Showdown - rock, paper, scissors (or scissors, paper, stone to you lot) against a markov chain in a sombrero
I made this!
Newbie here taking advantage of new status - I learned about you lot because something I did before ended up in your newsletter ( www.luduxia.com/whichwayround/ ) and thought you might have some curious opinions about this one.
(, Wed 8 May 2024, 0:06, Share, Reply)
That's what she said
I'm not sure what's up with that tbh, had the same reaction after posting it!
(, Wed 8 May 2024, 1:40, Share, Reply)
Well.
I like it ashully.
and I have clicked the "I like this" button therefore letting the board know that I like it.
(
TBL I posted this on, Wed 8 May 2024, 15:51,
Share,
Reply)
Thanks!
And thanks/sorry to whoever fixed the image size mess.
To celebrate this milestone I will be adding leaderboards to the old which way round game.
(, Thu 9 May 2024, 0:58, Share, Reply)
interesting
I assumed it was doing pattern analysis against an existing dataset, and won 6 in a row (interspaced with draws) by thinking "what would the move it would predict to be the most common be?".
This is where you'll tell me it's completely random and it was just dumb luck
(
cumquat may will not be commenting further on the allegations, Thu 9 May 2024, 2:13,
Share,
Reply)
Markov chains!
It really is a set of markov chains. There are a few of them of different lengths, and they estimate what your next move would be based on your previous moves (the number of moves differs according to the length of each markov chain). The results are then combined and then it picks the one to play against you. This is why it picks up patterns in what you're doing and you can confuse it if you radically change strategy.
This was a small fad a few years ago with stuff like www.rpscontest.com/ where people write bots. This one is based off ideas I got there but aiming more to be fun than excessively good.
(, Thu 9 May 2024, 2:22, Share, Reply)
ah, so just on my moves
thought it might start with coming up with a probabilistic response based on a dataset of all players choices, until the user has enough moves to establish a pattern. So just luck in first few choices. But the challenge after that was the same. Not just to be able to see patterns in your own choices but to to predict what the algorithm is most likely to think your next choice is, and this will change with each datapoint you give it. It's constantly changing your strategy.
A truly random choice would only settle on a 50% win/ loss, but we (humans) are quite bad at acting truly randomly unaided by things like dice, such as the "37" phenomenon www.youtube.com/watch?v=d6iQrh2TK98
(
cumquat may will not be commenting further on the allegations, Thu 9 May 2024, 4:59,
Share,
Reply)
Yeah, so it forgets everything if you create a new game or it just got far too annoying. The result is it is effectively totally random for the first 2 rounds, then it steadily becomes less random until after about 8 rounds randomness is only used to tie break on internal decisions.
You remind me, a couple of years back I bought urin.le hoping to create a sort of urinal themed wordle. The idea was to show configurations of urinals and their occupations and you would get to pick where to go, then if you picked the most popular one that day you win. Could never work out how to bootstrap it to get people in though.
(, Fri 10 May 2024, 12:19, Share, Reply)
I kept hitting paper and El Sombrero responded every time with paper. After getting into that loop I found that I could keep it going indefinitely, but throw in an occasional scissors and win every time. Stopped after 7x combo and 30 points. It seems that occasional pattern changes are not sufficient to trigger a change in the model's strategy?
(, Fri 10 May 2024, 16:28, Share, Reply)
You should check the leaderboards!
30 credits and 7 combo is just starting to get interesting. During development I'd regularly get hundreds of credits and double digit combos.
Markov chains do have mechanical ways to be beaten, and ultimately it would just become an endurance exercise.
(, Fri 10 May 2024, 22:09, Share, Reply)