Questions on submodule implementation

[xiao03]

Hi, I'm trying to implement MCTSnet recently, and your repo is very inspiring. I have several questions regarding the submodules.

1. The authors claim they use residual blocks in both the embedding network and the prior policy network. Each residual block contains two convolutional layers and a 'residual' step. Did you simplify it for the experiment?
2. The authors use MLP with ONE hidden layer in the backup network as well as the readout network. I think your code maps the input to output directly.

P.S. Have you tried to limit the path depth at each state?

[faameunier]

Hi,

Thanks for the questions !

1. For these two an implementation with residual blocks are implemented (Prior, Embedding). Looking at the annex A.2 of the paper, it specifies that they added a Conv2 input layer and Linear output, which is implemented here. We had troubles exactly understanding their architecture and tried our best to match what is described.

2. You are right, the model here is simplified. I can't remember exactly why I ended up implementing this one, but we went through a lot of back and forth and experimentation.

Regarding the path depth I don't think we tried limiting it by anyways. At each simulation we can in theory explore one step further (if the model always take the same path), which limits the maximum path depth by the maximum number of simulation.

Let me know if you manage to implement successfully the network !

[xiao03]

Hi, Thanks for answering my questions.

For Residual networks, I believe Appendix A.2 suggests that they use 2 and 3 blocks in prior policy and embedding network respectively. Therefore, there should end up at least 4 and 6 convolutional layers.

I attached my implementation for Residual Block, Embedding, and Prior network.
MCTSnet Submodule code.zip

[tklosek0]

@xiao03 Thanks for sharing your updated MCTSNet Modules.
Did you also updated the Backup & Readout Network? If so, would you mind sharing it? :)