DeepMind published three new videos with Ke Jie’s games reviewed by AlphaGo itself!
Since I was wondering which moves were discussed most frequently, I counted select words in the subtitles1.
Even though AlphaGo itself can’t talk, its thoughts were voiced by two very capable humans. So the number of times certain terms were uttered correlates approximately with how often AlphaGo considers those moves.
Of course, the sample is small and further research is necessary to yield more accurate results, but we can draw a preliminary conclusion: AlphaGo likes hane.
| Item | Game 1 | Game 2 | Game 3 | Total |
|---|---|---|---|---|
| hane | 38 | 31 | 22 | 91 |
| jump | 28 | 36 | 25 | 89 |
| extend/extension | 61 | 15 | 10 | 86 |
| keima | 22 | 35 | 22 | 79 |
| attach/-ment | 24 | 32 | 14 | 70 |
| invade/invasion | 38 | 22 | 10 | 70 |
| take* | 26 | 33 | 11 | 70* |
| peep | 36 | 28 | 5 | 69 |
| cut | 24 | 29 | 11 | 64 |
| atari | 13 | 44 | 4 | 61 |
| push* | 20 | 25 | 13 | 58* |
| nobi | 21 | 23 | 7 | 51 |
| 3-3 | 22 | 12 | 12 | 46 |
| block | 18 | 11 | 14 | 43 |
| pincer | 22 | 12 | 0 | 34 |
| descend | 7 | 7 | 6 | 20 |
| bump | 3 | 1 | 14 | 18 |
| kosumi | 2 | 14 | 2 | 18 |
| tiger/-mouth | 3 | 12 | 3 | 18 |
| ladder | 9 | 1 | 0 | 10 |
| retreat* | 5 | 0 | 4 | 9* |
*inflated due to usage in regular language.
(English verbs include all tenses and conjugations.)
From now on, I’m going to consider hane more often in my games, until new research brings a different conclusion.
It seems that AlphaGo also slightly prefers White over Black.
| Colour | Game 1 | Game 2 | Game 3 | Total |
|---|---|---|---|---|
| White | 141 | 78 | 34 | 253 |
| Black | 97 | 78 | 52 | 227 |
1Translated by yours truly.
