For those who haven't thought about it before: Mack's discussion of BAA vs ABA completely underscores what BestBot and Quackle are not doing, which is to assign weightings to the possible racks opponent might have. In this example, Mack says a human player will obviously infer that you have an S (or ?) if you play ABA because BAA is such an obvious play if you don't. Once someone codes up enhancements along those lines, that's going to knock another maybe 10% off humans' win percentage I'm afraid. Inference is kinda the only advantage you've got right now ☹️
I'm going to follow up my own comment by saying that I think that my statement of "kinda the only advantage" is a little too strong. There are at least another few more "human" advantages I can think of, some pretty common: 1. Misdirection: Look at the play where BestBot played UMM. Mack infers that Bot has an extremely bingo-prone rack and adjusts accordingly. My hypothesized rack-inference-capable bot would do that too because that's sort of the only situation where fishing for so few points makes sense. BUT, suppose you intentionally fish off the M not because it leaves you will a great rack, but because you are trying to mislead your opponent. That kind of meta-gaming is pretty rare, but I think it IS something that people like Josh (@axcertypo) and Jackson will sometimes try to exploit. 2. Word-knowledge exploits: Playing a word because you think opponent won't know it's phony. Or won't know it's good. Or won't know hooks. It's hard to see how to add to a bot's arsenal without some kind of real-world knowledge or a massive dataset, as well as meta-knowledge about your opponent. Not likely. Doing this is actually pretty common for humans. For example, if you're confident in a word and you think opponent doesn't know it, that's a good choice. Players do this all the time based on some assessment of their opponent. 3. Clock knowledge exploitation: For example, it's pretty obvious when people are working through blanks when they take a long time. Sometimes you see what looks like an innocuous play after a long think and it leaves opponent puzzled trying to infer what it means. In the case of BestBot, generally when it takes a long time, it's because there's no clearly stronger play and it needed a deeper sim. That is information that can be used. And next-level meta here would be to intentionally modify how long you take on a play to deceive your opponent. 4. Time pressure: This one's pretty self-evident. Your opponent is low on time, so intentionally don't give them more time to think, eliciting worse decisions. And related: 5. Using opponent's time: AFAIK, current bots don't think on opponent's time, so this is a pretty obvious enhancement. How it would work isn't so obvious though because it would require somehow assessing the relative values of different options and then somehow exploiting that old result based on less-complete knowledge after the new knowledge (what opponent's played -> opponent's rack inference) becomes available and the new board state. But it does provide twice as much time to the bot, so it does count as a valuable thing that humans do that bots don't today. 6. Probably more stuff I didn't think of.
KHI also creates a place for -SS or -ST bingoes (a lot goes over ID too) that would score a fair bit, that's what I thought you were talking about at first...
The bot had some good draws in the first few games of the series iirc, so it's only fair you get 4 s, the z, and both blanks this game. :) Merry Christmas Mack, love watching your stuff.
I was wondering about B3 WOFUL on the UNFOLD turn. It does give back plays to the left side of the triple, but there don't seem to be major threats there given the pool and that the bot is fishing. Is it just too much of an equity sacrifice? Keeps a good leave and doesn't really give much back.
There's also CLOUR, which is probably more concerning than ODOUR since more bingos start with a C than an O and there's also a C unseen. Also with a lead the W isn't so bad since it's a good scoring tile, so given that and the extra points of UNFOLD I think UNFOLD is a good bit better
At the point you played UNFOLD, you could have played it one square farther down, hooking the L onto OUR to make LOUR. I suppose the drawback there is that it creates a line where BestBot can start with a vowel on the TWS next to the D.
I just realized, there’s no draws so the only good outcome is winning both games in the week. Since you play 2 games a week vs bestbot you’re always going up or down with two wins or two losses. Never just by one. Does that affect your mentality at all?
Not really! I also always record 2 games at once, so what's nice is when I lose the first I still feel like I have a chance to redeem myself, and when I win I feel like worst case is a draw which isn't bad at all
It's a really busy time for me with my other job, and also a long trip, plus I just got home from 5 day tournament in Virginia... simply don't have the time or energy
For those who haven't thought about it before: Mack's discussion of BAA vs ABA completely underscores what BestBot and Quackle are not doing, which is to assign weightings to the possible racks opponent might have. In this example, Mack says a human player will obviously infer that you have an S (or ?) if you play ABA because BAA is such an obvious play if you don't. Once someone codes up enhancements along those lines, that's going to knock another maybe 10% off humans' win percentage I'm afraid. Inference is kinda the only advantage you've got right now ☹️
I'm going to follow up my own comment by saying that I think that my statement of "kinda the only advantage" is a little too strong. There are at least another few more "human" advantages I can think of, some pretty common:
1. Misdirection: Look at the play where BestBot played UMM. Mack infers that Bot has an extremely bingo-prone rack and adjusts accordingly. My hypothesized rack-inference-capable bot would do that too because that's sort of the only situation where fishing for so few points makes sense. BUT, suppose you intentionally fish off the M not because it leaves you will a great rack, but because you are trying to mislead your opponent. That kind of meta-gaming is pretty rare, but I think it IS something that people like Josh (@axcertypo) and Jackson will sometimes try to exploit.
2. Word-knowledge exploits: Playing a word because you think opponent won't know it's phony. Or won't know it's good. Or won't know hooks. It's hard to see how to add to a bot's arsenal without some kind of real-world knowledge or a massive dataset, as well as meta-knowledge about your opponent. Not likely. Doing this is actually pretty common for humans. For example, if you're confident in a word and you think opponent doesn't know it, that's a good choice. Players do this all the time based on some assessment of their opponent.
3. Clock knowledge exploitation: For example, it's pretty obvious when people are working through blanks when they take a long time. Sometimes you see what looks like an innocuous play after a long think and it leaves opponent puzzled trying to infer what it means. In the case of BestBot, generally when it takes a long time, it's because there's no clearly stronger play and it needed a deeper sim. That is information that can be used. And next-level meta here would be to intentionally modify how long you take on a play to deceive your opponent.
4. Time pressure: This one's pretty self-evident. Your opponent is low on time, so intentionally don't give them more time to think, eliciting worse decisions. And related:
5. Using opponent's time: AFAIK, current bots don't think on opponent's time, so this is a pretty obvious enhancement. How it would work isn't so obvious though because it would require somehow assessing the relative values of different options and then somehow exploiting that old result based on less-complete knowledge after the new knowledge (what opponent's played -> opponent's rack inference) becomes available and the new board state. But it does provide twice as much time to the bot, so it does count as a valuable thing that humans do that bots don't today.
6. Probably more stuff I didn't think of.
That was a very good analysis
Great analysis Mitch, that all makes a lot of sense!
Mack, could you please switch the GUI back to dark? It would be so much more viewer friendly. Especially when viewing at night.
KHI also creates a place for -SS or -ST bingoes (a lot goes over ID too) that would score a fair bit, that's what I thought you were talking about at first...
The bot had some good draws in the first few games of the series iirc, so it's only fair you get 4 s, the z, and both blanks this game. :) Merry Christmas Mack, love watching your stuff.
Thanks so much, Merry Christmas to you as well!
ABAC# apparently, that turn would be very different in collins
Merry Chartisms Mack!
Thanks, and Epitaphs Hyaloids to you!
nice to see another win. you’ve almost caught up.
i wonder does best bot continue to learn to get better throughout the tournament or is it static?
I think there are marginal improvements throughout but Cesar would have more details on this
I was wondering about B3 WOFUL on the UNFOLD turn. It does give back plays to the left side of the triple, but there don't seem to be major threats there given the pool and that the bot is fishing. Is it just too much of an equity sacrifice? Keeps a good leave and doesn't really give much back.
There's also CLOUR, which is probably more concerning than ODOUR since more bingos start with a C than an O and there's also a C unseen. Also with a lead the W isn't so bad since it's a good scoring tile, so given that and the extra points of UNFOLD I think UNFOLD is a good bit better
Suggestion: Scrabble Blitz Battle but you win if you get away with a phony
At the point you played UNFOLD, you could have played it one square farther down, hooking the L onto OUR to make LOUR. I suppose the drawback there is that it creates a line where BestBot can start with a vowel on the TWS next to the D.
Exactly, and it also gives up a lot of points
Aristotle totaliser
ARISTOTLE#
I just realized, there’s no draws so the only good outcome is winning both games in the week. Since you play 2 games a week vs bestbot you’re always going up or down with two wins or two losses. Never just by one. Does that affect your mentality at all?
Not really! I also always record 2 games at once, so what's nice is when I lose the first I still feel like I have a chance to redeem myself, and when I win I feel like worst case is a draw which isn't bad at all
Do you ever play fans on woogles csw?
I see you online sometimes and I'm always worried you're busy so I wanted to ask if I could match you sometime.
I don't play CSW unfortunately, sorry :(
EEW is bad, doesn't block any of the threats you want to block and REZ takes out your JUREL spot so ZEN may be better.
Why not play Albany New Year's? There are 117 players signed up, and some games will be broadcast live on UA-cam.
It's a really busy time for me with my other job, and also a long trip, plus I just got home from 5 day tournament in Virginia... simply don't have the time or energy
@@mackmeller It's best to fly. The flight lasts for about 3 and a half hours each way.