Houdini 3 Pro vs Deep Fritz 13

For those with older hardware, yes. But the glitch affects output analysis for the most part. The playing strength isn't affected too much. I'm actually using a Core 2 Quad, and Houdini 3 is demolishing Houdini 1.5a. (Which backs up the mail received from Houdart.)

So, you have not tested and do not intend to test Houdini 3, yet you make major claims about the glitch, which goes straight against all the people that used it on older hardware without any real problems.

I take your sarcasm as offensive.

@FEDTEL: Yes, I am Houdini's author.

@pfren: Your interventions are malicious and uninformed. There are no known problems with Houdini 3 other than the one that was reported and corrected about 10 days ago, and which was a fairly innocent issue that impacted a minority of the users.

Houdini 3 is not "perfect" - that is your description. It is about 70 ELo stronger than Houdini 1.5a, which means that it will still lose the occasional game and even the occasional match. But over-all there are no grounds for your negative comments.

Because Houdini 3 lost a single game to a lesser engine, that mean's it's worthless? Then all other engines are even more doomed - they lose or draw more than the majority of the time against Houdini 3. You're saying, Houdini 3 lost one match out of many, that means it's unreliable; and how incredibly reliable is an engine that loses all the time against Houdini? Even less reliable.

You are the ignorant one. You don't know the severity of the glitch and you spew out ignorant comments. Yes, let's bash the best chess engine because it had a MINOR glitch that was fixed TEN DAYS AGO.

 

Ahh, I see. He sure is persistent. (And rude.)

I am having trouble interpreting how/why Houdini is taking so much longer to score a high depth on Let's Check. It is making me wait about 15 hours to score 33 depth on Houdini when it would just take a few hours on Fritz 13.

My best guess is that Fritz' self-proclaimed depth ratings are a bit inflated.

Correct if I'm wrong: the depth/knodes etc cannot be compared head to head across different versions of the same engine, let alone different engines. (What I read.)

"Let's Check" takes the three highest depths regardless of which engine and ranks them so that they can be referenced by anyone. So maybe what you are pointing out is a flaw in the system, but this works for Shredder, Rybka, Houdini, Fritz, Stockfish, or whatever else has access to "Let's Check".

I was just wondering why DF 13 thinks it is going deeper than Houdini (in significantly less time), especially since Houdini takes up more CPU resources. I just assume that DF 13 branches out in a more limited way when it does its analysis, which enables it to post higher depth ratings (total speculation I don't really know). 

I think it's fun to try and get on the top three for early and major positions; I was just noticing that it takes longer with Houdini. No idea why.

I am very confused now about how many candidate lines I should be feeding the engine.

After 24 hours of having it evaluate black's response to the Zukertort with 4 active cores and 4 candidate moves, it ended up with 1.Nf6 as the best choice at 33-depth.

As a test to see if I could even get to 34-depth, I assigned 1...Nf6 as the only candidate move by pruning the other three lines, and using 6 active cores instead of 4. It has been on 33-depth all day long for 1...Nf6, and now it suddenly changed its assigned candidate line! (to 1...e6 - Queen's Gambit Invitation). This also happens to be what Fritz ended up with at 33-depth but I am thorougly confused at how it overrode the one it was working on all day and suddenly came up with a new one (after exactly 19.5 hours). Going to let it run another 12 hours or so to watch what happens.

I know many are against the method of trying to create a repertoire in this fashion, but if you were in favor of this method, how many candidate moves would you have it examine? I am intrigued that if I only gave it one it is capable of changing its mind much later on, but obviously it's not productive if I have to wait almost a full day for it to do this. So maybe I should continue letting it examine four. 

Update: @22 hours it switched to 1...c5 (Sicilian Invitation), changing the whole line! Then at 22.5 hours it actually hit 34 depth, sticking with the Sicilian invitation. At 32 hours it still is at 34 depth and 1...c5.

I just wanted to thank everyone (even pfren) for the input.

I have been realizing (like everyone was saying) just how unreliable engines are for the very first moves. I have been analyzing 1.e4 for a few days now and it has the French towering above all the other openings by a wide margin. I know the French is a sound response, but you can tell that the computer likes it so much because it is preparing itself as it would against a supercomputer. It reminds me of Kasparov's games against Deep Blue when he would just build a little French-like fortress to shield himself against heavy brute force calculations.

So for this project I decided just to go with the three openings that interest me the most, even though they are not favored by computer analysis. For black I am going to try the Dutch Leningrad for all first moves by white except 1.e4, in which case I will use Alekhine's. The idea behind these openings mainly (especially the Leningrad) is to overwhelm white with complexity instead of trying to equalize the game. "If both players are confused this is a moral vicory to Black, who started off with the worse position" (2004, Neil McDonald).This kind of complexity, however, would no doubt strongly benefit from engine analysis; so I will continue this project for the next 20 years or so, just restricted to these two systems.

I am still not fully decided on what to do for white (if anyone wants to make a suggestion) but I am pretty sure I will go with the English. Reason being is that I feel uncomfortable whenever I play against it (also when I play it myself). It is a positional nightmare requiring many subtle move-order dependent operations, and so seems like a perfect project for an engine-assisted repertoire.

However, the earlier in the line, the less I will consider the engine's input. But by about the 4th move, I will strongly consider the engine's results. Earlier on, I have to put more emphasis on how games with these lines actually turned up, and what grandmasters have to say about them. I ordered a few books on the Leningrad so I will use that as a starting point (while I spend 2+ weeks analyzing the first move for white).

I am still interested at whatever the computer thinks is the best move for all the first positions, seems like good data even though I don't know why. So I will still be stubbornly trying to hit record depth for all primary early moves. I may set aside this information in a section like "What the computer would do?" Anyway, I would be saving people time (that have "Let's Check access) who want to see results for early position—so they will be able to see 37-ply results in a second rather than spend week(s) calculating one move.

It increasingly more common these days. The hex-core i7-3930k is just under $600 which isn't too bad. Anything past that starts to become quite expensive and some are made only for servers. Most computers have 2-cores at minimum and many desktops are quad-core.

The i7's have a virtual core for each physical core, which mean that a quad core really has 8 logical cores. The i7-980x and 990x CPUs have 6 physical cores, which means they have a total of 12 logical cores. 

Unfortunately, I went cheap on my cpu and only got an i7-950. I originally built my system for gaming and CUDA programming (which uses graphics for calculating power) but apparently cranking out floating points isn't that helpful for chess engines, so none of the chess programs use CUDA.

Deep Fritz 13 offers the opportunity to use up to 8 cores, but since I went a little cheap on my CPU, I only use 4 or 6 depending on how much I will be multitasking. I technically could use 8 cores but so far have been nervous to do this because I was trying to avoid a crash.

And as advice to anyone wanting to buy a system for this (besides getting a flagship CPU with 12 logical cores) would be to get a lot of RAM (in the ballpark of 12GB). I have always been an advocate of low capacity but very fast and high quality RAM, but now I am feeling my stupidity there because I have to constantly stare at my RAM widget to avoid hitting 100%. This means I have to watch how many windows I have open while browsing etc. Right now my RAM is hovering at about 85% and I am only using 4 cores on Deep Fritz. I am 72 hours into a calculation though and *I think* that more RAM resources get used when you are in 34+ ply.

As a datapoint, I am using 3-channel Corsair Dominator RAM (2G X 3). It has low latencies and a high clock rate, but the lack of capacity is really killing me here. And to repeat since my post was a bit long and dry, my overkill-expensive graphics system is completely worthless to me for this, and I would have been better off getting a CPU with 12 logical cores. 

However, if you don't use more cores than your computer can chew, you should always be able to compete with the top calculation for any position in "Let's Check", regardless of computer strength, if you are willing to wait enough time. That and if you don't accidentally touch your mouse wheel with Fritz open because that will reset the calculation...

I just wanted to share an amazing development, which is also an example of a previous question that has still gone unanswered.

I began analyzing the first move for white as a showdown between 1.e4 and 1.d4 (using 6 cores). I assumed that none of the other positions would be able to compete with these two so I cut out all the other candidate moves.

8 hours later, I wake up and 1.Nf3 replaced 1.e4 as a candidate move!! Now 1.Nf3 is tied with 1.d4 (with the same score). 

First of all, wow. I haven't been giving much credit to the computer analysis of these very early lines, but if 1.Nf3 keeps performing strong a few days deep into this calculation, then I will be considering it as a new main line (instead of 1.c4 or 1.d4). I am very sick of the Queen's Gambit, and so am rooting for the Zukertort here (which I always assumed was inferior).

Anyway, this brings up the question of how the engine is working or what is going on here? I had it analyze two candidate moves and it just replaced it? Does this mean that it is unecessary to analyze three or 4 candidates because the best one will surface? How did it surface?? Is it related that I was using 6 cores for only 2 candidate moves?

Hopefully Robert can shed some light on that one. Would be happy to hear anyone else's ideas though. 

Impenetrable thickets.

I have a theory on this. The 1.Nf3 line seems to have been "derived" from the 1.d4 line. It looks like a transposition with almost the same move orders.

From the fourth move on, they are perfect transpositions (which explains them being tied with the same score). It's still pretty amazing that the Zukertort prevailed like this, and it would be interesting if this transposition ended up transforming into its own unique line, and then outperforming 1.d4! I will probably be running this calculation for the next full week, so will report back on that.

Back to that theory, just like how it shuffles around the second, third, and fourth positions, scrambling their move orders to try and find the highest-scoring sequence, it also shuffles that first move. And in this case, this transposition outscored 1.e4 so it kicked it off the list.

But since I'm only allowing two lines, now 1.e4 will no longer be evaluated. This kind of defeats the purpose of the contest between the two central pawns, but I guess 1.d4 won fairly since this appears to have happened around 33-ply?

By default, Deep Fritz 13 has you evaluating only two lines. But there seems to be a flaw in this if transpositions can permanently displace the other candidate move.

PS. Where it says 9628 kN/s in the screenshot, that is how fast the calculations are taking place. With 8 cores instead of 6, it will spike up to about 10,000 but not higher. This shows that there are signficant diminishing returns with extra cores (so no need to worry too much if you don't have that many cores at your disposal).

Virtual cores stink on Houdini 3. Read the manual; it states that virtual cores up the 'speed', but other factors make it actually slower. Make sure you use the auto-tune function in the engine, that the hash is set to the correct value for the analysis time allocated, CPU is overclocked, etc.

Having said that, having real 6 cores isn't very rare. While I think computers will outclass all players and play better moves than any human for vast majority of the time in middlegame, I do not think using a computer to analyze the first few moves is a good idea, no matter how strong your computer is. By the time it is, I think we're getting relatively close to solving chess.

I think the time could be better spent analyzing relatively common middle games that result from popular openings, or start in positions like poisoned pawn Najdorf. This type of analysis is good for adding new lines for opening books.

Just for kicks, a while back I let Houdini 2 Pro analyze the first move for 48 hours on a (then) stock Q6600, it went over 30 depth, I think 31 or 32, and suggested d4. Giving each move 1 day after that, I think it chose to play Queen's Pawn, Declined or something like that every time. 

Do we need to be explicit about why computer analysis of opening moves is relatively unproductive?

From my understanding: The opening has literally too many possibilities. The computer plays chess by calculating possible moves for other side, then our side, so on and so forth. While looking 30, 40, 50 moves ahead would be decisive in a middlegame, in the start, those 30, 40, 50 moves won't be accurate at all.

Houdini is very selective in its choices. It will probably only consider looking at a few lines, and miss any other moves that look bad to a person with no positional knowledge. Even if Houdini just looks at e4 alone, there are too many posibilities to even consider calculating that, because an opening move might look decent tactically at the start might end up with a pretty dumb pawn structure in the endgame, and the engine's not going to calculate from start to endgame.

Whereas a human inherits the knowledge of opening theory, of positional values, at the only time in the game where it causes the human to have an advantage. We've also collectively looked at countless openings and possible variations, and discarded ones we know are bad. 

You'd be hard pressed to make Houdini play Najdorf or any other Sicillian without an opening book no matter how long you leave it on to analyze.

So if a computer, which makes better moves by looking into all the possibilities, can find great opening moves at the start of the game where there are most possibilities, doesn't this foreshadow chess being close to solved?

I'm not on Pfren's camp, either. Engines have opening books for a reason: To crush players from start to finish. And those books are getting stronger day by day.

At least with the current technology contraints, it can only approach "solving chess" from relatively deep positions. The earlier you are in the game, the weaker an engine will be because there is just too much to consider. 

Anyway my computer crashed analyzing white's first move... so I reduced it to three cores but also three candidate moves this time. It has thus far ranked them in this order: 1.d4, 1.Nf3, 1.e4. 

What this at least tells us is that the Zukertort (1.Nf3) is both playable and respectable. PS. Houdini's definitive answer to the Zukertort after a few days and 34-ply actually is the Sicilian. 

But as long as you structure your priorities according to how deep you are in the game, you can keep engine analysis righfully adjusted. This means the deeper you are in the game, the more credibility/authority you can give to deep analysis. But early on, it's more important to see what grandmasters have to say about the strategic themes of each move, as well as to consider which moves have the best stats on the highest (human) level. 

Another point I dare say, which has not yet been brought up in this thread, is that our choice of openings should be more about the flavor of game that we prefer to play. We should find openings that interest us and generally provide the kinds of games we like to see/play. From there computer analysis becomes extremely valuable. If anything, memorizing as many computer-approved lines as you can will save you from falling for early traps and therefore give you a solid chance at either drawing or winning. 

update: 1.Nf3 now bumped 1.d4 off for the #1 spot! @31-ply.

The most it can say is the Zuckerfort is a decent opening, among all other other good openings. I don't see a point in spending time to analyze those moves.

Frankly, these chess engines are about as useful as asking a Porsche how to run faster.