[ThrustVectoring]

>But by definition, better players take better decisions Ev average.

Note that this is only true when expected value lines up exactly with "increases probability of winning". Lots of value calculations don't, and so better play can look strange in some situations.

Back when I was a middling chess player in High School, I'd pretty much always play King's Gambit as white when it was offered, especially against stronger players. I knew that I didn't study openings as much as other players, so I wound up in a place where I at least had as good opening book knowledge. Plus the opening is high variance, so it gave better opponents more chances to screw up and let me in the game. That said, the consensus high-level opinion on King's Gambit is that white does not get sufficient compensation for the pawn.

Or perhaps and even better example is AlphaGo Zero. When this engine is winning a Go game as it gets into endgame, it will play lines that confuse the professionals as locally sub-optimal. The pros later figured out that these moves would give up points in exchange for removing lines of play that are threatening and uncertain: instead of winning by 2.5 points with a small opportunity for the opponent to come back, AlphaGo Zero would win by 1.5 or 0.5 points without yielding any counter-play.

[rhlala]

Make a lot of sense, specially vs humans, if ev of two lines are similar, better take the one opponent is likely to have less studied.

In the case of alphago, all depends how ev is calculed, % winning or points.