search

rmattson1008 / dqn_asteroid_probe

Probing various layers of DQN to assess "difficulty" of decisions.
MIT License
0 stars 0 forks source link

defining difficulty #4

Open rmattson1008 opened 1 year ago

rmattson1008 commented 1 year ago

Atari has difficulty settings [0,3]. Probably best to create two probe sets, 1 for 0 and 1 for 3. Then I can compare results to see if any differences. I feel like I need to dive into the data and say like, this has 3 asteroids this only has 1. I think I have to train the model on difficulty=3, because this setting has extra game play the model needs to see. introduces other issues with the representational quality.

rmattson1008 commented 1 year ago

For such a short project I think it will be ok to just look at various state spaces and pick certain things.