About the evaluation function of houdini3

[Uri Blass]

I was not sure if to put it in this forum or in the general forum but I found that it is probably possible to find details about the evaluation function of houdini3 and by a legal way that is not reversed engineering

If I change the parameters of houdini to mate search in n then
it is obvious that it searches only to depth 2n with no extensions including no qsearch.

The result is that we can learn about the static evaluation function of houdini.

For example in the opening position if I ask houdini3 to search for mate in 1 then it can give the line 1.e4 e5 with 0.00 evaluation

I learn that probably houdini3 has no bonus for the side to move in the evaluation function.

If I ask it for mate in 2 it can give the line 1.d4 d5 2.Nf3 e5 with evaluation of 0.16 pawn against white(obviously does not see Nxe5 in the qsearch because of limited search).

[ZirconiumX]

[quote="Uri Blass"]I was not sure if to put it in this forum or in the general forum but I found that it is probably possible to find details about the evaluation function of houdini3 and by a legal way that is not reversed engineering

If I change the parameters of houdini to mate search in n then
it is obvious that it searches only to depth 2n with no extensions including no qsearch.
[/quote]

Naturally. If you're looking for a mate you want to find it as quickly as possible, and not have the results skewed.

[quote]
The result is that we can learn about the static evaluation function of houdini.

For example in the opening position if I ask houdini3 to search for mate in 1 then it can give the line 1.e4 e5 with 0.00 evaluation

I learn that probably houdini3 has no bonus for the side to move in the evaluation function.
[/quote]

This I have to disagree with. If Houdini had, say a +0.10 bonus for white, and a -0.10 bonus for black, then the two equal zero. This tells us nothing.

Matthew:out

[quote]
If I ask it for mate in 2 it can give the line 1.d4 d5 2.Nf3 e5 with evaluation of 0.16 pawn against white(obviously does not see Nxe5 in the qsearch because of limited search).[/quote]

[Uri Blass]

I was not sure if to put it in this forum or in the general forum but I found that it is probably possible to find details about the evaluation function of houdini3 and by a legal way that is not reversed engineering

If I change the parameters of houdini to mate search in n then
it is obvious that it searches only to depth 2n with no extensions including no qsearch.

Naturally. If you're looking for a mate you want to find it as quickly as possible, and not have the results skewed.

The result is that we can learn about the static evaluation function of houdini.

For example in the opening position if I ask houdini3 to search for mate in 1 then it can give the line 1.e4 e5 with 0.00 evaluation

I learn that probably houdini3 has no bonus for the side to move in the evaluation function.

This I have to disagree with. If Houdini had, say a +0.10 bonus for white, and a -0.10 bonus for black, then the two equal zero. This tells us nothing.

Matthew:out

If I ask it for mate in 2 it can give the line 1.d4 d5 2.Nf3 e5 with evaluation of 0.16 pawn against white(obviously does not see Nxe5 in the qsearch because of limited search).

If I see 0.00 for symmetric position then I learn that there is no bonus for the side to move.

If I see +0.1 for the position after 1.e4 e5 then I learn that there is a bonus of +0.1 for the side to move.

In the relevant case I saw 0.00 that is not a score that I see in every search so I learn that there is no bonus for the side to move.

[bob]

I was not sure if to put it in this forum or in the general forum but I found that it is probably possible to find details about the evaluation function of houdini3 and by a legal way that is not reversed engineering

If I change the parameters of houdini to mate search in n then
it is obvious that it searches only to depth 2n with no extensions including no qsearch.

The result is that we can learn about the static evaluation function of houdini.

For example in the opening position if I ask houdini3 to search for mate in 1 then it can give the line 1.e4 e5 with 0.00 evaluation

I learn that probably houdini3 has no bonus for the side to move in the evaluation function.

If I ask it for mate in 2 it can give the line 1.d4 d5 2.Nf3 e5 with evaluation of 0.16 pawn against white(obviously does not see Nxe5 in the qsearch because of limited search).

Sorry, but that is STILL reverse-engineering. You can reverse-engineer a black box, like an Intel CPU, without taking it apart. Just feed it every possible instruction, and see what is left where (registers, memory, etc) after each one. You eventually produce the complete set of opcodes and their descriptions.