ResNet Parameters

TurboZero trains a configurable Residual Neural Net with policy and value heads. The user may specify parameters governing the model's size. The project does not support using custom architecture at this time, but that feature will eventually be added.

TurboZero currently uses identical architecture to AlphaZero (depicted here), with a few configurable parameters.

Configuration Parameters

• res_channels: number of channels in each residual block
• res_blocks: number of residual blocks
• kernel_size: size ($n$ x $n$) of each convolutional kernel in the residual blocks (recommend 3x3)
• value_fc_size: size of final fully-connected layer in the value head
• value_output_activation: activation function prior to value head output, supported options are [relu, tanh, sigmoid, tanh0to1]. The empty string may also be provided, which will result in no activation.