Why do you play against GM Adams and not against Kasparov or Anand?

We play against GM Adams, because the match is in London and we wanted a "local hero" as our opponent. We think that GM Adams is one of the best - maybe even the best - player against a computer. He has a very sound positional style. Adams can play like Adams. Kasparov is one of the best players ever, but against Hydra he can not play like Kasparov. When it comes to tactics, to sharp positions, Hydra is better. One can only kill her softly. Besides this is GM Adam a real gentleman. We think the match will also be a nice event for anybody involved.
If the London encounter creates some interest, we would like to play with other top-players too. It is a matter of finding sponsors.

What are your expectations for the match?

As already noted above is GM Adams a very difficult opponent for any program. We are nevertheless optimistic that Hydra will still have a clean state in the tournament statistic against humans. 2 wins for Hydra, 4 draws.

Do you develop especially against GM Adams?

No. We play as usual against Shredder, Fritz and against a combination of the Hydra-chess expert GM Lutz+Shredder+Fritz. The goal is to improve the active, dynamic playing style of Hydra. This was our goal from the beginning of the project and is not directly related to GM Adams. Hydra shall simply play her game. We assume, GM Adams will try to do the same.

Is there some special opening preparation done?

Not directly. The Hydra opening book is very short. Typically 10 moves. After 10 moves we let the monster from the leash and rely on the playing strength of the program. We know the favorite opening lines of GM Adams and try to play the most active variations. But there are no attempts to develop some novelties or opening traps. We think it is nowadays almost impossible to find real good alternatives within the first 10 moves. Hydra shall find some new lines on its own after the 10th move.

On how many processors is Hydra-running?

The Hydra-Scylla Cluster consists of 64 Xeon CPUs. But for the Adams match we are using only 32 of them. The time critical part of the program runs on special purpose hardware cards (see also comparison of Hydra to Deep Blue). We have currently “only” 32 of them.

What is the key for the recent successes of Hydra?

The Hydra team-members are coming from and are living in different countries. We speak different native-languages, are members of different religious groups and have different professional backgrounds. But we are one team with a common goal. The Hydra team is the proof, that the so called “clash of the cultures” can be very productive and beneficial. All it takes is the ability to accept different backgrounds and the willingness to learn from each other.

How would you compare Hydra and Deep Blue?

The basic architecture is the same. Both are mixed software-hardware systems. The first part of the chess search is done in software, the last – time critical part – in special purpose hardware cards. The type of these special purpose hardware cards is different: Deep Blue used so called ASICs (Application Specific Integrated Circuits), Hydra uses FPGAs (Field Programmable Gate Arrays). An ASIC is traditional hardware. They are once designed, a mask is made and after fabrication they can not be changed any more. FPGAs can be reprogrammed with a high-level programming language. ASICs are faster than FPGAs, but they are not flexible. One can tune an FPGA program in the same way than a conventional chess program. FPGAs combine the best of the two worlds.

Is Hydra as strong as Deep Blue?

We think Hydra is considerable stronger. The Deep Blue team had a strong background in hardware-design, but they lacked the long computer chess experience of the Hydra team members. The Hydra team knew from the beginning what is important and what is unimportant in computer-chess. Additionally are FPGAs easier to program and especially to fine-tune than ASICs.

One example is the different search mechanism: Deep Blue and Hydra have about the same raw speed. Deep Blue searched in typical chess positions 12 moves (plies) deep, Hydra 18 plies. Deep Blue was the last brute-force searcher. Every possible combination was searched to 12 plies. Hydra uses sophisticated pruning techniques. She does not waste resources for irrelevant variations. Pruning away variations involves the risk to overlook some combinations, but the considerable greater search depth overcompensates this risk.

Hydra uses also the multi-processor machine much more efficiently The Hydra-team member Mr.Parallel Ulf Lorenz is a specialist in this field. The Univ. of Paderborn he works on is one of the leading research centers for parallel computing.

The evaluation is in Hydra more dynamic than in Deep Blue. Positional factors are evaluated higher and more accurate. DB chief architect F. Hsu noted, that the king-attack terms are in Deep Blue too low. But it could not be changed easily due to some initial design decisions. The Hydra design does not have these restrictions. The German top-GM R.Huebner wrote in an analysis of the Bilbao 2004 human-machine team match: Hydra seems to be the only program which plays systematic king-attacks.

Is Hydra something revolutionary or an evolution of Deep Blue?

It is the next step in the evolution of chess programs. We have improved a lot of details, use new hardware technology (FPGAs), but our starting point was Deep Blue. We are ahead of the crowd in one point: We created the first successfully working FPGA based High-Performance-Computer.

Is Hydra just a chess program or only a test bed for a new sort of High-Performance-Computer?

Our goal was to design the strongest chess-engine ever. But we have created a very powerful and cheap High Performance Computer. The Hydra chess engine is the proof, that this HPC is not just a “paper-tiger”, but a successfully working machine.

Can you say something about the Hydra-chess-evaluation?

Basically it is a dynamic evaluation. Dynamic evaluation means: An extended move generator is called for both sides. One can calculate from this information simple things like mobility (number of legal moves for each side), but also more involved ones like king-attack, pinned pieces, center control... There are of course some terms for pawn-structure and passed pawn. But the dynamic part is the most important one. Dynamic information is expensive to calculate in software, it is relative cheap/fast in hardware. The Hydra-FPGA needs only 2 clock-cycles (but not at 3 GHz, at mere 55MHz). This full dynamic information differentiates Hydra from the PC-programs. They have a dynamic evaluation too, but on the PC one has to implement a compromise between speed and evaluation accuracy.

"I guess Hydra does not learn to play better chess, except maybe by remembering its own games. - Or does it"?

Hydra has no automatic learning mechanism implemented. It is the team which learns from (bad-) experience. We play games against Shredder, Fritz and against the Hydra chess expert GM Lutz+Shredder&Fritz. Then we look at the games and try to identify weaknesses. Either the value of an existing evaluation feature is wrong, or the feature itself has a flaw or there is some feature missing. Then we adjust either the parameters or implement/improve the feature. Then we play again... No big science, but a lot of work and 15 years of experience in chess programming. We think automatic tuning can not - yet - compete with this experience. But maybe some day computers are not even the better chess players but also the better chess programmers. Then we will tune in the same way the learning algorithm.

"Do you find better results with more complicated or less complicated algorithms"?

An important aspect of the art of computer-chess-programming is to find the simplest possible and especially the most general solution. E.g. at the beginning of computer chess a lot of rules were tried to prune uninteresting lines. Then someone had the idea: Just make a passing move and do a shallow search. If the opponent can not refute the passing move, there is no point of trying out legal ones. This is the so called "Null-Move-Heuristic". A very general rule, besides Zugzwang detection, there is no direct chess knowledge involved.

The same holds for the evaluation. It is hopeless to implement in a program thousands of special chess rules. There will always be one important rule missing and it is practically impossible to specify human knowledge in a precise way. To give one example: Strong human players usually know for each type of position how to play a king-attack (or to avoid the king attack of the opponent). Hydra does not have this special knowledge. An important part of the evaluation is a general king-attack-term. Basically: Which pieces attack and defend the king. The search tries to optimize its own and minimizes the opponents attacking chances. GM R.Huebner noted, that Hydra plays a systematic, planned king-attack. This is not entirely true. The program has no direct notion of a plan. The optimization process of the search has a similar effect than the human planning process. It is playing along a plan without having one.