Scrabble GM vs. AI -- the Rematch! Game #42

[@-Mopo-]

I don't see the problem with JOWL. The top two static equity plays are (JOW)L or (JOW)L+LOC. There's no difference in what they give back, and they only differ by 0.1. So we don't need much more explanation than it essentially made the top equity play. The difference of 0.1 could be almost anything. It could be, for example, that it's got a higher bingo percentage if you open anything, or maybe keeping the C scores more next turn if it plays column C to (R)AI. It's so in the noise that who cares. It certainly isn't as unbelievable as everyone seems to think.

[@wfchannel4673]

yeah dropping just the l creates an 8-letter bingo lane starting with l from 9i, while loc shuts it down. sacrificing 0.1 points of equity to create a new bingo lane with a rack that doesn't score well makes sense to me. the croce bingo lanes are way easier to block than column i.

[@joeedley3936]

Thanks Mack. Excellent video and analysis! I had all the time in the world to find the eh setup after his tiz, but i didn't look hard enuf for it until AFTER he played ouphes. So i lucked out.

[@JDHinten]

Forget about JOWL for a moment - there's so much more confusing stuff here. Why can't I find AMIRITE in my Field Book of Practical Mineralogy? And just how many sides does a MARTAGON have anyway?!

[@justin_sr]

Hmm, a MARTAGON has the same number of sides as a TARRAGON, amirite?

[@coolnath99]

great game catching up

[@iwersonsch5131]

I'd like a gamemode where players see a full list of their legal moves, but they don't see whether their opponent's last move was legal so they still have to know the words if they want to challenge a phony off the board. and I'd like to see you go up against BestBot in that mode, I wonder how much of BestBot's advantage is in strategy and how much is in just seeing every legal play

[@rotervogel1523]

You would have to play against a human though. BestBot can't phony and it knows all the words, so you can't sneek a phony by

[@mackmeller]

That would be cool -- I know strong human players have tried this against Quackle championship player and are generally able to win more over time, since even sims have limitations that humans don't have. Once you take out the ability to never miss anything computers really lose the majority of their advantage.

[@ohtani2024]

There are many terrible plays made by BestBot. Glad you found another one and I believe you still have a shot to come back as long as it does not keep outdrawing all the time! You definitely have an edge except the endgame!

[@choco1101]

Maybe BestBot has a sense of humor. It saw no good plays and just drops and L to signal resignation.

[@Oblivion776]

Being allergic to mangos is tragic D:

[@mackmeller]

It's a shame since it was a later-onset allergy and I did used to like them, but like anything, you adapt :)

[@bro91wn]

I'm baffled by (JOW)L, too. You found everything that would play through the I, there's nothing on the left or with a Y draw to hook DEXY, and the only thing through the Ns is RECON(N)I(N)G, which fails because of GE*. Glad you could get the win! At the end, I would have probably gotten scared and played N(I)B for 11, but that gives another chance to fish for nothing.

[@AlexDings]

One thing that slows the bot down is when both blanks are still unseen. Obviously becomes more pronounced at later stages of the game. Also I hate English pronounciation, I've never heard CORDIAL spoken and would have rhymed it with DENIAL 😅

[@XANDOAndy]

As a native English speaker, I also do that all the time lmao I started learning some German and Spanish last year, and I have really enjoyed that when you hear a word, you can very likely figure out how it's spelled (at least in my experience so far)

[@kinderwaff]

Mango Meller allergic to Mack

[@rhysgriffiths9675]

I'm curious, does the bot always make the best equity plays, or does it makes the plays that give it the best chance of winning? For example, a human might make a risky, low equity play if it gives them a 5% chance of winning, rather than the highest equity play where they score more points but have 0% chance to win. Perhaps that's what's happening with the JOWL play - the bot is making the best equity play, rather than higher variance moves that open the board up in an attempt to make a comeback.

[@Goryus]

Bestbot simulates a bunch of full games for a variety of possible moves, and then picks the one with the highest win rate among the simulations.

[@AlexDings]

If it made the top equity play it wouldn't need to think - that's basically instantaneous. The top equity bot is HastyBot, which Mack beat in his earlier series.

[@mackmeller]

Yeah it does simulate outcomes to figure out which plays give it the best chance of winning, but even the sims have limitations, as they generally assume my next play is going to be the top equity play. But as a human who strategizes, I'm often not going to make the top equity play, as (especially as here when I have a lead) I'll want to prioritize keeping the board tight over maximizing equity. So basically BestBot is erroneously assuming I'm going to open the board more than I actually will with JOWL.

[@rhysgriffiths9675]

@@mackmeller Ahhh cool! That makes sense. Thanks for the explanation :D

[@DU1994]

Yeah, that's the disadvantage of simming. BestBot assumes you'll make your top equity move and open some space for its bingo. And in the game, you still made top equity move and didn't even open anything

[@mackmeller]

Yep, exactly

[@wfchannel4673]

i think you're misreading the jowl play as a fish. it's actually probably their best opportunity to open a bingo lane that's difficult to block on a closed board while their rack doesn't have many scoring opportunities, especially when there are already multiple s spots available so keeping jows available is of little value.

[@quinnbartlett7233]

But it doesn't open another lane?

[@wfchannel4673]

@@quinnbartlett7233 you can argue column i is already open because of the jow hooks, but there are already several s hooks on the board so this creates a lane for plays starting with L which would be LONICERA if bot draws one of the 5 remaining a's or LICENSOR if bot draws one of the 3 remaining s's or either if it draws one of the 2 remaining blanks. so 10 out of 46 tiles give bot a hard to block bingo.

[@joshuadorsam4619]

@@wfchannel4673 there are only 6 squares under 9i? sevens starting with L are only valuable there and JOWL gets rid of that

[@paulthompson1466]

@@wfchannel4673 There isn't space there for an 8 though.

[@wfchannel4673]

@@paulthompson1466 i totally miscounted the placement. you're right. so it's just rack leave i guess. bingo-y tiles, no where to play them so you just play off one in a play that is second best equity wise.

[@rachelblaquiere9134]

I presume JOWL was under the foolish assumption you would open a line. I'm just as confused

[@mackmeller]

Yep, exactly

[@whitesoxMLB]

Didn't have Mack pegged for a bigot. How can you be so intolerant of mangoes?

[@arcieplays9040]

Macondo Analysis time! H7 HAW sims 1.0% worse than H6 WHA, a slight mistake. QI and UPDO win their respective sims by over a percent. 11E AE narrowly sims 0.27% behind J5 AI. AIRTIME is correct. WIFE and DEX are correct. 7C GIT, 6C GAT and TAG and K5 GETTERS sim extremely close, After 500000 Iterations GIT 19.05% is ahead of GETTERS 19.04%, GAT 18.99%, TAG 18.94% BestBot's JOWL sims 4th?! After 230,000 Iterations .15% behind I9 LO, .4% behind LOC and .65% behind 8A CRORE CoRDIAL is best, And all other moves made are either best, or marginally behind the best move.

[@mackmeller]

Awesome, thanks as always!