Rybka 4 is here – and stronger than ever
The anticipation for the new Rybka was intense – some people paid a huge
amount of money to get an advanced copy of the engine for their match preparation.
Now everyone can have it and enjoy the new playing strength and features. We
went through some of the details with Rybka author Vasik Rajlich.
Question: How much stronger is Rybka 4 compared to Rybka
3? Which areas were improved especially? Is it possible to estimate the Elo
of Rybka 4?
Vasik Rajlich: Rybka 4 is improved in many areas. The search
is faster and more efficient and the evaluation is more accurate. I worked on
almost every area to some extent. One big area of improvement is in understanding
king attacks. As for Elo – that is a question for the testers. The exact
figure will depend on test conditions.
How do you work on improving Rybka?
I'll typically pick one area to concentrate on, and within that area try dozens
of little changes, each of which I reject or confirm based on the results of
an automated test procedure which I've gradually developed over the past few
years. The process is very incremental – there have been more than 3000
unique Rybka versions, each playing differently, since Rybka 3.
How much does Rybka improve, when more processors are used in parallel?
A good rule of thumb is that when you double the number of processors, Rybka's
'speed' increases by a factor of 1.7.
How do you describe playing Rybka's style? Is Rybka especially strong
in certain openings or types of positions? Do you recommend a special openings
book for Rybka 4?
Rybka's style evolves and I'm probably not the best judge of it. One thing
I'll say is that over time Rybka has gotten more aggressive and more tactical.
Rybka versions up to Rybka 2.3.2 were actually relatively weak tactically compared
to other top programs, and many grandmasters noticed it. When it comes to computer
chess, GMs in my experience tend to be an interesting blend of ignorance about
technical issues and some pretty savvy insight about where each program is not
reliable. Anyway, being more tactically aware has been one major area of focus
for me for the past few years.
As for the opening book – Jiri Dufek is the author of the Rybka 4 opening
book. He's been using Lukas' cluster to analyze and test and this book is much
deeper and more accurate than every book which is publicly available right now.
Note however that this book will be 'objective' – Jiri's goal has been
to find the truth, rather than to find variations which suit Rybka. This is
an intentional decision by our team. Our goal is to create objective analysis
tools. This is what users want, and it also simplifies Jiri's life somewhat,
as he doesn't have to worry about things like how Rybka is evolving.
How is the relation between search and knowledge in Rybka 4?
Knowledge guides the search – Rybka searches those things which her heuristics
tell her are important.
Do the pieces have always fixed material values or do they change with
placement or influence on the given position? Does e.g. the pair of bishops
get a special bonus? If yes, how large is it?
There are no real restrictions on Rybka's heuristics and how they are formulated,
they are interrelated in whatever way I think is best. Many heuristics involve
combinations of many features and they can become quite elaborate over time.
If I think a heuristic will work, I try it, and if it works I keep it.
The bishop pair is of course a major item, there are all sorts of heuristics
related to it. On average, the bishop pair seems to be worth around half a pawn.
The exact value depends on things like whether the position is open or closed,
how many pieces are still on the board, and how many minor pieces the opponent
has. Bishop pairs like an open board, and they don't like opposing minor pieces.
None of this is any secret, any grandmaster (or FIDE master, for that matter)
could tell you this.
Does Rybka have different evaluation functions for middlegame and endgame?
Yes, of course, but the transition is smooth. Abrupt transitions are very difficult,
to formulate and to maintain later – they tend to create unexpected problems.
Situations where abrupt transitions are appropriate are difficult for computers.
Is there a lot of special endgame knowledge implemented?
There is some, but I try to avoid really special cases. I look for heuristics
which trigger in (let's say) at least one position out of a thousand, or something
like that. Of course, the problem is that endgames are full of special cases.
This area needs a lot more work.
Does it help to use endgame tablebases? If not, when should they be used
when analysing endgames?
Sure, tablebases help. Rybka will misevaluate a lot of tablebase positions
without them. I think it's best to always use them – there is no reason
not to. They are sometimes probed even when analyzing middlegame positions.
Does Rybka evaluate and handle rook endings well or does Rybka often
underestimate the drawish tendency?
Definitely the latter. I did a little bit of work on this, but a lot of work
is still left.
How do you see the future of computer chess?
For the next two or three years, I'm planning to concentrate on playing strength.
This is still the most important issue right now. High Elos nonwithstanding,
Rybka is still too blind in too many positions. Together with ChessBase and
other publishers, we'll also add various simple analysis features during this
time. Our main criteria for analysis features right now is that they should
be intuitive, non-intrusive, and frequently visible. Most users tend to have
very little patience for hidden features which require a steep learning curve,
and even less patience when those features make other tasks more difficult.
Vas Rajlich with his wife Iweta in Budapest, where the two lived for some
After that, we will concentrate on real chess training. We will try to find
better ways to communicate the chess knowledge from the engine to the user.
A computer program should be just as good at training a human as it is at playing
a tournament game. Of course, computer training will always have its relative
strengths and weaknesses – a computer program has graphical capabilities
which no human trainer can match, while linguistic expression will probably
be awkward for machines for a long time to come. There are of course also areas
(for example, the psychology of practical play) which are relatively difficult
for a computer to understand.
The first load of Rybka 4 arriving at the ChessBase office in Hamburg
A screen dump from the new Rybka 4 program
3D boards with scoresheet and analysis board
Order Rybka 4 now in the ChessBase