EBF

[lauriet]

During my function MakeMove I increment "Nodes" or 'QNodes" depending on which function called makemove.
After each ply of search in my Iterative Deepening search I store the nodes in an array index by ply.
:TotalNodes[Ply] := Nodes + Qnodes.

At the end of the search I calc for each ply EBF[Ply] := TotalNodes[Ply] - TotalNodes[Ply - 1] / TotalNodes[Ply-1]

I then average the last 3 (plys} EBF.

Does this sound right ?
I only ask because the result I get seems to be too good to be true.

[Michael Sherwin]

During my function MakeMove I increment "Nodes" or 'QNodes" depending on which function called makemove.
After each ply of search in my Iterative Deepening search I store the nodes in an array index by ply.
:TotalNodes[Ply] := Nodes + Qnodes.

At the end of the search I calc for each ply EBF[Ply] := TotalNodes[Ply] - TotalNodes[Ply - 1] / TotalNodes[Ply-1]

I then average the last 3 (plys} EBF.

Does this sound right ?
I only ask because the result I get seems to be too good to be true.

"TotalNodes[Ply] - TotalNodes[Ply - 1] / TotalNodes[Ply-1]"
In every language that I know of division takes precedence over subtraction. Therefore the quoted code reduces to:
TotalNodes[Ply] - 1;
I can't imagine that was your intent.

[lauriet]

Yep, my mistake, I actually have the brackets

(Nodes - Nodes[i-1]) / Nodes[i-1]

"My Bad"

[lauriet]

OK, I think my issue is that the last ply does not get completed due to timing out.
If i ignore the last ply nodes and average the previous 3 ply nodes it seems to give believable values.

[Ras]

I then average the last 3 (plys} EBF.

You should always average over an even number of plies to take care of the odd-even-effect: https://www.chessprogramming.org/Odd-Ev ... _Deepening

[lauriet]

Hi again,

So what is the worst EBF ?
What is the best EBF ? (stockfish?)
What is a 'pretty good' EBF ?

In fact does EBF correlate with ELO rating at all ? I would imagine that it has some connection (all else being equal)

[Dann Corbit]

Worst EBF is pure mini-max. In the mid game, each new ply will take about 36 times as long as the previous ply on average

If you have perfectly implemented alpha-beta, it would reduce to something near 6. Of course, that assumes perfect move ordering.

Null move and LMR are also big wins, but the reduction will be not only a function of the quality of your implementation but everything about your program (how good is the eval, how good is the move ordering, how good is the hash table, how good is the ...)

pvs search also reduces branching factor, though I don't think it is called a pruning technique.

Razoring can be a big boost (the big speedup for Rybka came from brutal razoring)

Every pruning technique reduces branching factor (on average or it would not be pruning).

Stockfish is somewhere around 1.5 I think, though I have not measured it lately,

[Dann Corbit]

You cannot directly connect Elo to branching factor.
That is because it only works fabulously if it is implemented fabulously.
There was a version of ExChess which (for quite a while) had the smallest branching factor of any program.
But ExChess (while strong) was not a top ten program.
It is not just how much you prune, but what you prune.
The concept of pruning is to speed up the search by spending less time on the bad moves.
So what happens if you are pruning the good moves along with the bad moves (or worse yet, pruning the good moves and examining the bad moves)?
Obviously, the Elo would go down instead of up

"The Gambler" by Kenny Rogers sums it up nicely:
"You got to know when to hold 'em, know when to fold 'em, know what to throw away, know what to keep."

[lauriet]

Hey Dann,

Now Im worried, if stockfish is about 1.5.
My program would be way weaker but my EBF in the deeper plies can be anything from < 1 to
about 2 or 3.
Typically the early plies have an EBF of 5 or 6 and then the later plies reduce from that.
If you look at the earlier posts you can see how I calc it.
Ive change it to average the previous 4 plies starting from MaxPly - 1 (since the last ply is not completed typically)

Maybe, as you suggested, I am killing off good moves instead of bad moves......
Now if you could just tell me how to tell the difference I will be 'on my way'

[Dann Corbit]

I should mention about ExChess - Dan Homan is a great innovator. He did not invent Lazy SMP but he is the one that got it going. He explained what he did and he showed how well it worked. We owe him a debt of gratitude for that.

Don't expect 4000 Elo when your program is young. In fact, if it starts at SF level, we will all look at it with a stank-eye and wonder if something sneaky is going on.

Know this, it takes a very good programmer to write a simple chess program that works, So that is something to be proud of.

As you progress you can take great enjoyment in every step, and those who already walked in that path will smile with you. But beware, chess programming is a strange passion that can put you into a closet and make you program around the clock. But with Corona 19, maybe now is the perfect time to be a chess programmer. sure beats staring at the wall for 18 hours.