0
$\begingroup$

I'm currently working on a Deep Reinforcement Learning model for the game "connect 4". Before I started I read some rules and facts about "connect 4". Thats when I noticed after years of playing it, that there is actually no fixed play-field size. Common sizes are e.g. 8x8, 8x7, 7x8. A human might adapt to that quickly. Not so a NN.

My first question is about a fully connected NN. Is my assumption correct, that a model trained on one size will perform poorly on a different size?

Another idea would be the use of CNN. Could this maybe work? I could zero pad different games to e.g. 10x10. Then the model might be able to play different sizes.

Are there maybe other similar examples of generalization of different input arrays? Only found some experimental papers about CNN with a special "pyramid pooling".

$\endgroup$

1 Answer 1

2
$\begingroup$

It's hard to transfer a fully connected NN to inputs of differing sizes.

Convolutional neural networks are fairly flexible when it comes to input size. As long as you use some sort of global average pooling on the last conv layer, there is nothing which prevents you from feeding inputs of variable sizes.

If you try to generalize to different board sizes while training only on one, I'd expect it to do poorly, but better than random. If you're able to train on all board sizes, it may do well.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.