Interesting ugly shape lesson from bots
Posted: Fri Feb 22, 2019 3:54 pm
Watching miracle97 (Byun Sangil 9p) get his usual beating from BensonDarr / PhoenixGo on Fox, one move stuck out to me, white to play. After approach at 1 white attached with 2, a move beloved by ddks that they used to get told off for and pros only played rarely, but is back in fashion with the bots, particularly with a side stone marked (and black already strong on lower side so white doesn't mind strengthening black there to solidify the corner/side) and Byun wedged, a standard resistance to white's plans. White then ends up with 2 cuts, how to defend?
The following pro game pattern search from waltheri shows the usual shape choices, most common is simple and easy solid connect at 1, or there's the tougher corner block (and then usually black cuts). There was an interesting game recently between Mateusz Surma 2p and Stanislaw Frejlak 7d where Mateusz chose the block but connect was better so they had a fun fight, see commentary at https://explorebaduk.com/2019/02/13/vad ... rejlak-7d/.
But BensonDarr picked a move never seen in waltheri, and one sure to give Otake a heart attack for being such an ugly shape:
What's the idea? The most obvious is that compared to solid connection if black pushes at q3 white can safely block at r3. So instead a cut and fight is expected (during which r10 comes in handy) but it's really complicated. For more than that explore it yorself with lizzie.
I checked with a variety of bots, and many liked it too as #1 choice:
- LeelaZero #205 solid intuition, extend #1 within 1k playouts (solid connect -3%)
- LeelaZero #157 after 40k playouts, started out liking solid (<1% worse)
- LZ #157b (15 block 157 with 40b games) after 2k.
- LZ 191 (sees it as promising blue circle quite early but doesn't become #1 playouts until 8k)
- Elf v1 and v2 both like it within a few hundred playouts
- LMGX88, the 'Leela Master' trained on human games. After 20k playouts for connect ugly extend has 30 playouts, but then it realises it's good and by 47k total extend is #1
- MiniGo v15 990 cormorant. Doesn't discover it until 4k but by 10k is #1.
So I think it is very likely the best move, but one us humans with sense of good shape were blind to.
P.S. Speaking of Mateusz I just saw him on Fox (lost to FineArt) and he beat young Japanese hopeful Onishi Ryuhei 3p a few weeks ago, nice win!
The following pro game pattern search from waltheri shows the usual shape choices, most common is simple and easy solid connect at 1, or there's the tougher corner block (and then usually black cuts). There was an interesting game recently between Mateusz Surma 2p and Stanislaw Frejlak 7d where Mateusz chose the block but connect was better so they had a fun fight, see commentary at https://explorebaduk.com/2019/02/13/vad ... rejlak-7d/.
- BD m97 pattern search.PNG (222.82 KiB) Viewed 14480 times
What's the idea? The most obvious is that compared to solid connection if black pushes at q3 white can safely block at r3. So instead a cut and fight is expected (during which r10 comes in handy) but it's really complicated. For more than that explore it yorself with lizzie.
I checked with a variety of bots, and many liked it too as #1 choice:
- LeelaZero #205 solid intuition, extend #1 within 1k playouts (solid connect -3%)
- LeelaZero #157 after 40k playouts, started out liking solid (<1% worse)
- LZ #157b (15 block 157 with 40b games) after 2k.
- LZ 191 (sees it as promising blue circle quite early but doesn't become #1 playouts until 8k)
- Elf v1 and v2 both like it within a few hundred playouts
- LMGX88, the 'Leela Master' trained on human games. After 20k playouts for connect ugly extend has 30 playouts, but then it realises it's good and by 47k total extend is #1
- MiniGo v15 990 cormorant. Doesn't discover it until 4k but by 10k is #1.
So I think it is very likely the best move, but one us humans with sense of good shape were blind to.
P.S. Speaking of Mateusz I just saw him on Fox (lost to FineArt) and he beat young Japanese hopeful Onishi Ryuhei 3p a few weeks ago, nice win!
I wonder if the bots ever think this is a good move!
Waltheri indicates that this is a human blind spot. But I believe that John Fairbairn has shown us a similar human shape play, one that was not a descent to the third line, but an extension towards the side (on the 5th line? I don't remember exactly).
stone, my thought was the hanging connection on the third line.