A VERY INTERESTING...™ match between 2 galaxy brained Buckshot Roulette AIs

[@Grim-Grin]

Careful amicable smoochlad, I think this kissful delightboy guy is taking over your channel.

[@catinumpurrsona]

i cant believe this "delightful kissboy" character is trying to overtake our beloved smoochlad. he should sue them

[@AmicableSmoochlad]

This DKB character needs to get a life!!!

[@cwp24]

6:28 that’s the equivalent of stockfish finding an ABSURD move or sequence of moves in chess. I bet most humans wouldn’t have even thought about doing that.

[@audiosurfarchive]

Wayne: 🤔 Harry: 🤔

[@Fluugan]

So much DKB content dropping these days. You're my favorite new (to me) creator! I am absolutely hooked to your presentation vids and everything else.

[@AmicableSmoochlad]

Eyy, thanks so much! Really nice to hear :^)

[@scubasteven2222]

@@AmicableSmoochlad same here, the mix of statistical analysis, programming, strategy, humor, and personality are scratching my itch like nothing else. i'm so glad i found your channels recently man, thank you for your service Joe 🫡

[@scubasteven2222]

@@AmicableSmoochlad same here man, the combination of statistical analysis, coding, strategy, humor, and personality are scratching an itch i didn't even know i had. so glad i found your stuff recently, thank you for your service Joe 🫡

[@lizardizzle]

I love this kind of solved probably experiments in games! Whenever a game is said to just be random, if there are tools that the players can use at all, it becomes much more, and this really proves it. It's hard to visualize in the midst of the game without taking forever as a human player.

[@gemstone7818]

damn thats a really interesting analysis of how complex this game can be, will be nice to see a 3-4 player game analysis in the future

[@Kirn.]

2:40 An interesting play here would be to invert, shoot the opponent, and force them to shoot you so you can start the next loadout.

[@relthetraveller245]

I also think that would be the optimal play

[@AmicableSmoochlad]

Potentially correct; I've got the AI valuing adrenaline pretty highly though, so I think it's seeing the 1 missed damage + the used adrenaline as not being worth going first when at 3 health (out of handsaw range)

[@LittOnTheFifty]

Congenial Commentary Cobber with the heat 🔥🔥🔥🔥

[@Smallpriest]

That whole loadout where Wayne stole the beer was some excellent stuff, makes you realize how important starting first really is. Also at 30:18 if Harry used the jammer, would he have started the next loadout first?

[@Smallpriest]

Btw ans is yes, jammer stays active through new loadout

[@araanimations]

It’s crazy to me how the math is in the final round. I believe the scenario of Harry hitting a live from shooting at himself the 1st time works out better. He got the Worst situation of ejecting a bullet before his 25% chance of getting ahead failed. I believe the small chances of him going first regardless of taking 1 damage were why trying to hit 2 blanks in a row was the best to him.

[@AmicableSmoochlad]

Yeah, there's a lot to unpack there; really cool to see how much complexity can be in this seemingly simple game.

[@Phos9]

Burner phones with more than two players seems like a traveling postman problem

[@asw2s]

Can't believe Amicable Smoochboy got hacked! Who tf is Delightul Kissboy?!??

[@yey2101]

There was some other things in this video that I was gonna comment on but i fell asleep before writing this, something that i did remember though was that the ai would use a knife when having a 60% chance to hit instead of a guarantee or higher percentage. For example 3 lives 2 blanks, I feel like using a knife there when you don't know the shell isn't smart unless the enemy has 2 hp, but still. Either way this match was insane, loved it.

[@AmicableSmoochlad]

Yeah, the AI is usually pretty aggressive with handsaw usage, preferring to even use it in 50/50s. I think part of that is bc I have them gradually value them less and less the more they have since they tend to not stack well in multiples

[@imgoingtoquittheinternet]

why'd he pull the phone through 24:00 what if he saw the final option (assume it is red instead of blue) and then saw it again on his next phone I might be wrong though since I haven't actually played multiplayer yet

[@AmicableSmoochlad]

Looks like there was actually some kind of bug that was resulting in the AI thinking it had a guaranteed win even if it wasted a magnifying glass even though it didn't. Disabled parallelization and it made the correct move to get the guaranteed win. Without parallelization it thinks it only has a ~95% chance to win when wasting magnifying glass, which seems correct. Looking into why this would be.

[@AmicableSmoochlad]

With parallelization disabled, all the rest of the moves in the game are the same besides this one.

[@AmicableSmoochlad]

Ah, I think I know why actually. Has to do with parallelization allowing certain operations to clobber each other when updating the cache. I've got a fix for that.

[@rexrobotdude5578]

Whos this Delightful Kissboy fellow that keeps introducing himself at the beginning Wheres my lad Amicable Smoochlad…

[@marioisawesome8218]

4:30 it would be a guaranteed 2 if handsaw was still on the table because if the round is blank, wayne turns it on himself knowing there was only 1 blank left in the entire loadout, guaranteeing the next shot to be live. regardless, Harry made the best play in this situation by avoiding extra damage.

[@fragolastrawberry5920]

2:59-3:14 Is that the reason why wayne shot you at 24:00 of the video on the other channel (the video from 5/6 days ago)? Because there were two live ones (and no blanks) remaining, but instead of inverting and shooting himself, before shooting you, Wayne left the Inverter unused, even though it would've prevented him from taking 1 additional damage (so he would've gone into the next loadout with 2 lives instead of1).

[@josesteves]

Nice vídeo

[@relthetraveller245]

14:57 Why didn't Wayne invert there?

[@AmicableSmoochlad]

It's a pretty interesting situation, here's a copy/paste from my Discord: "I think the reason he doesn't invert: 2/3 of the time, it's blank and he wins if he inverts, 1/3 of the time it's live and harry starts the next loadout with adrenaline and inverter and Wayne with nothin. If he shoots, 1/3 of the time he wins. 2/3 of the time, he misses, then it goes to Harry. Harry can then invert, and 50% of the time loses or doesn't invert and loses 100% of the time. So the odds of winning are the same, but this way lets Wayne keep inverter and force Harry to use his inverter"

[@relthetraveller245]

​@@AmicableSmoochlad The strategy behind multiplayer is way more complex than what I initially thought. Thank you for your explaination, never thought about it in that way!

[@audiosurfarchive]

We have now reached a new level of DKB content: commentating the gigachad AI vs. AI matchup of a self programmed Backsho--I mean Buckshot Roulette matchup.. because the game wasn't hard enough. 🤔

[@kirbybrlivevoltamaster5223]

14:55 wayne has 66% chance of shooting a blank (giving 50% chance of losing and 50% of being the first on the next loadout) and has a 33% of winning, but why he doens't use the inverter? since in that moment, by using it his winrate would go up to 66% and the chance of insta losing would go to 33%

[@delightfulkissboy8926]

Answered this Q I in the replies to the comment before yours

[@tessenary_]

if perfect chess ends in a draw then perfect buckshot roulette ends in a win lose win or a lose win lose

[@Actimel730]

9:00 why didnt wayne first steal the beer and then the magnifying glass? cause then the live where he used the saw on the 50/50 would have been a known live for him. is it just cause magnifying glasses are higher rated than beer?

[@AmicableSmoochlad]

So I went back and checked what would happen if Wayne had seen live there instead of blank. He instead saws it off, then shoots rather than stealing a beer. So Wayne wanted to confirm what the next shell was before deciding whether or not to use beer.

[@droughtender4831]

It might be worth it to make the eval function for loadouts use a neural network trained via backpropagation, (though that may take a long time to train)

[@tortoise-chan]

I'd think regular old machine learning could do pretty well here. Just adjusting the evaluation function depending on whether the AI won or loss after reaching that position (so, reinforcement learning) would be a good starting point and should be pretty doable

[@Acedog11]

So why haven't you named this BuckStockFish or some other pun on Stockfish yet? It's right there :C

[@AmicableSmoochlad]

Lol you're right, could be like, StockfishInABarrel or something

[@tortoise-chan]

Do the AIs have any kind of endgame lookup tables or a dictionary of previously computed positions? I assume you're caching a lot of the work since minimax gets out of hand real quick otherwise. Im curious how much of this is strongly solved vs how much could potentially change based on the eval function. Either way, when i think of a replay function, i cant help but think of the game review from chess. Imagine buckshot with an eval bar (or, if we're shooting for the moon, a list of the different lines sorted by evaluation)

[@AmicableSmoochlad]

Yee, I have a dictionary that maps each game state to a value that's populated in evaluation. I don't extend evaluation past the end of the loadout bc the possibility space is too large, so on loadouts with no chance of winning or losing, the decision-making is based on the eval function quite a bit

[@gregoryderpwrld111]

I think something interesting you could do is try to program an AI that behaves like the base Dealer, but he knows the shell loadout, then pit him against Wayne An idiot who knows the loadout vs a genius who doesn’t

[@DkKobaADV]

You should use this ai to research the effectiveness of items. Id be curious how they rank

[@LittOnTheFifty]

Assuming there were a 1v1 Buckshot tournament, how should such an event be structured so that variance can be overcome by correct play?

[@LittOnTheFifty]

Could be formatted like a tennis tournament, or maybe a round robin... not sure which is best... but if there were a little prize pool (👀👀👀👀👀 such funds do exist), it'd be nice if the winner wasn't determined (entirely) by luck. Anyone have ideas about the number of games/sets/matches a player should be entitled to in order for skill edges to manifest?😊

[@Smallpriest]

Maybe best of 3 matches? 1 match is already a best of 3, so it's like best of 3 squared

[@THETV1RUS]

Are you using mini-max algorithm for the AI?

[@AmicableSmoochlad]

Pretty much; it's expectimax, so minimax but with non-deterministic games

[@audiosurfarchive]

HELL YEAH SEC9ND CHANNEL GANGGANG

[@Dendricklystable]

One small thing: I think that the game is too small, half the screen is just blank so if you could make it more compact that would be cool

[@AmicableSmoochlad]

The window is resizable, so you can have it be a square if ya want and it'll look more compact 👌 Not sure what to fill the negative space with in wide screen tbh

[@juistin872]

When choosing these seeds to review, do you just choose a random completed one, or is there some sort of metric that can clue you in to whether a particular match was interesting in some way? I know next to nothing about game coding/automation.

[@delightfulkissboy8926]

I've just been watchin a few games, and if I see a good one, I save the seed for future reference 👌

[@RedTail1-1]

I give up. I get it, you don't want to play multiplayer.

[@AmicableSmoochlad]

My condolences 😢

[@delightfulkissboy8926]

Mine too 😞

[@catcatcatcatcatcatcatcatcatca]

These AIs value starting a new loadout very highly. For humans thats when you have least information, and if the player starting the loadout fails to make a good play the second player has massive advantage. It would be interesting to see if limiting items would change this. If each player only got one item at the start of each loadout, I can’t see similar advantage for going first. Second alternative would be really really long loadouts, with 12 or more shells.

[@delightfulkissboy8926]

Yee, I think they're right tho, especially when life totals are low. Adrenaline, jammer, and cigarettes are all a lot better when you get to go first, and even though there's less info when you go first, usually there are more lives than blanks, so you'll get a pretty good chance to hit. If you or your OP are at 1 or 2 health, it's super important to go first

[@mindiblumenthal2051]

You are on Google

[@delightfulkissboy8926]

What do ya mean by that 🤔