2 open projects

[JJJ]

I think Leela will overcome his weakness with tactics Uri. It's just his way to learn at first. She might even improve against missed mate in 1 or 2 and so on.

10 millions game is just 25% of alphazero game played to train.

And anyway Leela might play 1 billions games if needed with bigger and bigger net. So let's see !

[Milos]

I think Leela will overcome his weakness with tactics Uri. It's just his way to learn at first. She might even improve against missed mate in 1 or 2 and so on.

10 millions game is just 25% of alphazero game played to train.

And anyway Leela might play 1 billions games if needed with bigger and bigger net. So let's see !

A0 after 10 millions training games already (allegedly because there is no proof beside that Google advertising leaflet) surpassed SF8.
You know where is LC0 at 800'000 playouts a minute - 400 Elo below SF8 on 64 cores and 1min/move.
You really think that increasing network size indefinitely will continue to yield strength improvements???
Sorry, but that's not how things work.

[jkiliani]

I think Leela will overcome his weakness with tactics Uri. It's just his way to learn at first. She might even improve against missed mate in 1 or 2 and so on.

10 millions game is just 25% of alphazero game played to train.

And anyway Leela might play 1 billions games if needed with bigger and bigger net. So let's see !

A0 after 10 millions training games already (allegedly because there is no proof beside that Google advertising leaflet) surpassed SF8.
You know where is LC0 at 800'000 playouts a minute - 400 Elo below SF8 on 64 cores and 1min/move.
You really think that increasing network size indefinitely will continue to yield strength improvements???
Sorry, but that's not how things work.

As far as I'm concerned that's EXACTLY how things are going to work. The policy improvement operator from MCTS works whether the improvement is in positional evaluation or in tactics, it just can't improve a net beyond its capacity. Raise the capacity, and the self-improvement resumes. If you don't believe that, just wait for let's say three months (next TCEC), and then we can discuss again whether increasing the net size works after all.

[velmarin]

Perhaps at some point a hybrid will be born, Eval, search and Syzygy from Stockfish with a neural network.
Or is it not possible?
What a performance it can have.

[noobpwnftw]

There are basically two categories of people riding the hype:
1. The new approach is so different and they found it interesting and worth a try, I'm neutral to them and I believe the strength of NN engines will reach top-tier because GPUs are relatively cheaper as in performance per buck.

2. Some people feel motivated when they finally get a chance to believe that their incapability of coding anything may turns out to be a profit, for those I enjoy turning their fantasy into a troll thread.

[velmarin]

troll

Here's this awful adjective again.

This is a forum, let's be nice,

[noobpwnftw]

awful forum

Nice quote.

[JJJ]

I think Leela will overcome his weakness with tactics Uri. It's just his way to learn at first. She might even improve against missed mate in 1 or 2 and so on.

10 millions game is just 25% of alphazero game played to train.

And anyway Leela might play 1 billions games if needed with bigger and bigger net. So let's see !

A0 after 10 millions training games already (allegedly because there is no proof beside that Google advertising leaflet) surpassed SF8.
You know where is LC0 at 800'000 playouts a minute - 400 Elo below SF8 on 64 cores and 1min/move.
You really think that increasing network size indefinitely will continue to yield strength improvements???
Sorry, but that's not how things work.

You know milos, you should pay attention about what you re writing. If you haven't noticed, you had never posted a single positive comment in this forum but the opposite. Negativity is gonna make you real sick somedays, perhaps you already are.

[Milos]

As far as I'm concerned that's EXACTLY how things are going to work. The policy improvement operator from MCTS works whether the improvement is in positional evaluation or in tactics, it just can't improve a net beyond its capacity. Raise the capacity, and the self-improvement resumes. If you don't believe that, just wait for let's say three months (next TCEC), and then we can discuss again whether increasing the net size works after all.

Whether further net size increase is gonna work short term or not is something that indeed we don't know and will need some time to see.
That there exists a point beyond which any further simple net size increase (such as increase in number of blocks and filters) doesn't bring anything is such an obvious property on NNs in general and more particular DNNs that claiming otherwise simply proves one doesn't know much about DNNs in general.
Based on your other claims about net so far I would say it is pretty obvious that your knowledge and experience with DNNs in general is very limited.
For your field of interest for example there is ton of indications that NNs offer subpar performance compared to other simpler ML techniques, so it wouldn't be much of a surprise that your DNN expertise is lacking .

[Milos]

You know milos, you should pay attention about what you re writing. If you haven't noticed, you had never posted a single positive comment in this forum but the opposite. Negativity is gonna make you real sick somedays, perhaps you already are.

You must be some psychologist then, regarding your "deep" insights. Well this is computer chess forum, so you better check you might be in a wrong forum my friend.
Your ad hominem attempt is pretty lame considering you can't offer any actual argument regarding the topic.