I have now downloaded some of the matches I have played from Galaxy into XG2. I have noticed some differences in the suggested moves and also a big difference in the PR for the games. I am relatively new to both XG2 and Galaxy so how do I get them singing from the same hymn sheet? I obviously cannot change Galaxy, but some settings in XG2 may need tweaked.

XG by default uses a higher level of analysis than Galaxy.

If you go to Options -> Settings -> Analysis and it looks something like this, the XG analysis is trustworthy and more trustworthy than Galaxy. I almost never look at the Galaxy analysis.



(I love the new image upload capabilities!)

My settings are identical, so what does this mean?

It means you should put much greater trust in the XG analysis than in the lower-level Galaxy analysis.

If you wish, you can do still better by selecting moves you're interested in and having XG analyze them at XGR++ level, and better yet by doing a rollout of some length.

Totally agree with Mr. Robertie. I used to love the convenience of Galaxy to review quickly on their site, but was convinced by my coach to download to XG (usually takes a couple of minutes or less). Many recommended plays are different, and even had the match where the PR’s flipped from player that was higher to lower. Use XG.

XG is the Piosolver of backgammon. Nothing (yet) is going to be a substitute for XG.

do you regard XGs evaluation of XGID=------a----------BBCCBBAn-:1:1:1:00:0:0:0:0:10 as BG solved?

you gonna find me a better one

Is your question: "show me a better bot". Than I can't. XG is the best bot on average (even if one takes into account that there is a bias when XG evaluates itself) but on average doesn't mean it's infailable.

When your question is: "show me a bot better at this position"? No problem, although BGBlitz doesn't understand (yet) that it is beneficial to slot an outside prime at the front it understands that the prime player is the clear favorite. And GnuBG (what has sometimes issues with very unusual positions) gets it better with around 50/50.

So BG is not solved yet although this is a common notion. That takes some further work.

Have you tried training a better containment neural net? Both Gnu and XG very consistently underestimate how easy it is to contain a checker after a late hit. Especially when there is a possibility of hitting multiple checkers. It's almost like you can add 0.08 to the evaluation of certain types of positions and end up with something that is guaranteed to be closer to a rollout result. That's how bad the bots are at these types of positions.

4-ply:


rollout:

very interesting that the cubeful evaluation changes are larger than for the cubeless values. Hm....

I don't know much about XGs architecture, the only info I'm aware of is that Xavier once told on bgonline "XG uses ***many** different NN based on the position type".

I don't know how many "many different NN" are, but this would be quite interesting (at least for me ) Or how he trains, from the same posting one might assume self play.
I'm a big fan of self play too but the devil is in the details, not every approach I tried work. And better containment is a goal for the next AI, probably this autumn (it's overdue...)

I'm sure you know this, but the cubeful equity change is so large because XG realizes on the rollout that the cubeless win % is high enough so that the cube is very efficient. So that's worth a lot in terms of cubeful equity versus the fairly inefficient cube in the 4-ply evaluation.

Someone once told me that the XG engine is basically an optimized Gnu BG. I'm not sure how true that is, but they do seem to have similar biases so I believe there's some truth to it and the neural nets are similar between Gnu and XG. I'm assume Gnu uses multiple neural nets as well?

Thanks for the info, that makes sense.

I know the rumour that XG is based on GnuBG but I don't see any supporting information for this. And why? This would imply somehow that it is impossible to come up with an implementation of it's own. And if Xavier is able to improve it, why not able to create something of it's own? The AI in XG is written in Assembly language (so a different language), many nets is probably more than 3 so the architecture is probably different too and he probably uses a different learning approach. No this is just a rumour.

GnuBG uses 3 nets. One for running positions, one for contact positions and one for crashed position (one side has most checkers on the low points and virtually only forced moves) and they don't use self learning but supervised learning.