For context, this occurred to me when I was trying to find a way to apply policy gradients to tic-tac-toe. I haven't compared it to the Gumbel-Softmax trick yet

For context, this occurred to me when I was trying to find a way to apply policy gradients to tic-tac-toe. I haven't compared it to the Gumbel-Softmax trick yet but empirically I can say that it works.