I made this project for fun. It is based on this video: https://www.youtube.com/watch?v=UGDQccLdZWc It lets you make really cool influence visualizations with a high degree of customization. Please enjoy it.

Main page: https://github.com/featurecat/Influencie
Download: https://github.com/featurecat/Influencie/releases

Note: there is no AI involved in generating the influence heatmap. It's a simple mathematical function with a basis in Go.

Neat.

In case you did want to use AI and have the influence map be much more precise, for example correctly ignoring very-dead stones, over the last few months, I've been experimenting with (among many other things) training neural nets that predict the final territorial ownership of each point on a board. Here's a screenshot of prediction by a neural net that is fairly small and whose playing strength is probably about 2 dan amateur or so.



I'm hoping in the not-too-distant future to (among many other things) scale up and get such a neural net up to beyond human pro strength and still able to do such visualizations.

That's very cool. I was thinking about this very problem (training a neural net to identify territory) recently. What was your training method?

I've been running some small-scale self-play training loops (like AlphaGoZero) with experimental modifications, such as recording also who owned each spot at the end of the game, and with a partial bonus for maximizing score during the self play, so as to encourage the bot to actually end up owning the regions that it "should own" and finish capturing dead stones, rather than being content to "give up" points as long as it's still winning. The ownership map was trained off of that.