Best move never changes

[Henk]

Best move at highest levels of the search almost never changes in my engine. So it wastes most of it's time. For if it would have stopped earlier the same best move would have been found. Looks like something wrong. I think it's because it reduces (applies LMR) after first move (best move of previous iterations). I don't know.

[vittyvirus]

Best move at highest levels of the search almost never changes in my engine. So it wastes most of it's time. For if it would have stopped earlier the same best move would have been found. Looks like something wrong. I think it's because it reduces (applies LMR) after first move (best move of previous iterations). I don't know.

Where's the code?

[bob]

Best move at highest levels of the search almost never changes in my engine. So it wastes most of it's time. For if it would have stopped earlier the same best move would have been found. Looks like something wrong. I think it's because it reduces (applies LMR) after first move (best move of previous iterations). I don't know.

Where's the code?

The last experiment several of us did, independently, starting with "Crafty goes Deep" by Monty Newborn and myself, and then "Dark Thought goes deep" by Earnst Heinz, found that our programs changed their mind on the order of 15%-18% of the time, each time we go a ply deeper. Obviously in some forced cases you never change your mind, so we did not include those in the test, we just took a set of 360 positions or so from the DB vs Kasparov matches and used those.

Never changing your mind certainly seems wrong as it suggests that additional depth is not revealing anything new to you.

[ZirconiumX]

Best move at highest levels of the search almost never changes in my engine. So it wastes most of it's time. For if it would have stopped earlier the same best move would have been found. Looks like something wrong. I think it's because it reduces (applies LMR) after first move (best move of previous iterations). I don't know.

Highlighted for clarity.

Do you do any tactical search extensions?

A possible improvement you could make would be "easy move", where if the best move has never changed (for the last 5 plies maybe) and it's 50cp or so better than the next move, then play it. Can be useful in situations where your engine is too blind to see the best move - play what we think is best, then we get more time to think.

Matthew:out

[Henk]

I think it reduces too much when applying LMR after first move, so it gets severe ply inflation.

[Of course a good move is a good move so it should not change.]

[ZirconiumX]

I think it reduces too much when applying LMR after first move, so it gets severe ply inflation.

[Of course a good move is a good move so it should not change.]

But alpha beta doesn't look for a good move, it looks for the best move. So something is quite clearly wrong.

Matthew:out