DDK trying a different strategy

[tonybe]

I’m currently rated somewhere around 12-13kyu on OGS. I don’t play humans very often, and when I do, it’s usually correspondence, so my rank progress is slow - but I don’t mind. My real goal is to become a stronger player, and I figure - once I put in the work - my rank will eventually reflect my abilities, so I’m not in a hurry. I spend months playing bots - learning and studying - and then play 3-4 human games per year to see how far I’ve come and how much I still have to learn.

I recently had an interesting game against a GnuGo bot, and I’m documenting this for other DDKs who have been trying to figure out how to level up and play using a more complex strategy. I’m sure there were lots of things I missed or could have done better, so if you have feedback on that I’ll do my best to learn from it.

A FEW WORDS ABOUT MY LEARNING STYLE

Early on in my Go journey, I realized that - when there is a skill difference between two opponents - the more advanced one is often playing a completely different game. They are following strategies and paying attention to elements that are often unthinkable or invisible to their opponent. Since I lack patience, I’m always racing ahead - trying to grow a new set of eyes to see those more advanced strategies that allow me to make big moves. The problem is - those big moves are just one piece of the puzzle. I still have to learn all kinds of discipline and structure to make profit from those big/risky moves after I’ve played them. Over time, my strategic thinking vs my actual tactical abilities have gotten spread out over a large skill spectrum. I find myself playing moves that are bigger than my britches, and then having a hard time making profit from them.

My main mode of studying in the past few years has been pitting a Dan-level bot (Leela, Katrain) against an SDK level bot, (GnuGo, Cosumi) and trying to distill individual exchanges into global strategies. I’ve written about this before - what I do is (whenever I have sente):

• Evaluate my strongest and weakest groups vs the weakest groups of my opponent
• Look for big move opportunities
• Find my top 3 moves based on that assessment
• Force myself to pick 1 of those 3 moves as my top choice
• Only then do I check with the Dan-level bot to see if their recommendations match my assessment.

If Leela or Katrain agree with me then my learning and confidence are reinforced. If the teacher bot picked one of my other 3 moves as their top choice, I take the time to try and understand what makes that move better. If the teacher bot finds a completely different move, I take a lot more time to figure out what new strategy this move represents, and try to see those opportunities next time.

After playing a bunch of assisted games, and feeling like I’ve absorbed some new strategy - I’ll play an unassisted game against the GnuGo bot on my phone to see if my new strategy actually works. My GnuGo is set to level 8 out of 10 - which is somewhere around 9-10kyu, and just slightly ahead of my own level.

This bot plays a fairly tight territorial game - it’s good at creating large extensions and developing those into big moyos early on. It’s merciless when I over-stretch my groups, and is good at cutting and chasing me when I make mistakes in traveling across the board or making shape. It teaches me to balance my big moves with careful / solid play (on a good day) or just shows me where I am likely to make mistakes (on a bad day).

HITTING THE DDK WALL

Recently, this method of learning has stalled, because the strategy I’m trying to learn VS the strategy that my Dan-bots are trying to teach me are so far apart. I’ve been trying to play the same kind of calm, territorial game that GnuGo plays. Instead of responding in kind, Katrain and Leela usually attack one of GnuGo’s weak groups within the first 50 moves and are able to capture completely - taking a 20-30 point lead right from the start.

When I try to play similar strategies unassisted, I always mess it up somehow. I am unable to capture those weak groups in the opening, so I end up chasing them. Then, in late mid-game, I’ll make some silly mistake in making shape or managing cut points, and get my chasing stones captured, erasing my earlier advantage.

Recently, I’ve been trying to play closer to my own skill level - trying to meet GnuGo on its own terms and playing the same type of territorial game (create large extensions, make big moyos, try to take 3 corners, do big moves in the middle, etc). The games would end up close - sometimes I lose by half a point, sometimes I win by half a point. But when I step back and look at the final result - it always seems like a typical, boring DDK game - big moyos, everything is alive, no big captures. I wanted to play a more interesting game, but I didn’t know how.

THE LIGHT BULB MOMENT

A few weeks ago, while I was on vacation and had more time on my hands, I found this old thread by AdamR about NHK commented matches with English subtitles:

Cho Chikun - commentated matches

This comment caught my eye: “The way they think about the game is inspirational. Watch all five for (money back) guaranteed five stones improvement.” I took that to heart and watched 6-7 different games - trying to pay attention to the (vast) difference between their strategic thinking and mine, and came away with some interesting ideas:

• After the first 6-7 moves of the opening, pro players will play risky/aggressive moves that deny eye-space rather than trying to create large extensions or secure territory,
• When those risky moves are chased or pincered - pro players will balance playing additional risky moves with solid moves that make shape or connect their weak groups to distant reinforcements
• Rather than settle or connect their groups quickly, pro players will create many unsettled groups and increase the complexity between them - often waiting until late mid-game or end-game to make eyes or create solid connections.

TRYING ON BIGGER SHOES

With these new ideas in mind, I tried a different approach to my GnuGo games - rather than playing safe moves to grab potential - I tried playing more risky moves to invade, pincer, or deny eye space, and then see if I could keep those risky stones alive. While playing, I realized I was more confident using this strategy because I’d gained a better understanding of when I can steal sente. Let me use this joseki variation to explain what I mean by that:

If White approaches my Q3 komoku at Q5 - this usually requires an immediate response to avoid losing that corner.

However, if White approaches that Q3 with an R5 knight’s move - I can steal sente and approach a different corner (i.e. tenuki for move 2) - knowing that even if White presses the attack at P4, I can settle the stones with this simple joseki. If White tenukis after I play M2, I can bide my time - wait until I have sente again, and then enclose that corner with S4 if there are no other corners to approach/enclose. If White secures the corner with S3, I can just play L4.

So, keeping all that in mind, please feel free to take a look at the actual game:

https://online-go.com/game/26869257

Throughout the game, I’ve annotated various turning points along the way in the comments and on the board - showing places where my old strategic eyes would have played safer moves to settle stones or grab more potential, and where I chose to play riskier moves instead to attack White’s eye-space (my old strategy moves are marked with Os)

I’ve also noted times when I made mistakes in my game, and then went back and played better moves that worked. The goal here is to try and learn from those so I avoid those mistakes in future games (mistakes marked with Xs):

Not only did I feel different the entire time I was playing this game (more confident / less worried), but when it was over I was surprised to find that I had won by 24.5 points:

To my satisfaction - the end result looked more complex and interesting than my previous games - it let me know I was on the right track. Here are my main take-aways:

• When you have a chance to steal sente, look for moves that threaten your opponent’s eye-space, or threaten to cut loose groups apart, rather than playing safe moves to extend or settle your own stones
• Make those risky moves in such a way that you can be fairly confident of keeping your own stones alive if they’re pincered or chased
• When you are pincered or chased, look for opportunities to counter-attack rather than immediately running to reinforcements
• When you have a weak group in the middle running to reinforcements - play thick moves to make good shape
• When your risky strategy is paying off in one place - don’t be distracted into making safe securing moves by a sente attack in another - continue to push your advantage with the risky strategy because it should net more profit, even if it means suffering a minor reduction elsewhere
• Don’t get distracted by moves you want to play, or moves that are easy to spot - you might be missing a much bigger move on the board
• Be patient and don’t expect your attacks to pay off immediately - you might set up a situation which creates cut points or a shortage of liberties at move 100 which might not pay off until move 200 - find the right time to make the most profit from that situation and don’t rush to complete it before it is ready
• If your weak group is attacked - see if you can steal one more sente move to attack elsewhere before you connect / secure the weak spot

One final caveat - I know that I didn’t actually “win” this game against GnuGo. Given that I made 7-8 mistakes, those would have probably cost me most or all of my advantage if I were playing against a human opponent with no take-backs and un-dos. Also, I realize that this is only my first step down the path this strategy represents. All through the game, I found that I’d often get threatened and revert to my usual / safer strategy rather than playing something risky, so there is still a lot more for me to learn.

However, as a tool for reviewing my moves and trying to come up with better strategy, I find this type of exercise very useful, as it allows me to slowly develop better perception abilities and playing discipline over time. Hopefully you will find it helpful as well.

[Gia]

@tonybe

When I try to play similar strategies unassisted, I always mess it up somehow.

I have a somewhat stupid question, that is probably obvious, but it doesn’t hurt to ask, right?

Were there any easy-to-spot differences between the groups the bot chose to attack and the ones you chose to attack? For example, closer to the center of the board or the borders, with or without ko threats, the smaller or bigger on the board, with a big or small advantage in liberties etc?

[tonybe]

So, if I understand correctly, you are asking about the groups that Leela or Katrain would successfully attack and capture.

The short answer is that - within the first 50 moves - Katrain either finds a way to cut or pincer to create a weak group, or would steal eye-space from a weak group and force it to run. A few exchanges later, it would go in for the kill, and cut that group off in such a way that it had little chance of making eyes.

Basically, a Dan playing against an SDK would find ways to play aggressively in the opening / transition to midgame and be able to grab a big lead.

I’m still unable to do this and successfully capture (well, that’s not totally true - I did manage to pull it off once in this game: https://online-go.com/game/22170326)

Though I have had some success with chasing and gaining profit through direction of play.

[Gia]

My question was mostly if you noticed any differences between the groups the bot chose and the ones you chose.

If maybe one of the reasons the bots do it more successfully is the way they assess which groups to target.

[tonybe]

Well yeah - that’s definitely part of it. The Dan-level bot is able to see all the ways a weak group can be killed, whereas I can only see a weak group that can be chased because - in my eyes - they are still fundamentally living groups. I need to grow a different set of eyes before I can figure out which are kill-able and which are non-kill-able groups.

I’d have to play through a few Katrain vs GnuGo games to find an actual example.

But yeah, my basic take-away is that I’m not yet ready for Dan-level strategy, but at least I’m starting to glimpse SDK-level strategy.

[Groin]

You may not achieve quickly victories but that’s not what is important yet, just a side effect and I do like and encourage you to continue! This new approach may guide you to complete it with new consideration on the way to achieve it, that you didn’t fully grasp yet (like direction of play and indirect attack but not only).

I am usually skeptical on using bots for studying, but everyone has his way and yours seems fruitful, good luck with that!

[bugcat]

It’s always interesting to read your long, detailed, thought-out posts that give us an insight into your approach to the game.

One thing that stood out to me is that you’re very focused on the idea that the key to being a strong player is to play “risky” moves. It seems to me that the moves you, as a DDK, perceive to be risky are not viewed that way by the player playing them, who has a better grasp of defensive technique and the extent to which the opponent is going to be able to counterattack.

From the diagram you gave, I think you’re still too focused on territory and your current idea of what constitutes a strong or safe group.

There are several things which have to be addressed about this diagram and your explanation of it.

if White approaches … I can steal sente

Your wording with steal seems to indicate that you think you’re taking an advantage by playing away. Certainly Black often does have the option of tenuki in this structure, but the press (3) is usually about as big for White as what Black has played elsewhere, if not more so, as well as sente. White also isn’t obliged to press. He can consider the taisha with (3) at O4, or some form of pincer, eg. (3) at O3 or N3.

In comparison, if you respond locally then you can take the initiative in this area. Suppose you are the one to take the key point at P4; now, if White wants to settle his group, he will need to make an extension on the side (eg. to R8 or R9) – which is not to say that this is, necessarily, urgent. Another idea is to kick with R4; now after White stands with Q5 and you fix your shape with O3, again White will need to make an extension if he wants to settle his group.

I can settle the stones with this simple joseki

It’s simple, yes, but not quite a joseki (even result). The exchange of (6) for (7) is somewhat bad for Black, since he’s strengthened White’s shape. For instance, without this exchange, the cut bQ4 Q5 - P5 P6 - O5 N5 is defended against by a ladder. With the exchange on the board, the same cut can be caught in a net with …N6. A net defence is usually better for White than a ladder, limiting his bad aji elsewhere on the board. Still, (6) – (7) is to an extent excusable because of certain aji which White would otherwise have at R3, which I can explain in more detail with a diagram if you’d like.

(8) is a move seen in certain recent professional and bot games, but afaik only in the structure that doesn’t have the (6) – (7) exchange. This is because the main purpose of (8) is eliminate the aji of White R3, which depends on the push-and-cut through Black’s high jump to N3 in that shape to capture the Q3 stones (ie. wR3 … wO3 - O2 N2 - M2 P2 - N1 Q2). If (6) – (7) is already on the board, then there is no advantage to (8), which should definitely be played on the third line.

I can bide my time - wait until I have sente again, and then enclose that corner with S4 if there are no other corners to approach/enclose. If White secures the corner with S3, I can just play L4.

S4 isn’t very big (although not exactly small either). Very often an extension, a development of a side formation, or a central move will be more worthwhile. Likewise, White S3 in the first fifty moves of the game will usually be gote – there is no way for White to seriously attack you with a follow-up move.

---

I think you need to adopt two modifications to your thinking:

1. You’ve only “stolen” sente if your opponent’s move was both small and gote, ie. a mistake. Otherwise you’ve simply “taken” it, and you shouldn’t assume that just by having done so you’ve gained an advantage.

2. The concept of a “risky” move depends inherently on whether or not the player who plays it thinks (not necessarily correctly) that he has an understanding of the position: which is to say the aggressive and defensive resources available, the defects of each relevant group and how they can be exploited, the amount of empty space in which eyeshape can be made, and so on.

[_KoBa]

“Stealing sente” sounds a bit funny for me, and i wouldnt call that “stealing” per se. What you really mean is being able to tenuki, because the move opponent played was not sente - approaching corners is big, and answering those approaches is also big, but when its just 1 stone vs 1 stone, both players have usually multiple ways to settle, regardless who plays there next.

One good question you should ask yourself is “am i screwed if my opponent plays another move here?”
If not, you can usually tenuki into something more urgent, or something bigger. (And urgent moves before big moves ^^)

If you have ever watched Haylee’s videos, you can tell how much she values sente. In fact, you can often hear the exact moment when her opponent loses the game and Haylee knowing she’s already won: “awww, is that even sente?”

ps. And grats, it sounds like you’re getting better at avoiding “puppy go”, ie. just following our opponents across the board and always answering locally, which i guess is one of the biggest mistakes we kyus tend to do.

[JethOrensin]

This was a very interesting post. I have not yet looked at the full game, but I stopped at move 45 and looked at the comments and I wanted to suggest that B14 by White would have been even more dangerous that B15 because you would instinctively want to strengthen your stone that just got attached. Then White can play B10 and now you have a group that has to run, while White casually builds a huge moyo on the bottom right.

Might be just a small piece of advice, but sometimes the indirect connections of stones, offer some pretty interesting results/possibilities/shapes

[tonybe]

@JethOrensin

I wanted to suggest that B14 by White would have been even more dangerous that B15

Oh, I agree 100% - B14 secures with the tiger’s mouth AND attaches. Always want to do 2 things if you can - absolutely!

That wasn’t what my commentary was talking about (side note: and - hey! why are you helping White anyway? Go proverb says: never interrupt your opponent when they’re making a mistake ; )

What I was trying to say was that White could have played one lousy securing move and my two stone group (B11 / C14) would be running. As soon as White omits that gote move, those stones are now alive and can attack from thickness - allowing me to be more confident in making move 49 at D8.