Leading Go player Lee Sedol says that despite losing four games to one to a Google algorithm, humanity isn’t necessarily outmatched by computers when it comes to reasoning and tactics in the game.
Lee believes his defeat may have been because of other human limitations, namely emotion and fatigue that affect concentration. He says it’s conceivable a human player with sufficient physical skills could overcome this disadvantage and then outplay the computer when it comes to the actual choice of moves.
We’ve previously covered the challenge of Go for artificial intelligence programs such as Google’s AlphaGo, but the short version is that the sheer range of possible moves vastly outnumbers that in chess. That range of possibilities to cover is what challenges computers, which struggle with the parallel processing needed to weigh up the options. One of the key factors in AlphaGo’s success is automatically picking a range of options most likely to succeed (based on previous games) and only considering those, rather than being determined to always find the “perfect” option.
One question now is how Lee was able to win the fourth game in the series. One possibility is of course that Lee winning one game in five was a matter of probability given the respective abilities of the players. Another possibility is that having already lost the series outright, Lee felt less pressure and could make more rational decisions.
Lee’s own belief is that he won because he played with the white stones. Normally the choice of stones isn’t considered a major factor: black has an advantage by starting first, but this is compensated for with a points handicap that is perfectly pitched to restore balance.
It seems that AlphaGo may have found it more difficult to play as black as doing so can lend itself to a more aggressive play that’s less about reacting to the opponent. Indeed, Lee even asked to play as black in the final game in the hope of beating AlphaGo in the circumstances most favorable to the computer.
Analyzing exactly what happened in the games isn’t the easiest of tasks: not only is the relative position of the players at any given time hard to quickly judge, but AlphaGo was set to favor options which increased the likelihood of winning (by any margin) rather than chasing the highest possible final score.
However, both analysis of the games and Lee’s own comments suggest that over the course of the series, the human did a better job of adjusting to his opponent’s style. While AlphaGo is all about learning from past experience, five games simply weren’t enough for it to better “figure out” Lee’s tactics.