•8 min read
The learning rate was the real lever, not the network's size
We grew the YINSH game AI to a larger brain and discovered one hardcoded number had been silently defeating the idea before it got started.
#yinsh
9 posts tagged with yinsh. This index aggregates both frontmatter tags: entries and inline #yinsh mentions.
•10 min read
270 generations of self-play: the clever tricks lost to raw capacity
270 generations of self-play on one GPU. A bigger network and a stricter referee won; the cheap clever tricks lost; the next lever is a hardware upgrade, not another idea.
•7 min read
The speedup that sped up nothing — when a flat line lies
After I made the YINSH AI's practice cheap, the referee became the slow part. I diagnosed the bottleneck, built a careful fix, proved it correct — and it made zero difference. This is the story of why a flat performance curve fooled me, and how to ship an honest negative result.
•7 min read
Winning sooner, not just winning — teaching the AI to value a quick kill
The YINSH AI thought a win one move away and a win a hundred moves away were exactly the same thing. Fixing that — teaching it to prefer winning sooner — made it both stronger and cheaper to run, and taught me the value network is a search multiplier, not just a scorekeeper.
•10 min read
Thinking smarter, not harder — a faster self-play loop that can't get less creative
The YINSH AI spends two-thirds of its training time practicing against itself. Three planned changes cut that time — the biggest replaces 800 brute-force look-aheads per move with about 50 clever ones — under one hard rule: it must not get less creative.
•8 min read
The AI took two moves when one would win — a YINSH value saturation story
At score 2:2, the staging YINSH bot had a guaranteed one-move win but chose a slower path — the diagnosis confirms value saturation, and a no-retrain MCTS-Solver fix cures it.
•7 min read
Two dead-end runs proved the YINSH AI was too small, not undertrained
Three 12-hour GPU runs to beat the YINSH champion gen-077 — two dead-ends on more data and momentum, then a network twice the size won in two generations from scratch.
•7 min read
Same wall, new fix — the AI broke its plateau by thinking harder, not getting bigger
A bigger network beat the YINSH champion, then stalled. The next breakthrough wasn't size — it was making the AI search deeper before it learned — plus a performance trap that almost hid the win.
•6 min read
The AI opponents players couldn't see, and the one line that hid them
Six trained AI opponents were live, but the game menu showed only three. A missing line in a startup script hid the rest, until the menu became config-driven.