Furure of chess engine programmers

[fern]

What our friends here that create engines as a closed system of routines, database, etc will do facing the emergence of these neural entities? What is the sense to work year after year to perfect an engine at the same time a new monster learns the game and develop to the level of a master or more in just some hours?

I assume that the neural things tends to be the same no matter who program it.

Fern

[brianr]

The machine learning neural net (NN) area is moving very rapidly with developments and new approaches appearing all of the time. Some "traditional" engine folks will lose interest, many will shift to the NNs, and there will also be new folks joining the ranks.

All NNs are far from the same. Just implementing one that plays good chess is far from trivial at this point. In addition, there are many different NN architectures and methods. Which ones are the best for chess? Nobody knows. How small a NN can play good chess and fit in a cell phone? Nobody knows. On top of the algorithm choices and the software options (which NN framework), there is also exciting new hardware to be used. Moreover, there is broader crossover from other areas besides just chess or games (like computer vision).

From my perspective, things were getting somewhat stagnant in computer chess and the new NN developments have actually rekindled a lot of new interest. While I can see how the high-level things tend to work, I don't pretend to understand a lot of the details. So, some of the results do seem a bit like magic.

[frankp]

I think it is wonderful to see a new approach.
The current NN (A0/leela) is unlikely to be optimal for chess - made for Go. There is much scope for improvement. And of course for hybrid approaches.
I guess the main drawback currently is the millions of games needed to train a net. WIthout DeepMind type hardware or a distributed project like leela, it may be a while before an individual can craft thier own engine at home, in a manner that is easily possible for a traditional AB search engine - with all the resource available to make it GM+ level at first flush once the bugs are removed.
Exciting times, i would say. Fail to understand the negativity displayed here at time to be honest.

One other aspect that struck me, is that even the weak NNs played seemingly sensible chess but was beatable - often by simple tactics and in the endgame. Much more enjoyable to play against for the casual hobby player perhaps. Or at least one of my very limited ability.

[hgm]

I assume that the neural things tends to be the same no matter who program it.

Actually this is what is the case with current engines based on alpha-beta search. This is so far developed that all engines are essentially the same, and any significant deviation from standard procedure makes it worse.

With the neural nets nothing is known, and hundreds of completely new tricks (that yet have to be discovered) can be used to improve the performance. How can the info on the position presented to the NN best be pre-processed? Which network topology is best? Which move-selection procedure in the Monte-Carlo search would work best for Chess? Basically AlphaZero was just a blind guess on any of this, perhaps guided by what was good for Go (a completely different game).

So finally there is room for creativity again, we no longer have to do all exactly the same, with a minuscule variation on it.

[Jhoravi]

In the future the Leela network will become like a Tablebase. All chess programs will interact with it.

[jkiliani]

In the future the Leela network will become like a Tablebase. All chess programs will interact with it.

Actually I don't think that the Leela network by itself will be that, but its training data will be the basis to train new networks for many other chess engines, probably on many different network architectures. A lot of deviations from the Deepmind neural net architecture have been suggested, which may be tried much more easily if there is a set of strong training data already available.

[Kotlov]

I think the engines based on handmade-programming will be called the "oldschool" ))

[frankp]

In the future the Leela network will become like a Tablebase. All chess programs will interact with it.

Actually I don't think that the Leela network by itself will be that, but its training data will be the basis to train new networks for many other chess engines, probably on many different network architectures. A lot of deviations from the Deepmind neural net architecture have been suggested, which may be tried much more easily if there is a set of strong training data already available.

If it is technically possibly to used existing training data on a different NN structure then an AI project to maximise the NN structure for chess (outcome) might be interesting, if a little scifi. Having dumb humans floundering around in the dark trying to discover the best NN structure for chess is perhaps not optimal.

[Henk]

Even if neural things would be the same the values of the parameters (mostly weights) are different.

[Henk]

New school using black boxes. Dangerous. Fortunately the output of chess is not harmful. Only might drive you crazy.