If the hardware were 1000x faster, how much time would #AlphaZero need to reach superhuman levels (starting again from scratch)?
Announcing #AlphaZero - our general RL system that starts from random and learns thru self-play to beat the top chess, shogi & go programs! Preliminary arxiv paper: https://arxiv.org/pdf/1712.01815.pdf … Games here: https://goo.gl/UEHcWz More details in full peer-reviewed paper coming soon.
-
-
-
It took 6 hours or so of wallclock time on 5000 TPUs. With 700 GHz machines (1000× faster), maybe 22 seconds?
- Još 5 drugih odgovora
Novi razgovor -
-
-
Tweet je nedostupan.
-
Games it lost were not started from beginning. They were tarted from fixed position (~bad position)
- Još 1 odgovor
-
-
-
Thanks for confirming English opening supremacy
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
So if each item of 4096 minibatch is an average over 800 'simulations', and you do 500K minibatch steps, then it has played 1.6 trillion games? if so is that about $50,000 of cloud training or what?
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
How did you decide the settings Stockfish used and what do you say to the notion that you handicapped it to ensure a more clear victory?
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
That's amazing. Can you ask it to figure out poker sometime? Thanks in advance.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Just outstanding playing by Alpha zero. Can we see some match against all best chess players in the world playing together and consulting to any top chess engine any point of time against Alpha Zero. It would be game like God against rest of the world.
-
They tried that with the game of Go : AlphaGo vs a team of 5.https://techcrunch.com/2017/05/26/googles-alphago-ai-defeats-team-of-five-leading-go-players/ …
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.