search

amaneureka / T-Rex

MLP based network which learns and play T-Rex game all by itself.
http://amaneureka.github.io/T-Rex/
295 stars 37 forks source link

[MLP] network gets confused #1

Open amaneureka opened 8 years ago

amaneureka commented 8 years ago

In a long run, sometimes the network gets confused and plays lame from next game.

amaneureka commented 8 years ago

Possible solution: Exponentially decrease learning rate proportional to confidence of network.

yask123 commented 8 years ago

Why does it get confused ?

amaneureka commented 8 years ago

Solution itself explains it. It's not good idea to move saturated weights with small momentum when they are already saturated. Hence it should decrease it's learning rate as network get confident on its decision.