Double-Dueling-DQN stops learning

[florath]

trafficstars

Running the Double-Dueling-DQN code results in a network, that stops learning after about 2000 episodes., i.e. the game results do not get better. Run the GridWorld example now four different times - and tried to adapt parameters: all result in mostly the same picture. The network has some 'good' learning curve at the beginning and then stops learning. For some results see:

1. run001
2. run002
3. run003
4. run004

I also started using Breakout-v0 - with mostly the same result.

Does anybody have an idea? Which parameters can be adapted to get better results?