thanks petri999. online cheating using engines is a problem in poker, too.
@HerkyHawkeye said in #18:
> Is it formatted for ipad?
It should work on an iPad, I recently added some fixes that target mobile platforms including iPhone and iPad. I don't own these devices so I can't test them out, if you give it a try please report back on your experience.
@petri999 said in #19:
> stocahastic game cannot be solved in sense checkers tic-tac-toe have been.
>
> Best engine is stronger than human. in 24 pt match expert backgammon would be about 2-1 favorite to win. Gnu Backgammon might be bit less favorite but would be favorite.
>
> I do doubt if engines can get better. Game is not as deep as Chess or Go and added with fact that dice plays a role even on long match. Even objectively better moves would give marginal gains anymore as most of moves are already objectively the best. Improvement since late -90's is very small.
>
> Due to engines and popularity of poker there are no longer sites where you can play backgammon for real money.
While researching existing backgammon engines while creating tabula (linked below) I got the sense that there is still potential for innovation in this space. Technology has come a long way recently, particularly in terms of the capabilities of neural networks, and so I think there are a lot of opportunities for engine authors to explore new and better ways of analyzing positions.
code.rocket9labs.com/tslocum/tabula
> Is it formatted for ipad?
It should work on an iPad, I recently added some fixes that target mobile platforms including iPhone and iPad. I don't own these devices so I can't test them out, if you give it a try please report back on your experience.
@petri999 said in #19:
> stocahastic game cannot be solved in sense checkers tic-tac-toe have been.
>
> Best engine is stronger than human. in 24 pt match expert backgammon would be about 2-1 favorite to win. Gnu Backgammon might be bit less favorite but would be favorite.
>
> I do doubt if engines can get better. Game is not as deep as Chess or Go and added with fact that dice plays a role even on long match. Even objectively better moves would give marginal gains anymore as most of moves are already objectively the best. Improvement since late -90's is very small.
>
> Due to engines and popularity of poker there are no longer sites where you can play backgammon for real money.
While researching existing backgammon engines while creating tabula (linked below) I got the sense that there is still potential for innovation in this space. Technology has come a long way recently, particularly in terms of the capabilities of neural networks, and so I think there are a lot of opportunities for engine authors to explore new and better ways of analyzing positions.
code.rocket9labs.com/tslocum/tabula
I checked just me .md file i was just wondering why you have some adjusted pip count? Properly trained simple one hidden layer NN will form something like that all by itself when trained. But programming your own bot is definitely good way to learn about backgammon and programming. But as I said scope for improvement is small. Even if you come with engine that creates better moves every now then gain is going to be small. Like understanding best opening roll moves has improved but in cases judgement what is best has switched gain is always less than1/100 of a point. And relatively few moves in the game adn take away moves that best move obvious you don't get many rolls where this extra 1/100 point can be gained. As for doubling it is a pure maths problem which solved quite some time a go. As long as Pwin/PwinGammon/PwinBackgammon provided by the NN are in the correct ball park there is no room any improvement. In particular as engines now search 4 ply for evaluation of cubing decision.
To put scale in when I played at fibs my rating was 1600 and top players were 2000 as were the engines. In in late -90s engines did not yet have power to do a forward search so they were not as strong as now. Using the FIBS rating formula for expected score for the weaker player that gives me whopping 20% chance a winning a 9 point match against top player. And I have actually won GnuBG with 7 points match so not really outlandish. Back then probably could have got 20% success rate obviously top player were not playing that much weaker players. I suck in backgammon just as much as in chess. I don't have one in 1/5 chance against Elo 2000 players let alone people who play professionally.
To put scale in when I played at fibs my rating was 1600 and top players were 2000 as were the engines. In in late -90s engines did not yet have power to do a forward search so they were not as strong as now. Using the FIBS rating formula for expected score for the weaker player that gives me whopping 20% chance a winning a 9 point match against top player. And I have actually won GnuBG with 7 points match so not really outlandish. Back then probably could have got 20% success rate obviously top player were not playing that much weaker players. I suck in backgammon just as much as in chess. I don't have one in 1/5 chance against Elo 2000 players let alone people who play professionally.
I explain my reasoning for creating the engine in more detail in the blog post linked below, but in summary it's a combination of needing an engine written in Go due to cross-compilation requirements (the client targets four platforms) and an engine that is explicitly *not* trying to be the best, but to be the most fun to play against foremost, while also being as challenging as possible without requiring more than a few seconds to calculate the next move on a low-powered Android device.
bgammon.org/blog/20240101-hello-world/
Neural networks are an option, and would be powerful with some training (the wildbg bot on bgammon.org is powered by NNs, linked below), but the focus then becomes training the NNs sufficiently in a correct manner. So it is just a personal preference of mine to not use NNs. This way the code complexity stays manageable, and each chosen move can be analyzed and understood from start to finish.
github.com/carsten-wenderdel/wildbg
bgammon.org/blog/20240101-hello-world/
Neural networks are an option, and would be powerful with some training (the wildbg bot on bgammon.org is powered by NNs, linked below), but the focus then becomes training the NNs sufficiently in a correct manner. So it is just a personal preference of mine to not use NNs. This way the code complexity stays manageable, and each chosen move can be analyzed and understood from start to finish.
github.com/carsten-wenderdel/wildbg
This topic has been archived and can no longer be replied to.