camille-004 / group-testing-rl Goto Github PK

Log "pretraining..." when win_rate = 0.0, before n_epochs == history_window_size

Find a better way to fix repetition issue, rather than agent learning not to repeat, as this adds unnecessary training time and does not satisfy the constraints of the original problem

More seamless ways to tune training hyperparameters

Change maximum epsilon threshold when win rate seems to stagger (important!). If increase, this might increase accuracy but also training time by a lot.