Stockfish 1.5.1

[George Tsavdaris]

Or do you plan do make it commercial one time in the future?

I just realized I forgot to reply to this question. There are no plans of a commercial version of Stockfish.

Perhaps this is a bad idea for you and for computer Chess.

For you because i guess if you make it commercial the number of buyers would be more than zero and for computer chess because if there is such a strong program hanging around then authors of commercial programs may feel intimidated about improving their program and even become disappointed, as also buyers of Chess programs will not prefer the commercial with the similar strength with the free one.

[Tord Romstad]

Thanks. My interest was if this was a genetic optimization/simulated annealing/etc method or a direct calculation kind of system.

In fact, it's more low-tech than either, but it still works quite well.

Due to the nature of the parameters (linear and quadratic) a direct calculation seemed likely, but I guess it also helps to keep the degrees of freedom low.

I'm not sure what to make of the contradicting statements whether this helps or not, but I guess it's easy enough to disable it and see the effect.

I'm not sure what the words "this" and "it" in the above sentence refer to, but in case it is the material balance evaluation: There are no contradicting statements. We believe that it is an improvement, but it's not as big as we hoped in advance. We also think that the current parameter settings, while better than the hand-tuned values we started with, are still far from optimal.

[Gian-Carlo Pascutto]

I'm not sure what the words "this" and "it" in the above sentence refer to, but in case it is the material balance evaluation: There are no contradicting statements. We believe that it is an improvement, but it's not as big as we hoped in advance.

Ok, I was confused because of this:

Material coefficents are one of the main improvments of SF 1.5

vs

In fact, the material coefficient tuning was largely a failure.

One man's failure is another man's success

[Tord Romstad]

Can't you just rip out the table? You probably don't need to bother with the difficult optimization step if you already got the results.

That would be an interesting experiment, but I don't think copying the table would work well in most programs. Optimal material balance scores depend too much on the rest of the evaluation function, especially mobility and piece square tables. Parameter settings which are good in program X usually wouldn't work well in program Y, unless the two programs have very similar evaluation functions.

[Tord Romstad]

Or do you plan do make it commercial one time in the future?

I just realized I forgot to reply to this question. There are no plans of a commercial version of Stockfish.

Perhaps this is a bad idea for you and for computer Chess.

For you because i guess if you make it commercial the number of buyers would be more than zero

That's one of several reasons why it should remain free. Buyers mean support, responsibility, web site design, and having to deal with things like taxes. The only form of compensation I would get is some money, which is irrelevant, because I have a job. Selling a chess program would lower my quality of life without giving me anything in return.

Of course I am only speaking for myself above.

and for computer chess because if there is such a strong program hanging around then authors of commercial programs may feel intimidated about improving their program and even become disappointed, as also buyers of Chess programs will not prefer the commercial with the similar strength with the free one.

Because Stockfish is free, commercial programmers can study it, reimplement the ideas they like in their own programs, and stay ahead.

It is still possible that it will some day be impossible to make significant amounts of money from computer chess, of course. If and when this happens, it is only because chess programming has become sufficiently easy and well understood that hobbyists like us can compete on the same level as professionals. This is a very natural development, and is happening in many fields. Trying to slow down progress in order to prevent it from happening is just silly.

[Dann Corbit]

Or do you plan do make it commercial one time in the future?

I just realized I forgot to reply to this question. There are no plans of a commercial version of Stockfish.

Perhaps this is a bad idea for you and for computer Chess.

For you because i guess if you make it commercial the number of buyers would be more than zero

That's one of several reasons why it should remain free. Buyers mean support, responsibility, web site design, and having to deal with things like taxes. The only form of compensation I would get is some money, which is irrelevant, because I have a job. Selling a chess program would lower my quality of life without giving me anything in return.

Of course I am only speaking for myself above.

and for computer chess because if there is such a strong program hanging around then authors of commercial programs may feel intimidated about improving their program and even become disappointed, as also buyers of Chess programs will not prefer the commercial with the similar strength with the free one.

Because Stockfish is free, commercial programmers can study it, reimplement the ideas they like in their own programs, and stay ahead.

It is still possible that it will some day be impossible to make significant amounts of money from computer chess, of course. If and when this happens, it is only because chess programming has become sufficiently easy and well understood that hobbyists like us can compete on the same level as professionals. This is a very natural development, and is happening in many fields. Trying to slow down progress in order to prevent it from happening is just silly.

I am not sure that the above list of improvements would actually harm commercial chess program sales.

The things you mentioned such as web sites, technical support, marketing, etc. are all needed by the consumer to have a successful technical product for the unwashed masses.

So even the existence of equal quality will not necessarily cause a dent in what the commercial programs rake in.

As evidence, I offer ChessMaster --> by a stupendous leap the biggest money maker in commercial chess programs, but several hundred Elo weaker than Rybka. But the many other things that CM does offer are no coincidence as to why the sales are so high:
Presence everywhere: (you will find CM at Fred Meyers and other department stores)

Nifty web site with simple digital download:
http://digitalstore.ubi.com/178/5154/pr ... n-Download

Technical support, reference materials, shiny box with neat looking old guy on the cover, etc.

[Dann Corbit]

Code: Select all

no Program Elo + - Games Score Av.Op. Draws 
2 Rybka 3 x64 2CPU 3155 12 12 2573 76.7% 2947 32.3% 
3 Rybka 3 w32 2CPU 3137 22 22 728 78.0% 2917 32.1% 
7 Naum 4 x64 2CPU 3056 17 17 908 62.6% 2967 42.2% 
8 Deep Shredder 12 x64 2CPU 3054 24 24 419 57.4% 3003 46.5% 
9 Naum 4 w32 2CPU 3039 18 18 742 59.4% 2973 48.0% 
18 Deep Fritz 11 2CPU 2978 15 15 1140 50.7% 2973 47.8% 
22 Naum 3.1 x64 2CPU 2960 12 12 1718 55.4% 2922 43.2% 
24 Thinker 5.4Ai x64 2CPU 2958 19 19 743 53.6% 2933 42.9% 
25 Zappa Mexico II x64 2CPU 2958 10 10 2671 50.9% 2951 43.5% 
27 Shredder WM Edition Bonn 2CPU 2948 16 16 1057 51.5% 2937 40.3% 
28 Fruit 2.3.5m p15 w32 2CPU 2946 16 16 1039 49.8% 2948 44.6% 
29 Naum 3.1 w32 2CPU 2939 23 23 507 53.0% 2919 44.4% 
31 Stockfish 1.4 x64 2CPU 2937 18 18 776 50.1% 2936 44.8% 
32 Deep Shredder 11 x64 2CPU 2937 10 10 2799 52.5% 2919 39.5% 
35 Deep Sjeng WC 2008 x64 2CPU 2929 14 14 1432 50.0% 2929 39.5% 
38 Hiarcs 12 MP 2CPU 2923 11 11 2483 49.9% 2923 40.8% 
41 Cyclone 2.3 2CPU 2919 24 24 454 50.3% 2916 43.8% 
42 Deep Sjeng 3.0 w32 2CPU 2917 22 22 548 49.6% 2919 43.4% 
47 Deep Shredder 11 w32 2CPU 2902 23 23 518 47.7% 2918 40.2% 
50 Toga II 1.4.2 JD 2CPU 2895 19 19 708 50.8% 2889 46.2% 
54 Zappa Mexico II w32 2CPU 2890 19 19 696 46.8% 2913 44.1% 
59 Glaurung 2.2 x64 2CPU 2881 19 19 724 45.9% 2910 41.3% 
60 Bright 0.4a 2CPU 2878 15 15 1230 48.4% 2889 41.4% 
71 Cyclops 2.0 2CPU 2826 32 32 302 40.1% 2896 33.8% 
73 Spike 1.2 Turin 2CPU 2818 8 8 4512 46.6% 2842 39.8% 
77 Deep Junior 10.1 2CPU 2813 11 11 2699 43.6% 2858 32.8% 
80 Jonny 3.08 2CPU 2804 22 22 579 50.0% 2804 38.5% 
86 The King 3.50 x64 2CPU 2759 20 20 722 49.9% 2760 37.4% 

[Tord Romstad]

I am not sure that the above list of improvements would actually harm commercial chess program sales.

The things you mentioned such as web sites, technical support, marketing, etc. are all needed by the consumer to have a successful technical product for the unwashed masses.

So even the existence of equal quality will not necessarily cause a dent in what the commercial programs rake in.

As evidence, I offer ChessMaster --> by a stupendous leap the biggest money maker in commercial chess programs, but several hundred Elo weaker than Rybka. But the many other things that CM does offer are no coincidence as to why the sales are so high:
Presence everywhere: (you will find CM at Fred Meyers and other department stores)

Absolutely. Too many people on this board assume that commercial success in computer chess is mainly about having a strong engine. GUI quality, support, features, entertainment value and marketing are all vastly more important than engine strength (and don't forget that writing a good chess GUI is far more difficult and time consuming than writing a top chess engine).

While there is a market for strong UCI chess engines, this is an extremely tiny niche market. Most chess players would have no idea whatsoever what to do with a UCI engine.

[mcostalba]

I'm not sure what to make of the contradicting statements whether this helps or not, but I guess it's easy enough to disable it and see the effect.

Tord got the idea of polynomial balance, me and Joona took the task of populating the corresponding tables with proper coefficents. To make it work you need _both_ . The idea of course, but alone is far from enough and me and Joona worked for more then a month to come up with seemingly acceptable coefficents.

Tord was expecting (and still is) more then achieved. For me it is already enough. We didn't verified ELO improvement with a broad range of engines that's the reason we cannot give you a single number. I have said that this is IMHO the biggest improvment in this release and, altough Tord was expecting more, I would think he agrees on this.

Regarding the general tuning metodolody we almost never use human heuristic to pick up proper parameter values. We almost entirely rely on automatic tuned values. And, yes, you need a good CPU power...but not a cluster

[mcostalba]

Can't you just rip out the table? You probably don't need to bother with the difficult optimization step if you already got the results.

That would be an interesting experiment, but I don't think copying the table would work well in most programs. Optimal material balance scores depend too much on the rest of the evaluation function, especially mobility and piece square tables. Parameter settings which are good in program X usually wouldn't work well in program Y, unless the two programs have very similar evaluation functions.

Well, in this case you can rip out also evaluation

Why not ?