Program crashes with `min_prevalence = 90` and `language = binary`

[LuciecLaurent]

Description

When using min_prevalence = 90 combined with a language = binary, the feature selection results is not to small, so why is not working ? Instead of exiting gracefully with a clean error message, the program panics. The program works with min_prevalence = 90 and when language' is not binary`. Why is not working with those parameter ?

Steps to Reproduce

1. Set min_prevalence = 90
2. Use a language = binary
3. Run the program
4. Observe the panic

param.yaml

Actual Behavior

The program panics:

thread '<unnamed>' (38) panicked at src/ga.rs:104:9:
Initial population size is too small (30<50 individuals)!
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

thread 'main' (19) panicked at src/main.rs:157:29:
Thread panicked!: Any { .. }

Logs

2026-03-05 23:13:19 [INFO] Training using Genetic Algorithm
2026-03-05 23:13:19 [INFO] Selecting features...
2026-03-05 23:13:19 [WARN] Class "PD" has only 9 significant features based on required threshold ! All features kept for this class.
2026-03-05 23:13:19 [WARN] Class "PR/CR" has only 5 significant features based on required threshold ! All features kept for this class.
2026-03-05 23:13:19 [INFO] 14 features selected
2026-03-05 23:13:19 [ERROR] Initial population size is too small (30<50 individuals)!

We can see there are 14 features, it should work.

Environment

• min_prevalence: 90
• language = binary
• Affected files: src/ga.rs:104, src/main.rs:157

Labels

bug, feature

[eprifti]

Fix applied in commit 48edec8

Root cause: With 14 features and binary language, only ~30 unique individuals can be generated. The hard minimum of 50 was too strict.

Changes:

• Lowered hard minimum population size from 50 to 2 (minimum needed for crossover).
• When population is < 50, a warning is now emitted instead of crashing.
• Added descriptive error messages that report feature count, language, and prevalence settings when the population is too small.
• Backtraces are suppressed in favor of clean [ERROR] log lines.

A run with 14 features + binary language producing 30 individuals will now proceed with a warning.

[LuciecLaurent]

Context

This is a follow-up to the ga.rs:641 panic ("Vec must have exactly 2 elements") that occurs with language: bin and feature_minimal_prevalence_pct : 90.

Finding 1 — population_size has no effect on the panic

Reducing population_size does not prevent the crash:

population_size	select_elite_pct (default: 2)	param.yaml	log	Result
1000	2	param.yaml	round_2026-06-17_09-09-55_2026-06-17_09-09-55.log	panic
100	2	param.yaml	round_2026-06-17_09-09-55_2026-06-17_09-09-55.log	panic
50	2	param.yaml	round_2026-06-17_09-09-56_2026-06-17_09-09-56.log	panic
9	2	param.yaml	round_2026-06-17_09-09-56_2026-06-17_09-09-56.log	panic

Finding 2 — select_elite_pct is the key parameter

With population_size: 9 and select_random_pct: 0:

select_elite_pct	elite parents = floor(9 × pct%)	Vec content	param.yaml	log	Result
2, 11, 12	0–1	[]	param_elite_pct_2, param_elite_pct_11,param_elite_pct_12	elite_pct_2_round_2026-06-17_08-28-44_2026-06-17_08-28-44.log, elite_pct_11_round_2026-06-17_08-28-45_2026-06-17_08-28-45.log, elite_pct_12_round_2026-06-17_08-28-45_2026-06-17_08-28-45.log	panic
22, 25, 34	2–3	[1 feature]	elite_pct_22_param.yaml, elite_pct_25_param.yaml, elite_pct_34param.yaml	elite_pct_22_round_2026-06-17_08-28-46_2026-06-17_08-28-46.log, elite_pct_25_round_2026-06-17_08-28-46_2026-06-17_08-28-46.log, elite_pct_34_round_2026-06-17_08-28-47_2026-06-17_08-28-47.log	panic
45, 56	4–5	[1 feature]	elite_pct_45_param.yaml, elite_pct_56_param.yaml	elite_pct_45_round_2026-06-17_08-28-47_2026-06-17_08-28-47.log, elite_pct_56_round_2026-06-17_08-28-48_2026-06-17_08-28-48.log	panic
78	7	—	param.yaml	round_2026-06-17_08-28-48_2026-06-17_08-28-48.log	OK ✓
100	9	—	param.yaml	round_2026-06-17_08-28-49_2026-06-17_08-28-49.log	OK ✓

With population_size: 1000 and select_elite_pct: 78 → also OK ✓ (param.yaml, round_2026-06-17_08-31-16_2026-06-17_08-31-16.log)

Workaround

ga:
population_size: 1000
select_elite_pct: 78
select_niche_pct: 0
select_random_pct: 0

Caveat

This workaround keeps 780 out of 1000 individuals as elite parents each generation, severely limiting the GA's exploration capacity. It is a patch, not a fix — the root cause (no graceful handling of an empty or near-empty discriminative feature set) remains in ga.rs:641.