Beneficial to play against AI?

Published 29 May 2020 by Raivo (last edited 30 May 2020)
tags: training, ai, katago

With how accessible strong AI are, is it worth it to play handicap games against them as a way to get better?

In martial arts there are often set forms that are practiced, Kata. And inspired by this I have been playing alot of handicap games versus KataGo. So far I have gone from not beating katago even with 9 stone handicap, to now beating KataGo with only 8 stones in only 1 week of playing a single slow game against katago a day and reviewing it. In the reviews I try to focus on the direction play, not so much the sequences that the AI suggests.

My thinking behind it is that KataGo is strong and plays well. So the moves that are played against me in these games are good moves, thus I'm constantly being shown proper ways to play, and punish weak/slack moves that I play.

I try to balance this AI play with faster games on OGS/IGS to get some repetition in, and my conclusion after this week of 'KataGO Kata' is that it has improved my play. But I'm worried that I shall develop bad habits from too much play vs. AI.

There seem to be tangible benefits to this type of training, but are there any pitfalls or drawbacks?
What are your experiences or thoughts on using these new breeds of open AI's in your go training regime?


Comments (4)


antti wrote 3 years, 10 months ago:

Playing the AI with a high handicap is probably analogical to playing a strong human with a high handicap, although after the game you don’t hear the human’s comments but instead get to check the AI’s reading.

I think the setting works in that it gives you an idea on how a strong player would play in particular local situations. The downside is in the ‘particular’ word: high-stone handicap games feature notably different situations from even games, and therefore the sequences that you learn in high-stone handicap games will not always translate well to even games. As you get better at defeating the AI, however, you will gradually reduce the handicap and learn techniques more relevant to even games.

What I would be careful about is the phenomenon that playing a lot of high-handicap games as black can systematically shift your playing style to become too defensive. If you have 9 or 8 handicap stones at the start, you can make significant concessions during the game and still win easily. In an even game, you need to know how to keep a whole-board balance, and so the general way of thinking is quite different from handicap games. Additionally, in even games you sometimes need to know how to create complications when you’re behind; and this is something you really cannot learn while playing handicap games as black.

My above feedback is from considering your situation as analogical to playing with a strong human, but it is possible that there actually are notable differences that I cannot see or think of. Therefore, I think it would be an interesting project to see how far you can improve with your current method!


MarcelGruenauer wrote 3 years, 10 months ago:

Yes, the whole-board strategy in high-handicap games is very different from even games or low-handicap games, but when the game progresses to local corner and side situations as well as endgame, KataGo often shows me sequences that I had not thought about.

I like to extract these local situations, expand them with AI analysis and put them in a file as a reference to broaden my repertoire and also to deepen my understanding of certain common shapes.


bellicose wrote 3 years, 9 months ago:

Hey, are you playing KataGo on a server (like OGS) or did you download and play on your own machine (using something like Sabaki)? I have a Mac, and I am not very savvy with things like homebrew and So far have not been able to figure out getting it on my system.


antti wrote 3 years, 9 months ago:

I’m running it on my home Mac computer with Lizzie. Unfortunately the installation is a bit tricky – the easiest way is to install Homebrew and then brew install KataGo, after which the newly created KataGo folder will have a readme file that explains the rest (although the information is pretty well hidden). If you get this far and feel like trying the installation out, scroll down to ‘Compiling KataGo’ in the readme and follow the steps for a Linux installation; they worked for me on a Mac, too (opt for opencl over CUDA unless you happen to have a Mac with an NVIDIA graphics card).

By the way, if you have a newer-generation iPad (or iPhone) that has an A12 chip or newer, it’s almost more efficient to just get A Master of Go from the App store and use that instead!