How to initialize network to output zeros?

Hi,

I’m currently trying to implement the algorithm found in this paper : https://arxiv.org/pdf/1811.00164.pdf.

However, at some point, the algorithm need to initialize the neural network to return 0 for all inputs (line 1).

So my question is simply how do I do that ? At first I thought to initialize the weights of the output layer to zero but it will prevent it from learning. I’m sure there is a simple method but I can’t find it.

It won’t, this only applies to hidden layers (that generate “artifical” features). For output layer, loss (eq.: distance to target) provides a non-zero gradient.

Hi, I was wondering if you had any luck figuring this out? I came across this topic because I was actually trying to implement that same algorithm, and I had the same question.

I tried your suggestion regarding the output layer weights, and didn’t get any improvement in performance. My advantage networks end up with a loss of around 30k mean squared error and I converge around 500mbb exploitability. I can’t figure out where I’m going wrong in replicating the neural net of the paper, but this is one line that I don’t understand. I’ve never seen such a technique discussed anywhere.

Hi, initializing the weights of the output layer to zero seemed to work in my case. You can still look on public implementation of deep cfr such as :

or

Hope that helps