why n_tasks * alpha?

[mlresh]

Hi,

I noticed that in the experiments, epo_train.py scales alpha (task weight) by the number of tasks but not for the other methods like pmtl_train.py and linscalar_train.py. epo_search.py in toy_experiments also uses a different multiplier x = x - 10. * max(lp.mu_rl, 0.1) * step_size * d_nd.

Is there any particular reason behind these choices?

Thanks.

[wjm-wjm]

Hi,

I noticed that in the experiments, epo_train.py scales alpha (task weight) by the number of tasks but not for the other methods like pmtl_train.py and linscalar_train.py. epo_search.py in toy_experiments also uses a different multiplier x = x - 10. * max(lp.mu_rl, 0.1) * step_size * d_nd.

Is there any particular reason behind these choices?

Thanks.

I also have this question, do you solve it now ?