Seeing how this 'Twixtbot (a) thread is short, I've decided to post my thoughts on the recent 'Twixtbot vs. The World' experiment here. As my first post above divided my subsequent posts into the topics:
a) Twixtbot playing in tournaments
b) Twixtbot junior
c) Using Twixtbot for analysis
d) Cheating in Twixt
e) How Bony Jordan can make money from this
I will point out that the original subtitle for this post, which I neglected to include in the topic line - ('Twixtbot playing tournaments') is in line with some of what I have to say.
My thanks to all who participated in the experiment, and these players have been cited in another post. In as much as it was 'fun' to view this game as 'pride of humanity' vs. 'the advent of machines', and it WAS FUN, it really was not my intention. I have learned enough about Alpha Zero programming (reading articles, interviews, YouTube posts, etc.) to get a handle on the type of analysis and decision making being used. There is really NO strategy - no techniques - no 'learning' being applied in these games; rather, it is NOTHING BUT A MATTER OF STATISTICAL TABULATION. Play 50,000, or 2,000,000, or whatever sample games (moving initially at random), and see which moves result in wins; build a neural net based on those moves that WIN MOST OF THE TIME, which can continue to improve as more results are tabulated; and in live games, continue the tabulation by playing out moves in the actual game to enhance the neural net, and more tabulation. (Correct me if I am not accurate.)
There is nothing wrong with this; in fact, I'm somewhat IN AWE at this development in computer science, and I do believe that this type of programming can have far reaching consequences - hopefully good ones, but like all tools for good, one can envision it being perverted to evil uses. But for the good, imagine benefits for world economics, ecology, medicine, space exploration, etc. that might come out of this approach!
However, my feeling is that the 'Alpha Zero' approach to playing strategy games is NOT LIKELY to lead to improvements in the way HUMANS play these games. Humans can not internalize databases of millions of games, or tabulate billions of results to decide on moves in a game. We do NOT work that way, CAN NOT work that way, and WILL NOT EVER work that way. (Commander Data wins a strategy game on an episode of Star Trek TNG - but then again, Data is NOT human. Data wins at Stragegema ) Humans STRATEGIZE; they learn 'rules of thumb' in a game - what makes good positions, what are bad approaches to take; how to look immediately ahead for a few moves; no human internalizes millions of positions and tabulates them.
So Twixtbot is basically playing statistically. That the approach wins games can be seen from it's record, but what occurred to me that it still is not actually playing 'positionally' or 'strategically'. How does it win, then - in a game with as much complexity as Twixt (or Chess, Go, and most others), it is only a matter of TIME, statistically speaking, before a human player will play less-then-optimal moves. These moves might still allow that human player to win games against another human player, since neither of them can fathom it's weakness; only the machine can see that after 50,000 games, the moves loses, say, 42% of the time. After a string of less-optimal-moves are played, the machine enters territory where it is spotting moves which win, say, 95% of the time, so plays them. The human is lost.
But these 'less-than-optimal' moves ARE mistakes, of a sort. They are too subtle for most human players to notice; not without playing through thousands of variations, and what human has the time or energy for that. But mistakes, they are - so without the machine, CAN HUMANS AVOID THESE MISTAKES BY EMPLOYING A TEAM APPROACH? That was the experiment. Individually, none of the players on The World team can boast a good record of playing against Twixtbot, but NOW, the moves are being DISCUSSED, CHECKED, RE-CHECKED, and ANALYZED, using HUMAN UNDERSTANDING of the game. No tactical blunders will be made. No unwise decisions will pass scrutiny. The Team worked with a goal - analyze and capitalize on the weakest part of Twixtbot's statistical approach, and plan accordingly, using knowledge of the game, strategy, positional analysis, and limited looking ahead.
IT WORKED! WE WON! Frankly, I had my doubts going in - based on what I had seen, I would not have been surprised at all by a Twixtbot victory. I had ONE editorial planned in case we lost, another editorial planned if we won (this one), and I have no idea what I would have said in case of a draw. But we won, so here we are.
Should Twixtbot compete in rated play and the championship? Well, NOW, I think that things are different. Twixtbot CAN be beaten. We have proved that. It is not at all easy - but winning a championship is never 'easy', anyway. Opinions above are leaning towards saying that TB should be removed from all rated play, because it can not be defeated. This has changed. For me, personally, at my level of play there is NO WAY I would ever be able to play with the degree of care and profundity that is required to win such a game. But perhaps the top players here would rather be put to the test?! So my feeling about the championship NOW is that it really should be left to those players - perhaps in a championship-by-championship basis: TB CAN BE BEATEN. Can you do it? If you do, then it means even more if this occurs on the way to winning the Twixt Championship. However, beatable as it may be, TB remains a bot; some players simply do not with to test their skills against bots, ever. (on the other hand - how many players of reasonable strength enjoyed beating the tar out of the old TB, before the upgrade? The old TB made a nice 'punching bag, and I didn't see many objections to it's presence - perhaps because it was easy and fun to beat????)
Things have changed somewhat in this past week, now that bonyjordan has upgraded again to TwixtbotSlow. There may be some clamour to retry the experiment against this new 'monster' - well, anyone who WANTS to do this, CAN do this. I won't be running another experiment personally, as I believe that I have already 'proved my point'. What point is that, you may ask? Well, it is that the 'statistical' approach may very well win games, but is not useful as a learning or training tool, and actually is not even playing all that well! Humans can defeat the machine by using normal, human analysis, as long as sufficient effort is made so as to not blunder the game away with one or more poor moves. Furthermore - based on this, I would very much like to see if a team of the best Chess players in the world - working with analysis, time, databases, opening books, etc. can defeat the best Chess engines - you know, Stockfish, Leela Zero, etc. These engines are all rated over 3200, where Magnus Carlsen is a measly 2800, so generally NOBODY stages human vs. computer matches anymore. Even the best human players come right out and say they will be crushed! But I am not aware of what I am proposing EVER BEING TRIED! If anybody knows of such a case, do tell. Let us see how ambitious I am - perhaps I will describe the results of our 'Twixtbot vs. The World' experiment to the 'powers that be' in the Chess world, and see if anybody wants to test it out. Again, my feeling is that 'positional' and 'strategic' chess moves, when applied WITHOUT ERROR, will defeat the mathematical approach taken by the machines. I may post this remark over at the Chess Forum to see if anyone can further this discussion.
(meow)
Little Golem