Open TheMightiestCarrot opened 1 year ago
Hi, thanks for the question.
Therefore i would like to incorporate some kind of reinitialization of whole search (basically starting trainer.run with multiple different seeds).
The seed
that trainer accepts is mainly for environmental randomness.
For policies, you can programmatically supply different initial values (e.g., here is the code in PGPE)
Let me know if this answers your question.
Hello,
i have a task with unknown global optima and since optimizers can stuck in local optima i want to make sure the achieved optima is reached from various random starting points. Therefore i would like to incorporate some kind of reinitialization of whole search (basically starting trainer.run with multiple different seeds). Is it even necessary? Does SimManager -> eval_params -> _for_loop_eval -> policy_reset_func perform reliable reinitialization of policy state?
Thanks in advance for your advice.