Computer Chess Strength

[jack512]

In 1982, Thompson estimated that a factor of two increase in computing power is worth about 100 ELO rating points (Advances in Computer Chess 3). My question is, with changes in algorithms since then, such as LMR and null move, has this correspondence changed? Can anyone point me to recent studies that measure the increase in rating points as a function of computing power?

[bob]

In 1982, Thompson estimated that a factor of two increase in computing power is worth about 100 ELO rating points (Advances in Computer Chess 3). My question is, with changes in algorithms since then, such as LMR and null move, has this correspondence changed? Can anyone point me to recent studies that measure the increase in rating points as a function of computing power?

Yes, it has changed quite a bit. It is more like 50 Elo today for an additional ply, which is often less than a 2x time doubling as well. Today's search is more selective, which means a single ply adds depth and inaccuracy, where the depth gain is good, but the inaccuracy is bad.

[jack512]

In 1982, Thompson estimated that a factor of two increase in computing power is worth about 100 ELO rating points (Advances in Computer Chess 3). My question is, with changes in algorithms since then, such as LMR and null move, has this correspondence changed? Can anyone point me to recent studies that measure the increase in rating points as a function of computing power?

Yes, it has changed quite a bit. It is more like 50 Elo today for an additional ply, which is often less than a 2x time doubling as well. Today's search is more selective, which means a single ply adds depth and inaccuracy, where the depth gain is good, but the inaccuracy is bad.

Ok, thx. So the more selective search gives you fewer Elo per extra ply, but more extra ply per doubling of computer power.

How does it end up if we just look at how many extra Elo for a doubling of computer power?

[bob]

In 1982, Thompson estimated that a factor of two increase in computing power is worth about 100 ELO rating points (Advances in Computer Chess 3). My question is, with changes in algorithms since then, such as LMR and null move, has this correspondence changed? Can anyone point me to recent studies that measure the increase in rating points as a function of computing power?

Yes, it has changed quite a bit. It is more like 50 Elo today for an additional ply, which is often less than a 2x time doubling as well. Today's search is more selective, which means a single ply adds depth and inaccuracy, where the depth gain is good, but the inaccuracy is bad.

Ok, thx. So the more selective search gives you fewer Elo per extra ply, but more extra ply per doubling of computer power.

How does it end up if we just look at how many extra Elo for a doubling of computer power?

That has actually been fairly stable. It is a bit lower than 30 years ago for sure, but it helps. Typically today, double time limit gives 50-70 Elo gain...

But obviously it is a decreasing scale.