How the New AI Changed Move-3 Theory, and 10 Strategies
We rebuilt bocch's "Strongest" CPU. In 400 games against the previous Strongest, it scored 293 wins, 83 losses and 24 draws, a 76% win rate. This article covers how the new AI was made, and how theory and strategies changed when seen through its eyes.
From a scoring table to an AI that sees the whole board
The previous AI looked at a position and decided "who is better" with a scoring table. A stone that had arrived on the far row was worth 249 points, a loner was a penalty, having many legal moves was a bonus, a stone that could arrive next move was a bonus, and so on. The weights of these items were tuned on data from 26,000 games, but the items themselves were thought up by humans. Good and bad shapes that humans haven't noticed can't be written into a scoring table.
The new AI uses no scoring table. It looks at the whole board and directly outputs how likely that position is to win. It is a neural network. Nobody told it what to look at; it played hundreds of thousands of practice games against itself and learned on its own which shapes win.
Play practice games, review the positions it lost, then test itself against its previous self. If it wins, the new version takes over. We repeated this cycle 9 times, and it took over only 4 times. The stronger it gets, the harder it is to beat your previous self.
Along the way we also tried making the AI's "brain" bigger. Its judgment of positions got a little more accurate, but computing became more than twice as slow, it searched one move less deep, and it lost the match against the original AI. On the other hand, a version that was simply made faster beat the original 57% of the time. In bocch, reading deeply matters more than seeing cleverly. The current Strongest reads 15 moves ahead in 3 seconds per move.
How move-3 theory changed
We re-examined opening theory up to move 3 through the new AI's eyes. Moves 1 and 2 barely changed: the main line is still the Snuggle for Black and Shoulder to Shoulder for White. What changed was move 3, and we learned two big things.
1. Corner Guard vs. Corner Guard: Snuggle, as theory says
Black plays Corner Guard (6f→6e) and White answers with Corner Guard too (1a→1b). Theory's third move is the Snuggle on the other side, 1f→2e. The new AI chose 6e→5e, sliding the Corner Guard stone one step further in.
In 100 games each, theory's Snuggle won 80% and the AI's move 49%. Here theory was right. In the opening, building a cluster before spreading out wins more often.
2. If the opponent comes into the center, push forward
Next, Black has played Corner Guard and White has flown into the center with a Migrating Bird, 5a→3c. Theory again said Snuggle here. The new AI pushed a stone two squares forward with a Lone Spear, 5f→5d.
The result: theory's move won 57%, the Lone Spear 88%. When the opponent comes into the center, don't shore up your defenses; go forward and try to decide the game. That was the new AI's answer. Based on this, we rewrote the Strongest CPU's opening book and the move-3 table in the game's "Learn" tab.
First move + second move = a strategy
Until now, bocch has given names to single moves, like the Snuggle or the Lone Spear. But re-examining them with the new AI showed that even with the same first move, the win rate changes completely depending on the second move.
So we call the combination of your own first and second moves a "strategy", and gave names to the 10 most common ones. The name is the same for Black or White as long as the shape is the same. With a name, you can sum up the opening you just played in a word, which makes it easier to remember and to talk about with others, just like Yagura or Fourth-File Rook in shogi.
The diagrams below show Black after two moves (White's moves are left out). We had the new AI play 200 games against itself from each strategy.
Solid: strong even when the opponent answers correctly
- Shield & Spear: build a shield with the Snuggle, then push the stone beside it straight out with a Lone Spear. Even against the best reply it wins 47%, the highest of the 10. It was the main line in the old theory too.
- Escort Bird: send the stone beside the Snuggle flying diagonally with a Migrating Bird. It wins 86% against casual replies and 42% against the best reply.
Surprise: big wins against opponents who don't know them
- Twin Walls: bring in both edges. Against casual replies it wins 87%, the most of the 10. But it gets tough if the opponent answers with Shoulder to Shoulder (the Twin Walls article).
- Vanguard: go to the center first, then bring in an edge.
- Corner: Corner Guard, then bring in the edge on the other side.
- Twin Blades: combine the Snuggle with a Migrating Bird on the other side.
An oddity: the Tempo-Loss Opening
This strategy steps forward one square with Corner Guard, then steps in one more. The resulting shape is exactly the same as playing the Snuggle in one move, so it loses a whole tempo. Even so, in positions where the opponent answered with Shoulder to Shoulder it won 35%, not as bad as you might expect.
Strategies you shouldn't play
- Lost Child: break out with a Drifting Cloud, then bring a stone in. Even against casual replies it wins under 40%.
- Dragon Spear: charge two stones forward with a Lone Spear and a Frontal Breakthrough.
- Vagabond Cloud: a Migrating Bird and a Drifting Cloud that leave your stones scattered.
Dragon Spear and Vagabond Cloud win 6% or less even against casual replies; they almost never win. If you leave your edge stones where they are and fling stones into the center, the new AI will always punish it.
Summary
- In the opening, the basic idea is to firm up the area around you.
- But if the opponent comes into the center, push forward.
- Among strategies, Shield & Spear and Escort Bird are the solid ones. Twin Walls is strong, but switch to another move if the opponent answers with Shoulder to Shoulder.
When you play a strategy with your second move in the game, its name appears on the board. Give the new Strongest CPU a try.
This article is a written version of episode 24 of the YouTube series "bocch定跡研究". The win rates for Twin Walls against Shoulder to Shoulder have been replaced with the figures from a later re-measurement (episode 25).