Werf et al. have solved 5x5 Go with minimax search (Alpha-Beta pruning). The best opening for black is C3, with which it can easily win the entire board. For the more interesting openings, for example, C2, the best play sequence is shown to the right (use the left/right arrow buttons or keys to step through the game). Black wins by 3 points.
I am working on an AlphaZero style bot. The question is: can it find this optimal sequence by itself. Just for background, even for 5x5 board, minimax search took about a day to finish, with all the optimizations they can put in. It is hopeless for larger board sizes.
Starting from scratch, my AlphaZero bot has quickly learned that white should take C3 as the second move. As training progresses, it started exploring the "cut" pattern, as seen in the optimal sequence (C2-C3-D3-D2). At one point, it even estimated that the final score of this pattern is between 2~3 points.
However, in the next iteration, the bot seemed to have lost interest in this pattern. It explored alternatives at move #4, for example, C4 (shown right), which leads to the final estimated score around -1. Clearly this outcome is more favorable to white, so it's reasonable that it would follow this path instead of playing D2 in move #4. Again, you can follow the self-played game by clicking the left/right arrow.
Why did the bot deviate from the optimal play? What went wrong? This shows the complexity of Go, even on a 5x5 board. To the right we show the likely optimal move #5 for black (E2 is the key, rather than D4). Now black is secure. White needs to worry about its own survival. Since it cannot play aggressive, the result is at least B+3 (Black winning by 3 points).
This little exercise shows the difficulty in AlphaGo training. The bot has no knowledge, at least not at this stage, of life and death. Many of its board valuations might be off, some completely wrong. With more training, it might find the optimal play.
Just for fun, we show here an interesting game the bot found during self-play. White survives without a single eye; two black blocks survive with one eye each. So we cannot just code up the conventional wisdom either. Self-play ended there with the result of B+1, although the actual margin is a bit larger if black captures the two isolated white stones.
This game feels a bit alien-like. But the principle is widely applicable, just that it is more relevant for 5x5 board than for bigger boards.