Discuss optimizer when tensor computation on top of PyTorch.

[linjing-lab]

When initialize subclass, it's only determine None values of subclass data and complete inheritance of TorchOptimizer, which actually completes direct use of torch's optimizer in tensor computation under the premise of torch.optim alignment.

evotorch/optimizers.py#L165

super().__init__(torch.optim.Adam, solution_length=solution_length, dtype=dtype, device=device, config=config)

This behavior on top of PyTorch must provide high compatibility for reinforcement learning when adam or sgd used in workflow, but simply wrapped optimizer didn't served more concrete prompts for usages in evotorch, especially examples in torch and rl perspective integrations, so examples are project whose convergence given by Adam or SGD and executed steps when None check before typing values of TorchOptimizer. This issue stands for evolutionary traits important than specific optimizer in local search.