AppliedDataSciencePartners / DeepReinforcementLearning

A replica of the AlphaZero methodology for deep reinforcement learning in Python
GNU General Public License v3.0
2.02k stars 758 forks source link

About learning rate and loss decrease #32

Open xuehy opened 5 years ago

xuehy commented 5 years ago

Sorry to bother you! I am making an alphazero implementation similar to yours, which is also for the Connect4 board game. The training went smooth at first, however, after 70+ iterations, the loss can no longer decrease. I manually set the learning rate from 1e-3 to 1e-5, but the loss still gradually increases. Then I came across your blog about your implementation, and I find it very similar to mine. Have you ever met this case in your experiments? Hopefully you could offer me some advice :)