few questions about setting hyperparameters

[Nightbringers]

In go game, the default value for "max_num_considered_actions" is 16. I want to know if this value should increase with the increase of num_simulations？Also maxvisit_init ? What value is better for max_num_considered_actions when num_simulations is 400 and 800 and 1600? There are a few other hyperparameters that I'm not sure how to set optimally. I hope you can help me check how the settings are. Do I need to adjust qtransform? for go 19*19, my settings is: policy_output = mctx.gumbel_muzero_policy( ... num_simulations=400, max_num_considered_actions = 32, qtransform=partial( mctx.qtransform_completed_by_mix_value, value_scale=0.1, maxvisit_init=50, rescale_values=False, ), gumbel_scale=1.0, )

[fidlej]

Thanks for asking. The rescale_values=False is good for Go. You can then use value_scale=1.0. Otherwise, I would keep the defaults unchanged.

I do not have many results from 19x19 Go. On 9x9 Go, the max_num_considered_actions was not very important, as you can see from Figure 3b. https://openreview.net/pdf?id=bERaNdoegnO