AI enthusiast and amateur player here: Michael Redmond made a great point yesterday, if the algorithm is only interested in maximizing probability of win and ignoring margin of victory, shouldn't there be some override for weak moves played when the lead is sufficient? AlphaGo played some weak moves when it perceived it was sufficiently ahead yesterday in the end game. A truly intelligent opponent will play strong moves even when sufficiently ahead, no?
I think the idea is that AlphaGo decided that the "strong" moves, while it may have increased its lead, would have been more risky than moves that just fortified its current lead.
I think the point is the seemingly stronger move (giving a bigger margin) can have follow ups that can lead to a lower chance of winning for the AI evaluation
