OK so they're so damn strong now maybe it's not necessary but I've been wondering whether bots could be coded to win more against humans.

Bots are programmed simple to play with the lowest error rate as defined by self play. If a bot were programmed, when conditions were right, to give up small amounts of equity by making sub-optimal moves which resulted in more complex positions for the human opponent I believe it would win more games.
How would one do this?
Well firstly you'd have to define what complex meant.
The bot could look at the replies and based on how many moves got evaluated at the final ply could measure a basic level of complexity.
If more replies were deemed worthy of full evaluation then the human opponent would also have to consider more moves which in theory would mean greater chances to err.
I don't know enough about neural nets to fully elaborate but I believe by looking at the weights a bot can determine the type of position.
An example would be if the best move turned the game in to a running game it might be better to give up a small amount to equity to maintain a contact game.

There would have to be some trial and error to establish the correct parameters but a programmer could launch various incarnations on an online backgammon server for testing.

IB

I've often wondered what the bots would do about doubling choices if you somehow assigned opponents specific doubling tendancies like most humans have. Would be interesting to see what they came up with imo.

Ask this question in the bgonline forum. Perhaps you will get a valid answer from XG's developer.

I toyed with implementing something, but the ROI did not seems enough to justify the large amount of work needed.

There are several things that can be done:
* Skewed MET: probably the most important part as most gain the bot can get against weaker player is to avoid raising the cube too much.
* Measure of complexity of the position, this could be generated by a Neural network and used to change the equity of a move: the more complex the better the equity from the bot perspective.
* use user profile to determined specific weakness: this seems difficult and also dangerous.

All in one, yes the bot will win quite a lot more against weaker player, but that seems useless for me (if one day TD allow bots to enter tournament, that would be a different story). The main interest of implementing this is to analyze game where players' skills are quite different.

Hello Xavier
Did you get anywhere at all on this part even something as basic as how many moves made it through the SI to 3-ply?

Like you said though more important features to implement right now.

Also does XG play on any online servers?

IB

I've wondered about this also. I suspect that there are lots of players that do better against a weak opponent then a bot would. Picking up on cube patterns from game to game seems like a huge advantage to me. Especially if they are large flaws in a weaker players game.

As I said, I did not put much effort into that. I have several idea in determining position complexity, the easiest is to check position where 1-ply and 3-ply disagree. Another is to use Rollout duration, the longer the rollout the more complex the position is.

XG plays only on GammonSite: www.gammonsite.com which is our online site XG has been playing there for about 8 years :-)

OK. I don't play online.
What versions (strengths) of XG play on Gammonsite and what are the elos?

Thanks

They play using 3-ply and their Elo is around 2150.

Neil Kazaross is around 2130 currently, but he does adapt his game to weaker player. I think XG is losing about 30-50 Elo by not doing the same.

Thanks.
Where can I see the rankings?
Any plans to unleash XGR+ on Gammonsite?
When is XG2 beta due out?


Thanks!

IB

Ranking: you can see them by connecting to the site: you get a 30 day trial period.

XGR+ on GammonSite: people are complaining already the bot wins too much

XG2 Beta, maybe by the end of this month.

I'm sure when you first released XGR+ it always cheated against me but as my PR got lower it stopped .

Looking forward to the Beta.

Thanks Xavier