Twixtbot (c) analysis TWIXT PP
1 replies. Last post: 2019-11-19Reply to this topic Return to forum
MisterCat at 2019-11-19
c) I have questions – more like doubts, as to whether the current version of Twixtbot can prove to be a useful tool for analysis and improvement in human games. I briefly discussed this over at the commentator: Twixtbot is playing statistically, and humans can not play that way. Humans play strategically. David Bush disagreed at the time, but here’s more of my meaning: Twixtbot acquired it’s neural net by playing millions of games against itself, and statistically evaluating which moves win more often in each position. This is the brilliant idea fostered by the creators of Alpha Zero. Brilliant, it is, I say – this type of analysis may, one day, allow computers to cure diseases, stabilize economies, produce world peace, and so on. It is, indeed, profound, in it’s potential.But no human being will EVER be able to ‘think’ that way, since a statistical analysis of millions of games is not possible by a human brain. A human must evaluate positions – look ahead a few moves at a time, analyze STRATEGIES, like weakness in corners, setups, etc. The reason analysis of Twixtbot’s games will fail, in my opinion, is because of ‘the butterfly effect’: just ONE peg in a different square at the corner of the board might not look important in short term, strategical analysis; however, it makes all the difference 20 moves down the road. This is also why I don’t think it useful to catalog ‘best’ opening moves, based on Twixtbot. From what I’ve been seeing, the bot can make seemingly THE WORST OPENING MOVES IN THE WORLD, and STILL win!I think that in Chess, the top grandmasters who HAVE been using Stockfish, Fritz, Komodo etc. as analysis tools will TRY to use Leela Zero, and eventually realize that they can not learn from this – for the same reason.