Hi guys. The link is: https://withablink.coding.me/goPolicyNet/ .

It's GPU-accelerated, search-free (pure policy network), and runs at ~0.2s per step on a desktop. If you are using a mobile device, please open it in Firefox / Chrome for maximal speed.

It's using my own DL library ( https://github.com/BlinkDL/BlinkDL ) . With some further optimization I estimate the speed can be 10x ~ 100x faster.

In the future we can also plug in Leela Zero's policy network.

Thank you, I enjoyed playing with it a few games (by always playing the #1 options for one color, and my own moves for the other color).
It is quite strong (I got in trouble a couple of times when I was not careful!), but it can also make some easy mistakes when fighting.

How was it trained?

Hi it's trained from human expert games (gogod, kgs, etc.)