Week 10 - Distraction

 I'm. So. Sleepy.

This Week

I didn't end up playing a lot of games this week, but I did put a lot of work into Mahjong. I've been transferring all my notes from reading the books over into a Google Doc.

Yu_song sometimes takes screenshots of Maka's judgments then puts them into a document with what he thinks Maka is trying to say. I eventually decided, hey, I wrote the Umamusume Reference Doc. Why don't I just make a Riichi Reference Doc?

So, I've been putting a lot of time into that. It's at 60 pages right now, full of notes from my book collection and organised nicely, covering everything from tile efficiency, to how to judge if you can safely deal into someone in South 4, to discard reading that I don't even understand myself. I'll get an 8-dan player to read through it as a sanity check once it's done before publishing it.

My sleeping troubles have also resurfaced. After shrinking my sleep, I tried slowly increasing it, but that made my issues with falling asleep come back. Maybe seven hours is enough sleep for me? Maybe my thoughts being full of writing the document also impact it. I'll figure it out eventually.

And I did some reading about the OK plateau. It seems like the best way to break out of that state is targeted practice, which is mostly what I've been doing. But, I might need to push it to more extremes. I downloaded Mattari Mahjong so I can do things like tile counting and calling decisions in a condensed, high speed environment.

I also did some reading about memory during my off day, trying to see if there would be a way to remember the counts of each tile discarded each round, to know all the kabes without checking. Stu Ungar counted down a six-deck shoe. The mahjong tiles even fit onto a keyboard: 123456789 for manzu, qwertyuio for pinzu, asdfghjlkl for souzu, zxcvbnm for honors. I kinda wonder if I could have a mental keyboard that encoded the numbers of each tile I saw. People can memorize a deck of cards in 20 seconds, if someone put effort into memory, it would probably be fast enough to use in a mahjong game.

But, I don't know. Just idle curiosity at this point. I started going a little crazy on the day with no mahjong.

Next Week

I think I'll just give myself permission to keep working away on the document. I have some slight obsessive tendencies when it comes to things like that. And, it'll be useful for me, too. So, I'll just yap about AIs for a bit.

When I'm playing well, I get S grades from Maka. However, getting S+ with any consistency would probably mean playing like Maka. And I don't wanna play like Maka. To get more meaningful grades, I think I should switch to using NAGA.

Maka is...

1. + Free to use (for my game volume)
2. + Fast and easy to use, almost effortless
3. - Not the style of player I want to be
4. - Not very granular. Maka gives grades from E to S+, while others like Naga or Mortal give scores from 0-100.

If we think about Naga, it's...

1. + Strong, for sure
2. + The playstyle I want (Kagashi or Omega, at least)
3. - Costly (~20$ a month for my game volume)
4. - Cumbersome to use while playing on Mahjong Soul

Naga's not designed for Mahjong Soul, so from what I can tell, you have to use the custom analysis, which charges per round instead of game. If it's 10 NP per round, then that's only about 40~50 games a month, even with the 20$ tier. Getting the replays from Mahjong Soul to Naga would also be annoying.

You may wonder, "How much does the AI choice matter?"

Well, correct play in Mahjong doesn't have a good definition. Even strong play doesn't. First, look at the stat differences between various AIs (I don't know the stats for Maka):

All 5 Naga versions outperform the version that reached 10 dan. However, they have plenty of individuality. Here's a graph showing what percentage of decisions each Naga type agrees on, when compared to another type:

Hibakari (the lowest call rate) and Kagashi (the highest call rate) in particular are extreme outliers compared to the other three types. I think the low disagreement between Kagashi and Omega is quite interesting, considering their numerical stats are very similar.

You could play perfectly according to Hibakari, and you'd have 10-dan level strength, but only have an 84% agreement with Nishiki. What would 84% agreement with Maka give you? Can we estimate that through Mortal...?

A grade replay in Mortal: 73% match, 82 rating

S- grade replay in Mortal: 74% match, 86 rating

Different S- grade replay: 81% match, 85 rating

S grade replay in Mortal: 85% match, 92 rating

S+ grade replay in Mortal: 92% match, 89 rating

Guess not. In any case, finding the AI that's closest to your playstyle seems valuable. ないおトン similarly says in the book that, "If someone has a high match percentage, you can say their strength is similar to Naga's, but if it's low, you can't say they're weak."

Apparently, each Naga version other than Nishiki was modelled after a specific Tenhou player, but the book doesn't say who, except mentioning Omega was ないおトン. I suppose just based on stats...

High call rate Kagashi:（≧▽≦）or ASAPIN?

Low call rate Hibakari: お知らせ or CLS?

Low deal-in rate Gamma: 火時計を押せ！ or すずめクレイジー?

Well, that's all just speculation, and the book just says "Tenhou player," not necessarily "Tenhoui."

Maybe it's about time to switch over to Tenhou. Deciding to reach Master first was just an arbitrary decision. If NAGA will help me get stronger, I might as well do Tenhou first and then go back to Mahjong Soul once I'm stable.