Open cosmicBboy opened 4 years ago
idea 9: use Random Network Distillation, applied to the value of the next time step
idea 10: try Q-actor critic method instead of advantage function
idea 11: use simpler policy architecture, with multivariate normal to jointly produce all hyperparameters instead of sequentially with an RNN
idea 12: use model-based RL to estimate the reward function (function approximator can even be gaussian process!)
Noting these down for the neurips bbo challenge
n_suggestions
ranked by value.n_suggestions
and set the reward for the lowestm
suggestions to -1 (this should happen in the observe step)x
factor more suggestions thann_suggestions
and send topn_suggestions
ranked by value. Use all of the suggestions to update the controller, setting the rewards for the suggestions that didn't make it to -1 (this should happen in the suggest step)