[flyingdoubleG]

Description

SIBEA's environmental selection keeps more than pop_size survivors, so the population grows over the run:

import torch
from evomo.algorithms import SIBEA
from evomo.problems.numerical import DTLZ2
from evomo.workflows import UnifiedWorkflow

torch.manual_seed(0)
algo = SIBEA(pop_size=100, n_objs=3, lb=torch.zeros(12), ub=torch.ones(12))
wf = UnifiedWorkflow(algo, DTLZ2(d=12, m=3))
wf.init_step()
for _ in range(30):
    wf.step()
print(algo.pop.shape[0])  # 343; it is 100 after the fix

Per generation: 139, 100, 145, 159, 164, ..., 318, 334, 343. After 200 generations, the population holds 4385 solutions.

Cause. non_dominate_rank returns zero-based ranks (see AGENTS.md), but the front counting follows PlatEMO's one-based FrontNo:

ranks = torch.arange(1, max_rank + 1, device=device)

So rank 0 is never counted, and the cumulative counts start at the second front. As a result:

• max_f_no is at least 1, and mask_keep = front_no < max_f_no always keeps the whole first front.
• num_needed = N - num_kept goes negative once the first front is larger than N, and sort_idx[:num_needed] then drops only the last few solutions of the next front instead of selecting nothing.

Fix. Count fronts from rank 0. With that change, num_kept < N <= num_kept + |last front|, and exactly N solutions survive.

Testing.

• test_sibea_keeps_population_size runs 10 generations and checks the population and fitness sizes. Before the fix it fails with 38 != 20; with the fix it passes.
• unit_test/algorithms/test_moea_generated.py passes.
• After 200 generations on DTLZ2 (m=3) the population stays at 100, with IGD 0.073. On ZDT1 the IGD is 0.0056, compared with 0.0052 for NSGA-II under the same budget.

Checklist

• I have formatted my Python code with ruff.
• I have good commit messages.
• If adding new algorithms, problems, operators:
  • Added related test cases.
  • Added docstring to explain important parameters.

[Zhenyu2Liang]

Thank you for fixing the bugs.