That is a useful example, thank you, because it shows how to update the workers.
However, I don’t know which callback is suitable. Ideally, I would do it once and only once. This would suggest on_algorithm_init, but I believe (to be confirmed) that this is called after the init(), but before restoring the training, which presumably is calling load_checkpoint(), which will then override the policy_mapping_fn.
Do you have any thoughts on this matter please?