Network gives NaN after 30k trainings?

What is wrong?

I am working on a neural network, which implements reinforcement learning to play tetris. At one point during the training process with the learning rate set to 0.3 the network after 30k trainings gives the result 0 and then goes to NaN. Tried to figure out why Nan, but can’t seem to.

Where does it happen?

During the training process, when running a network with:

50 - input neurons
700 - first hidden
200 - second hidden
1 - output

How do we replicate the issue?

Just run the network with the [700, 200] neurons and learning rate of 0.3.

How important is this (1-5)?

2

Expected behavior (i.e. solution)

It should give at least 0 if nothing else

Other Comments

The entire code is here https://github.com/PetarIvancevic/node-neural-tetris . The code is run with:

node index.js -g NUM_GAMES -n NAME_OF_FOLDER

Ou yeah…the network LEARNS how to play…but it gives NaN xD…just weird. The blocks are fixated to always be OBlocks for testing and comparison.

[mubaidr]

Sure, will do some testing.

[PetarIvancevic]

You should “play” around 30k games, then it breaks. My latest guess is that it might be somehow node related, but I have no idea how 😕