Even more on AlphaZero

Apparently AlphaZero has a predecessor named Giraffe. The developer Matthew Lai now works for DeepMind. Surprise?

The paper on Giraffe explains everything in much greater detail. For instance it took 72 hours to calibrate the net on a workstation with 2×10-core Intel Xeon E5-2660 v2 CPU. If we use a rough shortcut and divide 72 by 4, then Google’s high-end cluster with 64 TPUs did the job 18 times faster.

In 2016 Giraffe peaked at Elo 2410 in engine competitions, which is remarkably weak as even the current version of the good old GNU Chess is rated Elo 2800+. Given the initial results, the decision to keep investing in the idea is quite remarkable too.

As I wrote in the previous article, my Stockfish, casually running on just one core on an Intel i7 4760 3.60 GHz, took roughly 75 minutes to find the star move at depth 41. Hardware is the bottleneck. Just for comparison: Massive hardware upgrades almost doubled the playing-strength of AlphaGo. It simply expands the search-horizon.

Looking at the difference in pure hardware power this reminds me of David vs. Goliath. Running on identical machines, Stockfish should beat AlphaZero easily. Drawing 72 games with such a handicap is actually amazing. Let’s not kid ourselves, Stockfish on 4 TPUs would beat Stockfish running on much weaker hardware convincingly too.

One thing is clear: It will take a few CPU-generations until mere mortals like you and me will be able to run AlphaZero at home.


Update (21.12.2018): The results by Leela are so encouraging, there may not be a need to run AlphaZero at home anymore.