never delete trial data mid-run** — archive after study is complete
|
|
|
|
## Algorithm Selection (REFERENCE)
|
|
| Variables | Landscape | Recommended | Notes |
|
|
|-----------|-----------|-------------|-------|
|
|
| < 5 | Smooth | Nelder-Mead or COBYLA | Simple, fast convergence |
|
|
| 5-20 | Noisy | CMA-ES | Robust, population-based |
|
|
| > 20 | Any | Bayesian (Optuna TPE) | Efficient with many variables |
|
|
| Multi-obj | Any | NSGA-II or MOEA/D | Pareto front generation |
|
|
| With surrogate | Expensive eval | GNN surrogate + CMA-ES | Reduce simulation count |
|
|
|
|
## Common Failures
|
|
| Symptom | Cause | Fix |
|
|
|---------|-------|-----|
|
|
| Mesh not updating | Missing `*_i.prt` load | Load idealized part first |
|
|
| NX crashes on journal | Using `cmd /c` | Switch to PowerShell |
|
|
| Baseline trial missing | CMA-ES skips x0 | Explicitly enqueue baseline |
|
|
| Optimization finds unphysical optimum | Surrogate + gradient | Switch to CMA-ES or add validation |
|
|
| Study can't resume | Absolute paths in script | Use relative paths |
|