search

roclark / super-mario-bros-dqn

DQN model used to train and beat Super Mario Bros. for the NES using PyTorch
MIT License
32 stars 12 forks source link

Unable to reproduce the results shown in the demo using the .dat files provided. #4

Open squishytofu042 opened 4 years ago

squishytofu042 commented 4 years ago

I'm using the following code in my terminal:

python train.py --action-space complex --environment SuperMarioBros-1-1-v0 --transfer

using the SuperMarioBros-1-1-v0.dat file provided by you in the pretrained_models folder, in order to view the results from the parameters stored in the .dat file. However, the results are not at all similar to the ones shown in your demo under the 'Progress' section.

Could you please double check this. I'm currently having this problem with your super-mario-bros-dqn project. I haven't yet tested it on the super-mario-bros-a3c one.

squishytofu042 commented 4 years ago

UPDATE: I noticed the agent does perform almost as good as the gif when looking at the video in recording/run1. I've left the network running via:

python train.py --action-space complex --environment SuperMarioBros-1-1-v0 --render --transfer

But looking at the video now (currently on run53), the agent is dying on the first Goomba now.

Why does this happen?