search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.61k stars 569 forks source link

Is there a katago function that can do dead-end questions？ #841

Open awsjgy opened 1 year ago

awsjgy commented 1 year ago

Although katago is very strong in chess, it is too strong to be of much help to humans. However, some extensions to katago might be better to help us improve our game. For example, if we could implement chatgpt to interact with katago, it could tell us why we should prioritize this move through some natural language. More basic information on this

Can katago do dead or alive problems, but he thinks about it based on the big picture. Although I know there are some ai by @kaorahi Lizgoban that can solve dead or alive problems, but it's not perfect yet. For example if some questions are positively solved for mugging and killing, then ai may consider the global maximum value of the place and not choose to calculate on top of this dead-live question. Also although there is a setup for looting, but if the number of loot is not enough, it will not consider the loot kill. Can you set the number of robocalls to be greater than the number of robocalls in the dead-or-alive question itself, and also if the value of the dead-or-alive question itself is set to be greater than the value of the robocalls

Can you prioritize the net kills so that you don't have to consider robbing the kills, and the net lives so that you don't have to consider robbing the lives?

When encountering a dead or alive problem, but there are identical pieces around, if you consider only choosing to eat a certain piece, or only making a certain piece alive.

HackYardo commented 1 year ago

Looks like ur code age is older than ur AI age. AI is not general purpose, it's niche yet: chatgpt is for chatting, katago is for playing Go. Katago can't say why (s)he play stones here or there because (s)he can't chat, chatgpt can't do it too because (s)he can't play Go... If u'd like to learn about how soon the perfect AI you want come ture, see: https://arxiv.org/pdf/1705.08807.pdf

HackYardo commented 1 year ago

As for Life&Death questions, LizGoban is using the API of KataGo, LizGoban is not a AI, it is a GUI APP with KataGo built-in. There're still many api of KataGo not be used actually. At last, L&D questions is unsolved, see #261

awsjgy commented 1 year ago

I know chatgpt can't play go directly or communicate directly with humans. But I've used the chatgpt plus version, which has a lot of plugins and extensions, and there are quite a few developers who use the chatgpt api to enable interaction with software of various functions. It's not complete, but it's dawning.

I read your https://github.com/lightvector/KataGo/pull/261, this kinfkong is our Chinese, he mentioned the function of understanding dead or alive problems, selected area to analyze dead or alive problems, already implemented on top of lizzieyzy go software on github, but not smart enough to Consider dead and alive problems that include robbing aspects, and dead and alive problems that have already been selected for a given piece.