Random mover rank

What rating would a random mover bot have?
Here’s a game played with random moves: Random moves

1 Like

30-40 kyu?

1 Like

bot should be able to end game properly or question makes no sense.
for example:
it shouldn’t be allowed to play inside own eyes and passes when no place to play left
in Chinese rules I think it would be stronger than 25k, but weaker than 20k

2 Likes

I agree. 25k player’s moves can be quite random, or worse-than-random, as they explore what the rules actually mean.

1 Like

Any 25k would crush a random move bot, every time, no doubt. If you have a true 30k human who is not trying to lose on purpose, there is no way that he loses against a random bot, a random bot is no doubt much worse than 30k.

Basically the human “understands” the capture rule (I am assuming that we are talking about at least, a human who has just learned the rules, but never played before in their lifes, so first game ever). Even if the human still does not understand the capture rule perfectly, they can learn during the game by trying moves and checking which ones capture (I assume we have a referee or system like OGS that at least tells you if a move is legal).

This alone is a huge advantage against a random bot, that understands nothing. Something as simple as “simply choose an enemy group and start filling its liberties to capture it, one by one, completely disregard your own safety” will work marvels and crush a random bot, as it will not defend and play random stuff somewhere else. The bot may get very lucky and capture some of your stones, but you will almost always soon capture everything with very little effort.

Trying to extend the scale to such bad play though, worse than “a very bad human player that just learned the rules” does not make too much sense, as weaknesses are so great that there will probably be too strong “rock paper scissors” effects, that do not occur at higher levels. Say, badbot1 always wins against badbot2, badbot2 always wins against badbot3, and badbot3 always wins against badbot1, by simple peculiarities of their move sequences (and say, they play always the same sequence everygame, and don’t learn). How do you “rate” such bots? Is is pretty meaningless at such bad play (say maybe badbot1 always picks the top-left-most legal move, badbot2 always picks the first in a different order etc, and it just happens that by chance the triangle happens).

2 Likes

I guess we’re about to open the can of worms of “what is 25k” again.

At OGS, 25k includes total beginners who never played before.

Edit: actually, this is not quite true, since they will be [?] till they lost a few.

Nonetheless, I’ve seen 25k people building two eyes in the corner first…

1 Like

If the beginner never played before but understands the capture rule, I have no doubt at all that the beginner will totally crush a random bot.

Even if the beginner is not explained the rules very well and just told “here it is, OGS, you place stone by clicking, try to surround to capture, get more stones to win”, the beginner has a very high chance of defeating a random bot. Look at the above examples to see the things that random does…

4 Likes

I kind of want to run the experiment now, pick a random person on the street and pit them against random bot on OGS after giving them such a simple description of the game XDXDXDXD I am obviously pretty convinced that the human will almost always win :stuck_out_tongue:

1 Like

As an experiment, I showed an ongoing game (9x9) to a work colleague and with absolutely no explanation of the game, (Nothing about the rules. Nothing about the aim of the game.) I asked them to show me where the “blood is” on the board. They correctly identified the 2-liberty group that could not be saved. Go is very intuitive. :slight_smile:

4 Likes

LZ selfplay ratings started at 0 ELO = random play, iirc. Too bad its rating graph is not too informative, because it is highly inflated by illusory differences between networks.

(Since it had an 55% promotion threshold, every lucky promotion of a nearly-identical strength net still counted as X ratings points “gain”.)

1 Like

I start 6 year old novices at 42k in my youth club, which in my club means I give 8 handicap on 9x9. They somewhat understand the capture rule and they like to attach to stones, but other than that they play quite randomly. But I think they are a still a lot stronger than a truely random bot.

I’m pretty sure I can give a truely random bot much more than 8 stones handicap on a 9x9 board and maybe even those 6 year old novices can give it 9 stones handicap on a 9x9 boards, especially once they learn how to exploit it. So how weak would such a bot be? I think a lot weaker than 50k, maybe even 90k?

4 Likes

After 62000 self-play games, Leela Zero was still crushed by idiotbot on KGS: https://www.reddit.com/r/cbaduk/comments/7ekjxk/leelazero_vs_idiotbot_round_2/

I don’t think any human who understands the rules can be weaker than idiotbot.

2 Likes

I seen many real OGS 25k, they often play 1st line, self atari and other nonsense. I don’t believe they tried to read rules.

1 Like

Oh man, that’s insane, that’s painful!!! :scream:

I don’t even know why it finished! It could have gone on forever, filling their own territory, being killed and then playing again in the new empty space.
Seeing that game made a hole in my brain! :rofl:

Oh my, those first line moves in the opening! I swear I’m gonna have nightmares!

2 Likes

I disagree.

The random mover only needs to detect legal moves and randomly choose between them: just consider “pass” another legal move to choose from and use Tromp-Taylor.

No need to detect eyes or suicide, as that wouldn’t be random.

2 Likes

and “resign”

1 Like

I suspect that @icecream17’s random bot tries ten or twenty moves, and passes if none of them are legal.

I doubt it’s that easy. Since pass is NOT just another legal move, if you treat it like this you get very different strength (a suicidal bot basically) than with actual random moves.

It’s not clear which of those would be the “real” random play. Randomly choosing from board plays is not the same as certain suicide, which should mean negative strength (and not 0).

1 Like

100.0% random bot is just not interesting question. Human that knows rules at least a little can just create 2 eyes group in a corner and then pass until random bot destroys itself.

Instead it makes much more sense to think about slightly less random bot.

Then its actually possible to measure strength of such bot. If 20 kyu OGS player will have 50% winrate against it after giving 8 stones handicap, therefore almost random bot has 28 kyu rank.

what is slightly less random?