| |
BGonline.org Forums
Neural net, data, code, bug, etc. - Is XG violating GNUBG GPL?
Posted By: Jack Mack In Response To: Is XG violating GNUBG GPL? (niuniuch)
Date: Tuesday, 21 March 2017, at 8:42 a.m.
"Looks like you ignored my post in the other branch, so let me reply here."
I didn't mean to ignore you but posting several long responses in one night against a mob is tiring in every meaning but let me apologize anyway and try to give you a caring answer, in enough detail that others may benefit from it also.
The term "neural net" is misused in bg bots to refer to a static set of data, maybe updated between software versions, as opposed to being continuously, dynamically updated by the bot's learning as it keeps playing.
In bg, I suspect that the only scientifically, empirically trained neural net was perhaps TD-Gammon's playing billions of cubeless games by making purely random checker moves.
By the time you introduce the cube into the training, you are half way away from empirical science and that much closer to folk science.
And then everybody comes up with their creative ideas to improve their bots with arbitrary formulas, which make it all more and more subjective.
With that, you would expect the bots to diverge further away from one another in time. Especially if their "neural nets" get retrained again and again using those separately "improved" :) formulas.
Of course, this doesn't mean that they would never evaluate hundreds or thousands (who knows) of positions the same way but the more they agree, the more likely it is that they share data and code.
However, if you are lucky to come across even just a few provably strange plays by two bots, you may be well justified to conclude the two bots share a "bug", which a very strong proof of data and/or (most likely) code sharing. (sort of a truncated roll out??:)
The only argument in relation to the above was made by me here was in the thread titled "Is Snowie still relevant?" as an addition to Murat's arguments in RGB which are based on what the bots do after they evaluate positions.
All decisions and actions are made by the "code", based on the data and calculations.
It's a process similar to filling tax forms. :) "If your income is less than $10,000 skip the next question and go to step 9", etc...
When the bots is on roll, it would first ask if it owns the cube, based on which it will either try to figure out its winning chances and make a cube decision or skip that step.
Then it would roll the dice and try to figure out the best move. If one stands out as the best, it will send that move to the next step of code execution which will make that move. If all moves are equal (which may be zero or -1 or however no chance of winning is expressed), it will send any one of those moves to the next step which will make that move which may look (or actually be) randomly selected.
Now comes the subject of resigning because the above stage may be reached several rolls before the end of the game and a good human or bot player would be expected to resign in order to not waste time or irritate his opponent.
Unlike earlier bg bots like Jellyfish and Snowie, current bots like GNUBG and XG do actually try to resign correctly (before considering cube action or rolling the dice) and at least some of the time they succeed.
This proves that both bots calculate their chances of winning (or at least check that it's not zero) as the first thing when they are on roll.
Resigning is a "yes or no" decision. There is no room for "maybe".
But then, when it's a resign position (i.e. zero chance of winning, after evaluating the position!) GNUBG sometimes resigns, sometimes it doesn't.
That in itself wouldn't raise any red flags, but in resign positions XG also sometimes resigns, sometimes it doesn't.
That where Murat's arguments come on the stage, which is that both bots' erratic resigning behavior has to be due to a shared bug, thus shared code.
Furthermore, when both bots share truly unexplainable behavior such as resigning the same arbitrary number of points regardless of the score (which I illustrated in this thread with images), it becomes quite undeniable that the two bots share bugs and thus code.
As no amount of evidence or logical argument would ever be enough to convince believers that there can be no god, none of the above is likely to have any effect of XG admirers.
Personally, in the absence of any counter evidence, logical arguments or even mere admitting/denying/promising to comply/etc. statements, for me it's more than enough to convince me and to sink my esteem for XG and its developers to 0.00000!
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.