Is Allie a 'Leela net player'?

[Madeleine Birchfield]

You have a misunderstanding of how training works in leela land. The engine does not perform the training. A tensorflow framework is used to train nets using either self play or other data. My training framework is written using pytorch, not tensorflow, but all of them look much like image classification resnet training frameworks with lots of boilerplate, I.e. if you had to call it a clone, it would be a MNIST clone. Otherwise you can call them Alphazero type networks, or Allie networks, or Scorpio networks, or a0lite networks, which is most appropriate, as the latest Bad Gyal networks are not compatible with leela.

Even if we restrict to data only, I think a0lite is already using Bad Gyal in TCEC, so according to the rules Allie wouldn't be able to use Bad Gyal. Since Tiny Gyal uses the same data that Bad Gyal uses, if TCEC applied its rules fairly to all engines, Allie wouldn't be allowed to use Tiny Gyal either. The current Stein network that Allie uses in TCEC used Leela data for training, so if TCEC applied its rules fairly to all engines the Stein network would be disallowed and Allie would have to find or create another net to use. That being said, I think the new TCEC rules require the data used to be from the engine itself, so if TCEC applied its rules fairly to all engines, Allie would have to use a net trained using Allie data, and a0lite would have to use a net trained using a0lite data. And Scorpio would have to use a net trained using Scorpio data.

[Madeleine Birchfield]

Chess engine programmers really should have a pile of sand in the backyard plus a kiln and make their own CPUs

What is this even supposed to mean?

[dkappe]

Even if we restrict to data only, I think a0lite is already using Bad Gyal in TCEC, so according to the rules Allie wouldn't be able to use Bad Gyal. Since Tiny Gyal uses the same data that Bad Gyal uses, if TCEC applied its rules fairly to all engines, Allie wouldn't be allowed to use Tiny Gyal either. The current Stein network that Allie uses in TCEC used Leela data for training, so if TCEC applied its rules fairly to all engines the Stein network would be disallowed and Allie would have to find or create another net to use. That being said, I think the new TCEC rules require the data to be generated by the engine itself, so if TCEC applied its rules fairly to all engines, Allie would have to use a net trained using Allie data, and a0lite would have to use a net trained using a0lite data. And Scorpio would have to use a net trained using Scorpio data.

Why would Allie want to use a tiny 16x2 network or a 128x10 network in TCEC? It makes no sense. I just wanted to point out that not all networks — like the Gyal and Ender networks — are leela networks. Some of them won’t even run in lc0.

What is a0lite data? You seem confused.

[the_real_greco]

Good luck writing specific uniqueness rules to handle every case. Especially the ones you haven't dreamed of yet. You'd wind up rewriting the rules every season, to fit whatever engines you wanted to include anyway.

So what I guess what I'm saying is, including Allie is more fun than having two engines compete for two SuFi spots? And I'm not sure there's any other question to ask.

[Tony P.]

^^ Before the NNUE hype started, many thought that small nets wouldn't work, but now that can be reconsidered.

That being said, I think the new TCEC rules require the data used to be from the engine itself, so if TCEC applied its rules fairly to all engines, Allie would have to use a net trained using Allie data, and a0lite would have to use a net trained using a0lite data.

There's however another option that makes sense from the aggregate viewer satisfaction standpoint: to persuade Dietrich to withdraw a0lite from QL in exchange for credit as an Allie co-author if Allie does manage to benefit from a small net.

[Madeleine Birchfield]

Good luck writing specific uniqueness rules to handle every case. Especially the ones you haven't dreamed of yet. You'd wind up rewriting the rules every season, to fit whatever engines you wanted to include anyway.

If new technologies do appear then the rules would have to be rewritten anyways to accommodate the new technologies. We also aren't talking about handling every case. We are just talking about handling two cases, the NNUE network architecture and the Leela network architecture. And I'm arguing that the two library-esque network architectures are similar enough in function that the rules that apply to one should apply to the other.

So what I guess what I'm saying is, including Allie is more fun than having two engines compete for two SuFi spots? And I'm not sure there's any other question to ask.

Then why have uniqueness rules in the first place? Why doesn't TCEC do what CCCC does and say, 'here are the engines that are playing this season and we as the organisers have the right to choose which engines get to participate in the tournament and which engines do not, to ensure maximum fun in the tournament'.

[dkappe]

^^ Before the NNUE hype started, many thought that small nets wouldn't work, but now that can be reconsidered.

That being said, I think the new TCEC rules require the data used to be from the engine itself, so if TCEC applied its rules fairly to all engines, Allie would have to use a net trained using Allie data, and a0lite would have to use a net trained using a0lite data.

There's however another option that makes sense from the aggregate viewer satisfaction standpoint: to persuade Dietrich to withdraw a0lite from QL in exchange for credit as an Allie co-author if Allie does manage to benefit from a small net.

So Allie can play GM strength (weak) on TCEC hardware, and I can have my name as author on an engine I had no hand in writing?

[Madeleine Birchfield]

Why would Allie want to use a tiny 16x2 network or a 128x10 network in TCEC? It makes no sense. I just wanted to point out that not all networks — like the Gyal and Ender networks — are leela networks. Some of them won’t even run in lc0.

Are there any networks that would run on Allie and wouldn't run on Leela? And are there any networks that would run on Leela but wouldn't run on Allie?

[dkappe]

Why would Allie want to use a tiny 16x2 network or a 128x10 network in TCEC? It makes no sense. I just wanted to point out that not all networks — like the Gyal and Ender networks — are leela networks. Some of them won’t even run in lc0.

Are there any networks that would run on Allie and wouldn't run on Leela? And are there any networks that would run on Leela but wouldn't run on Allie?

Don’t know. Scorpio will run just about anything. You should talk to the authors of the engines about their capabilities.

[Tony P.]

So Allie can play GM strength (weak) on TCEC hardware, and I can have my name as author on an engine I had no hand in writing?

I assumed that Allie's search would be heavily rewritten. With the current search, it would indeed not benefit from running the Gyals at TCEC.

I'd also assume that an AB engine would be better off using a graph NN than a 2d CNN, but having seen how Leela beat the classical SF, I'm prepared for surprises.