OMFG! STOCKFISH NOOO! ( AlphaZero vs Stockfish ) Thoughts

Imagine waking up this morning ,

Having your childhood hero who you view as the perfect role model

get forced from behind balls DEEP .

I don't like it one bit
Feelsbad #Pray4Stockfish

i wanna have zero

it play's like tall and many more great players i really like say ferwell drawsssss in chess

We have seen in other threads that the comparision was not fair. While Stockfish used 64 cores and no opening book and no endgame tablebases, Alpa Zero used something equivalent to 400 to 2000 cores and something equivalent to an opening book. Also, they didnt use the latest development version of Stockfish but an older Version.

Some said even Fritz could beat Stockfish under such circumstances and further that Alpha Zero is probably equally strong.

Some said that this just proves how optimized modern chess engines are, as Alpha Zero beated the engines in the other games much clearer.

Still it is cool that an Algorithm can learn how to play chess like this in just four hours, even if they used Power comparable to 500.000 to 1.250.000 cores to train Alpha Zero (im quoting a dude here who said they used 5000 TPUs and one TPU is said to be comparable to 100 - 250 modern cores, depending on how you count, correct me if i am wrong here).

For comparison, my laptop has four (outdated) cores.

Sources:


groups.google.com/forum/?fromgroups=#!topic/fishcooking/pcFRIurN_l4

It's NOT fair unless you use similar hardware and similar things like end-game databases. have Alpha0 play on even terms.

lets see the potential here, this was the first test in Chess for a learning Neuronal network. Remember, no human knowledge was given to Alpha Zero, it learned this Stuff on its own. From a Scientific Point of view this is quite exciting.

I dont know how far this goes but this in general means that in a few generations computers will do all the work for us. We will then just have hobbies and do what we want.

Thats quite a cool vision, all we have to take care of is that there will always be a panic power-off button, which the computer can not prevent us to press.

Imagine.

The playing field was far from equal. Makes one wonder why SF developers accepted the challenge. A large $ kickback perhaps. With no table bases to rely on, SF, Komodo, Houdini are just "blunder checkers" anyway. The result was easily predicted. The "neural computing" however is exciting.

The thing is, DM and normal engines evaluation is different. Deep mind doesn't need to bruteforce positions as much to achieve an advantage, hence why we have 70,000,000 positions for SF versus 80,000 from DM, and won't really matter hardware or hashsize since DM's genetic algorithm is only layered on learning phase, not when playing. And should be considered that DM had only four hours of training under 8 TPU's, it could've trained for days. The four hours of training were deemed "enough" to go against all the database SF8 have been collecting since its conception. There is no comparison between them really. SF8 can be improved as much as they can, its going to be a losing battle.

*4 hours of training under 5064 TPUs
That's millions of dollars of hardware.

arxiv.org/pdf/1712.01815.pdf Fascinating stuff, FAAAAAAAAR from a level playing field though, but that's not really the point!