We know, and some of us - not me - even understand why, that bot reliability with no-komi and handicap games is subject to a good deal of interpretation. So far as I can make out, the consensus seems to be that a bot trained on 7.5 komi can offer pretty good assessments early in the game, but beyond that it all gets rather woolly in that the bot would probably still win but would not always be picking the best move and could even occasionally lose.

But that is presumably because the balance is being tilted heavily in Black's favour.

If a change is made that affects both sides equally, is reliability impacted as much?

I am thinking specifically of group tax in old Chinese games. It's a sizeable and fairly important factor which we know shows up even in josekis. Keeping your stones connected, even very early on, was considered to be worth a lot. But does this effect attenuate in the middle game (only to resurface late in the game)? And whether it attenuates or not, would the overall effect be less than the absence of komi?

On a separate tack, it has been assumed that a bot search with more playouts produces more reliable moves. But that presumably is dependent on deep searches that are too hard for humans. For humans, is the initial array of suggested moves not a better guide, as they are presumably compiled using factors that are more amenable to human understanding? And would those instantly suggested moves not be better (in the sense of general strength improvement) in most cases for humans to ape and, if so, especially in no-komi or other tilted games?

Is there anything in the Chinese literature about group tax and bots?

The problem with no-komi games is that a bot trained on 7.5 komi will try to win on 7.5 komi, which does not necessarily mean winning on 5.5 komi for example, let alone 0 komi.

But since the bot's knee jerk reaction (the policy, the list of candidates) depends on komi much less than the position evaluation (the value, ie. who is winning), that output can be still somewhat reliable.



Right, the longer the search, the greater the impact of position evaluation is (which is incorrect if komi or sth else is different). The bot slowly shifts its attention from moves initially thought to be interesting to moves whose average evaluation (again incorrect) is highest. OTOH the problem with policy and shallow searches is that bot level drops heavily without deep reading, much more than for a human pro. It can still be useful for kyus or low dans, but this limits the usefulness from another direction.

Also what bots need lots of time/playouts to read is not the same as humans. A bot may take 1 minute to realise a ladder is not working that a 10 kyu knows in 1 second, but plays better fighting shapes in 1 second that a 9p does in 1 minute.