Can't get results for LunarLander

[TrentBrick]

Hi thanks for sharing your code and implementation.

However, when running your notebook with the LunarLander-v2, even with 1,000 epochs it doesn't seem to be learning anything:

Can you share the hyperparameters necessary to reproduce the LunarLander?

Thank you.

[BY571]

hello @TrentBrick !
Sure, here the hyperparameter:
"horizon_scale" : 0.01,
"return_scale" : 0.025,
"replay_size" : 500,
"n_warm_up_episodes" : 10,
"n_updates_per_iter" : 100,
"n_episodes_per_iter" : 20,
"last_few" : 75,
"batch_size" : 768,
"layer_size" : 128,
"learning_rate" : 1e-3

[TrentBrick]

Thanks for getting back to me and sharing these parameters. I just tried them but they also dont seem to be working for me.

[BY571]

oh, I just noticed that there is an older network architecture. try to add two linear layers, it should work then.
Thanks for noticing, ill update the code!

[TrentBrick]

Just heads up I've moved on to using other repos for this. Spent too long trying to run yours. You should double check that everything now works though for future people that come across this repo.

[tangzk]

oh, I just noticed that there is an older network architecture. try to add two linear layers, it should work then. Thanks for noticing, ill update the code!

@BY571 Have you updated the latest code? I got results similar to TrentBrick in LunarLander-v2 env.

[TrentBrick]

@tangzk if you want to use my repo for this... https://github.com/TrentBrick/RewardConditionedUDRL