search

luwo9 / bomberman_rl

Reinformcement learning for Bomberman: Machine Learning Essentials lecture 2024 final project
0 stars 0 forks source link

Neural Network designs and features #8

Open luwo9 opened 2 months ago

luwo9 commented 2 months ago

Maybe it is worth making a CNN part of the Network, for processing the 2D field arrays of the states. Maybe its beneficial convert all the features where its possible to such an array and make it feature maps of the "image".

That is, i suppose a neural network might not learn good if one input are lists of x,y coordinates and the other one is a field array (~ one hot version), field arrays could work better and maybe CNNs are worth it. (Local correlations like bomb/coin/wall/crate next to agent may be of interest and somewhat translationally invariant and it is maybe more effective?)

RuneRost commented 2 months ago

I think one-hot encoded input representing whole grid should be easier to learn for neural network and helps because otherwise we don't hav fixed size. However it may be worth trying if a FCC works as well since feature maps are relatively small (6x6?).

luwo9 commented 2 months ago

I agree, probably CNN and fully connected networks are both worth trying out. Looking at the states again i am not sure if there really are things that are not best one-hot/map-like encoded.

luwo9 commented 2 months ago

Besides more informed/biased feature transformes, which are expected to be neccesary for less advanced regression models and probably worth trying out for neural networks aswell, one should probably also try to transform the less informed features well.

For example in the image-like state encoding (which is not very informed) maybe a bomb should not be encoded by its position but by its possible blast coordinates.