[Bug] ValueError: high is out of bounds for int32 on Windows (vecrl.py, SyncVectorEnv)

[FredericThoma]

Description

Running any script that uses VecGymNE (e.g. with PGPE or CoSyNE) on Windows fails with the following error:

ValueError: high is out of bounds for int32

Root cause

In evotorch/neuroevolution/net/vecrl.py, line 1661:

seeds = self.__random_state.randint(0, 2**32, self.num_envs)

On Windows, NumPy's default integer type is int32, so passing 2**32 as the
high argument to numpy.random.RandomState.randint exceeds the maximum value
for that dtype and raises a ValueError.

On Linux (and WSL), NumPy's default integer type is int64, so the same call
succeeds.

Steps to reproduce

1. Use Windows (native, not WSL)
2. Run this script:

from evotorch.algorithms import PGPE
from evotorch.decorators import pass_info
from evotorch.neuroevolution import VecGymNE
from torch import nn


@pass_info
class LunarPolicy(nn.Module):
    def __init__(self, obs_length: int, act_length: int, **kwargs):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(obs_length, 16),
            nn.Tanh(),
            nn.Linear(16, act_length),
        )

    def forward(self, x):
        return self.net(x)


problem = VecGymNE(
    env="LunarLander-v3",
    network=LunarPolicy,
    num_episodes=1,
    num_actors=1,
)

searcher = PGPE(problem, popsize=10, radius_init=4.5, center_learning_rate=0.2, stdev_learning_rate=0.1)
searcher.step()

print("OK — no error (you are likely on Linux/WSL)")

4. Observe the ValueError

Full traceback

Traceback (most recent call last):
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\examples\scripts\reproduce_bug.py", line 42, in
searcher.step()
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\algorithms\searchalgorithm.py", line 390, in step
self._step()
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\algorithms\distributed\gaussian.py", line 354, in _step_non_distributed
fill_and_eval_pop()
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\algorithms\distributed\gaussian.py", line 295, in fill_and_eval_pop
self.problem.evaluate(self._population)
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\core.py", line 2559, in evaluate
self._evaluate_all(batch)
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\core.py", line 2577, in _evaluate_all
self._evaluate_batch(batch)
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\neuroevolution\vecgymne.py", line 738, in _evaluate_batch
self._evaluate_subbatch(batch)
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\neuroevolution\vecgymne.py", line 752, in _evaluate_subbatch
obs_per_env = env.reset()
^^^^^^^^^^^
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\neuroevolution\net\vecrl.py", line 492, in reset
reset_result = self.env.reset(*args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\neuroevolution\net\vecrl.py", line 1776, in reset
seed_kwargs_list = self.__pop_seed_kwargs()
^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users<USER>\Documents\EvotorchFork\evotorch\src\evotorch\neuroevolution\net\vecrl.py", line 1661, in __pop_seed_kwargs
seeds = self.__random_state.randint(0, 232, self.num_envs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "numpy/random/mtrand.pyx", line 799, in numpy.random.mtrand.RandomState.randint
File "numpy/random/_bounded_integers.pyx", line 2881, in numpy.random._bounded_integers._rand_int32
ValueError: high is out of bounds for int32

Proposed fix

Add dtype=np.int64 to the randint call:

seeds = self.__random_state.randint(0, 2**32, self.num_envs, dtype=np.int64)

Workarounds

• Recommended: Run inside WSL (Windows Subsystem for Linux)
• Manual patch: Edit .venv\Lib\site-packages\evotorch\neuroevolution\net\vecrl.py
  line 1661 as shown above

Environment

• OS: Windows (native)
• Python: Python 3.12.10
• NumPy: 2.4.4
• EvoTorch: 0.1.dev80+gcebcac4f2