Hacker News Reader

Comment by dragontamer | Hacker News Reader

AlphaGo engineers didn't really know how to configure Stockfish correctly.

The computer itself was strong. But the weird timing setup, the disabled databases (ie: both opening book AND endgame book was disabled. Stockfish normally plays PERFECTLY when the board is reduced to 6 pieces or less, as well as perfectly knows the winner / loser in every 6-piece setup. But that was disabled for the AlphaGo games)

Without the ability to run AlphaGo on our own and recreate the test, we have no way in knowing how AlphaGo would work under "fair" conditions.

-------

And btw: 1GB of RAM is a lulzy setup. You put all the CPU time you want, but gimping the RAM down to 1GB is... weird. Its a VERY suspect "test" that none of us can replicate.

View on HN

It's not that they didn't know how to configure it correctly; some of it was purposefully (and rightfully) disabled to see real difference between stockfish and AZ.

Remember that if stockfish uses tablebases and opening books it's really "cheating" by using human knowledge.

Although I am surprised that they reduced the hashtable so much and also used stockfish 8 when stockfish 9 was available. It was not RAM but just the hashtable size also their "test" has been replicated by plenty of people.

We have our own "replica" of the NN playing chess in lichess.

Opening book I agree, but tablebases are definitely not human knowledge. They are a bruteforced database of every possible position, or in essence a precomputed position. Put simply, giving a massive time advantage to the user during the late middle and endgame.

So if A0 wasn't built to access them, but stockfish did in the match, then that would be an unfair advantage.

> So if A0 wasn't built to access them, but stockfish did in the match, then that would be an unfair advantage.

Why is that an unfair advantage? AlphaZero developers wanted to prove that their AI was better than Stockfish.

They failed to do that. If AlphaZero wasn't built to access Tablebases, then they should have built it to access tablebases. Don't unfairly gimp Stockfish because you're lazy at programming.

Alpha zero can access the tablebase. This proof of concept was to let both the engines play with minimal interference of outside pre computed knowledge.

The programs are different programs. You can try it both ways, with and without the opening books. Because there are different configurations, in order to see who win must be specified by the configuration in use.

One way is to do according to the most capable configurations of each program, so that means Stockfish with all of them. Other way is to specify the same limits of time, memory, etc of each one, although that is still not the excuse to avoid the opening books and tablebases and so on unless they use up more memory than AlphaZero does.

> Remember that if stockfish uses tablebases and opening books it's really "cheating" by using human knowledge.

I thought the "point" of AlphaZero was that it was better than the sum of human knowledge.

By the way: tablebases are computer generated. There is no human alive who knows all combinations of 6-piece endgames. Tablebases are brute-forced endgames, created by computers to be used by computers.

Stockfish itself was only really designed to play the midgame. Its designed to be used with both tablebases and opening databases.

> It was not RAM but just the hashtable size also their "test" has been replicated by plenty of people.

Citations please. I'd love to see a "proper" test of AlphaZero vs Stockfish.

You can estimate pretty well even on average hardware by keeping the same 875x ratio using time odds. However, what is a fair comparison to either side depends on who you ask.

Reply on news.ycombinator.com