Author Topic: The flaw in AlphaZero  (Read 356 times)

Arjo Schreuders

  • Newbie
  • *
  • Posts: 12
    • View Profile
The flaw in AlphaZero
« on: February 18, 2019, 11:37:12 AM »
I was kinda surprised there is no topic on AlphaZero on this forum...
So let me start one by sharing what I noticed by playing trough the games it lost against Stockfish...

In (almost) all the games it lost cause it refused a draw by repetition in a totally drawn position...
Guess this is because it is programmed to obtimize its winning chances...
(probably it wins more games than it loses by this)
And taking a repetition ofcourse gives you 0% winning chances...
So it takes an inferior move instead...

Stephen Ham

  • Newbie
  • *
  • Posts: 4
    • View Profile
Re: The flaw in AlphaZero
« Reply #1 on: June 09, 2019, 02:44:12 AM »
Hi Arjo,

It seems you're speculating without knowing the facts.

A0 and fellow neural network engine, Leela, are superior to the latest Stockfish iterations in overall match play. But, while they manifest superior strategic skills, they're notoriously weak tactically, and even worse in endgames. Some endgames are played at a nearly childish level. So, this endgame weakness surely explains poor decision making in drawish endgames.

Engines never optimize winning chances by playing inferior moves and making poor decisions. Neural network (NN) engines instead make their decisions based upon prior play against themselves. But, if they lack endgame technique, then their prior play will be riddled with errors. And these erroneous endgames are the basis upon which subsequent decisions are based.

At present, the best chess engine would be Stockfish with Alpha Beta heuristics in the opening and endgame, and A0/Leela (NN engines) in the middlegame. To date, that hybrid has yet to be created...although it surely will with time.

Sincerely,
-Steve-
« Last Edit: June 09, 2019, 02:48:48 AM by Stephen Ham »

Arjo Schreuders

  • Newbie
  • *
  • Posts: 12
    • View Profile
Re: The flaw in AlphaZero
« Reply #2 on: June 09, 2019, 07:48:31 AM »
Its not a speculation... Just an observation...
After repeating a position twice it makes a different move which seems to be 0.00 (or 50% winning chances) too,
but loses in the long run...

About your SF/A0 hybrid... It wont be as strong as you think...
It will take double the coding and so will be (very) slow...

Arjo Schreuders

  • Newbie
  • *
  • Posts: 12
    • View Profile
Re: The flaw in AlphaZero
« Reply #3 on: June 10, 2019, 02:46:38 PM »

A0 and fellow neural network engine, Leela, are superior to the latest Stockfish iterations in overall match play. But, while they manifest superior strategic skills, they're notoriously weak tactically, and even worse in endgames. Some endgames are played at a nearly childish level. So, this endgame weakness surely explains poor decision making in drawish endgames.

Engines never optimize winning chances by playing inferior moves and making poor decisions. Neural network (NN) engines instead make their decisions based upon prior play against themselves. But, if they lack endgame technique, then their prior play will be riddled with errors. And these erroneous endgames are the basis upon which subsequent decisions are based.



I wasn't talking about endgames in my post...
Most are middlegame positions...

Stephen Ham

  • Newbie
  • *
  • Posts: 4
    • View Profile
Re: The flaw in AlphaZero
« Reply #4 on: June 30, 2019, 05:18:48 AM »
About your SF/A0 hybrid... It wont be as strong as you think...
It will take double the coding and so will be (very) slow...

You continue to misunderstand NN engines. They have no heuristic coding. Their only code involves the rules of chess. After that, they simply play millions of games. AI is then utilized to "learn" to play well.

But even IF double coding exists, which it obviously doesn't, you'd still be wrong because only one engine would be operational at any given time. As already stated, it's Stockfish in the opening and endgame and Leela 0 in the middle game.

-Steve-
« Last Edit: June 30, 2019, 05:24:46 AM by Stephen Ham »

Arjo Schreuders

  • Newbie
  • *
  • Posts: 12
    • View Profile
Re: The flaw in AlphaZero
« Reply #5 on: June 30, 2019, 06:54:50 AM »
"You continue to misunderstand NN engines. They have no heuristic coding. Their only code involves the rules of chess. After that, they simply play millions of games. AI is then utilized to "learn" to play well."

And based on those games they write their own code....
Seems like you are the one misinformed...
But I will stop this silly discussion...

Arild Haugen

  • Newbie
  • *
  • Posts: 8
    • View Profile
Re: The flaw in AlphaZero
« Reply #6 on: July 17, 2019, 02:47:49 PM »
I think endgame technique can be improved by training with Tablebases.