| |
BGonline.org Forums
what would you like 2c in XG3?
Posted By: Simon Knight
Date: Friday, 6 January 2012, at 11:13 a.m.
The title says it all – this thread is for anyone who wants to see the world’s best bot grow even better in the future. I have some ideas and maybe you do too, in which case I hope you’ll share them here.
1. Learning Lab
Since BG is a game of pattern recognition, the trick to getting better is to repeatedly see those patterns you don't understand, in a way which keeps you involved and interested. Dissecting the corpse of a dead game, wincing at the brainless horrors XG accuses you of in angry crimson and reproving green, is a slow and demoralising way to improve, and reminds me of the way I was taught Latin as a schoolboy. Instead of pointing out failure, how about rewarding success, like this:
XG Learning Lab stores all your bad plays, both with checkers and the cube, in its own special database. Your task is now to clear them from its memory by getting them right next time. Whenever you decide to go into Lab Mode, XGL calls up the positions randomly one by one, and asks you for your move or your decision. Get it correct, and XGL chalks you up a brownie-point. Get it wrong again, and you can be sure the position will appear a lot in future. Once you have finally made a good choice three times running (or however many times you specify), over whatever timescale it might take you to achieve this (minutes for an expert, months for a beginner) then your misplay is considered rectified and erased.
This would become an addictive game in its own right – purge yourself of all your past backgammon sins and a victory-screen appears, along with a small synthetic fanfare... : - ) Additionally, XGLL would score your results in both PR and elo, and also graph your progress over the long-term, just the way it already does with ordinary games.
2. Openings Trainer
XG2 already contains the most definitive openings database in existence. At the moment it is purely an exercise in geekery, its contents known only to a few high priests of the game. However the information could easily be disseminated in a way that is educational and useful. The principle is exactly the same as for the Learning Lab, namely guidance through stimulus and reward, but here the aim is to play all the opening rolls and replies perfectly, either for money-play or tournament or both.
The program will soon believe you know how to play an opening 3:1, and will subsequently never ask you this again, but 6:4s and 2:1s et al at different match-scores, and your best reply when you roll second, would again take anything from minutes to months to master, depending on your standard.
3. CubeSky
Extreme Gammon knows everything there is to know about the cube, yet teaches nothing on the subject whatsoever. Piles of numbers to two decimal places are not a form of enlightenment. Imagine instead the market window as the archetypal child’s window, a square with four equal panes. Through the window you see rain below the horizontal crossbar and blue sky above it. Plotted along the skyline is the past course of your game, with the move currently under question being the sun.
Say it’s the first roll of the game, so the sun is still on the far left of the window. You’re down 7-away 2-away, so the sun has appeared far higher in the sky than usual. Three rolls in and you’ve had nice dice, your opponent hasn’t. Maybe the sun is already on the window-bar, neither stuck in the rain nor overshooting into the clear blue sky, making this a perfect double. But now click the mouse and CubeSky will auto-scroll through all other scores, as quickly or as slowly as you like. Had you been up 7/2 instead of down, for example, the sun would barely scrape above the windowsill, making that same perfect double a hideous blunder, with the sun exploding into supernova in the blackest part of the raincloud.
This extreme example would only help a beginner, obviously, but it would help him enormously. Going further up the rankings, surely any player would love to know whether, for instance, the mistake they made when one point down would have been correct at level-score, or if the cube had still been in the centre. Regardless of experience and ability, who wouldn’t want to see striking visual proof of the cube’s dynamics, alongside the numerical evidence?
It’s hard to describe a visual aid without the use of a visual aid to describe it, but the horizontal beam across the middle of the window represents the right time to offer a double, and this beam magically gets fatter or thinner according to the early/late cost ratio associated with the score and current level of the cube. The state of the weather shows whether it's a pass or a take.
This idea would only be of moderate value to moneyplayers, but would really come into its own for those of us who prefer matches. And if you want to make CubeSky a learning game as well, plug in your own estimates of winning and gammon chances of past correct decisions taken by XG, and see how the vista changes before it reveals the true picture; the better your guesses, the higher your score.
People learn through sights and sounds, not stats. If you agree, and would like to see XG develop into a fabulous teaching aid, then you need to say so or it will never happen. Xavier is brilliant at what he does, is receptive to ideas and listens to his customers, but all that only works if people tell him what they want.
--
When I posted this on 2+2 the other day, I was advised to place it here instead. One of the responders to that thread then mentioned this…
- - - “Let’s Check” is a revolutionary new feature of Fritz 13 that will change the chess world. With it Fritz 13 users can join a world-wide community that will put together a giant knowledge base for chess... - - -
… and said he would like backgammon to have something similar. Geeky mass rollouts are of no interest to me whatever, but the concept offers something infinitely more valuable: it would give XG access to the wisdom of crowds – or make that the wisdom of clouds**, in this case.
Imagine a beta of the Learning Lab in which a group of good players all categorise their own mistakes, firstly into broad classes: disconnection, cowardice, blindness, over-ambition, etc., and then into more specific detail: blocking when you should have run, running when you should have blocked, racing when you’re behind, ditto vice versa, overlooking the faraway blot, etc. Six main categories and eighteen sub-categories should cover just about everything, plus a ‘miscellaneous’ tab for the remainder.
Once XG has a few thousand of these under its belt to analyse, it can start giving your own mistakes the appropriate tags automatically. It can then compare your error rates in the various categories against the global average made by players with a similar rating to yours. And this, in turn, would enable it to generate a written report on your play. ‘You are exceptionally good at taking worthwhile risks to make home-points, but you tend to make passive low anchors needlessly…’ etc. Armed with all this, it can summon from its cloud precisely the kind of positions you most often get wrong, and test you on them until you scream for mercy. This might sound pie-in-the-sky, but turning stats into English isn’t actually that hard – the program simply translates figures, say 15% over average, into phrases, i.e. ‘exceptionally good’ , ‘buy new glasses’ etc.
** http://en.wikipedia.org/wiki/The_Wisdom_of_Crowds
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.