RNG leads to repeated random numbers or non-reproducible results

[tautomer]

Problem:

• When a seed is provided, the package resets the random number generator (RNG) states multiple times within a single campaign:

  • The optimizer, model, designer, and DOE components each reset RNG states independently.
  • During a normal workflow (without explicitly defining the optimizer/designer), random seed reset happens when
    • the optimizer is initialized (torch and numpy)
    • the designer is initialized (torch only)
      • if DOE is used (numpy only)
    • every iteration when each model is instantiated (torch and numpy)

• When no seed is provided, there is no RNG control at all, resulting in no reproducibility.

Impact:

Repeated RNG resets cause correlated and repeated noise that biases results. In the model fitting stage, the model and scipy/Adam optimizer will always start from the same initial conditions. Additionally, global seed control is fragile, as unexpected random draws between steps break reproducibility unless perfectly replicated.

Proposed Fix:

• Centralize RNG state management to avoid multiple independent resets.
• Ensure the simulator noise RNG is initialized once per campaign and not reset every iteration.
• Provide consistent RNG control whether or not a seed is specified, enabling reproducibility.
• Refactor the code to use a single RNG state or RNG generator instance passed explicitly to components that require randomness.
• Add a package level toggle to switch back to current behavior for backward compatibility.