Robot Wars

[diane]

This topic was raised before, here - bot tourney, and although a few people expressed an interest - and thought it would draw a crowd - it got lost in the original topic i think.

Since it has been mentioned at least three times over this weekend - I wondered if it would be possible.

If we could schedule a time - one weekend perhaps - with plenty of warning, could the bots - as many as possible, stop their matches (or a little before - so no dropping   ) and play each other in a tourneybot run tournament?

I dont know who to contact for a lot of them, but webbie - I guess it would come down to you to sort out the MGNU's.

Would it be possible - is it difficult - and could we do it??  And who would be interested to watch - me for one!!

[EddieVedd]

Yeah I'd watch  I wouldn't have a clue who to contact either but let me know if i can help any other way.  

[lewscannon]

It would be interesting to see how many doubles were rolled in this tournament.

[diane]

Well this kinda struck me - analysis of the played matches would be interesting - just how lucky they get against each other  

Course, playing back the analysis might be the only way to actually see what some of the moves were - given how fast some of them can play.

[webrunner]

It is possible for the mgnu bots. I control 4 of them.

[spielberg]

Well as webrunner controls four of the bots I suggest a tourney , either  7 / 9/ 11  or 9/ 11/ 17 (match lengths), for those four bots  . If we capture all the matches then the time the tournament is held becomes less of an issue - anyone interested could "see" the games whenever they wanted.

How "fixed" are these bots btw webrunner - are they still learning ?

[diane]

Whilst four does constitute enough for a small tournament - it would be nice to see some others participating - bonehead, parlorbot and donzbot come to mind.  It would proabably all be a bit beyond motif - and montecarlo is limited to 1 pointers - so thats probably out. Any other ideas?

[lewscannon]

I'd be pulling for harmony to win the battle of the bots

[tryout]

How "fixed" are these bots btw webrunner - are they still learning ?

I haven't heard so far that any maintainer of a bot based on GNU Backgammon (gnubg) has trained the neural net further than what he got from the gnubg team at that time. The neural net of gnubg is trained before release and then stays fixed for this version. It's just a quite complicated evaluation function with the position and match information as input and some percentages as output.

Since all bots play on a pretty high level I'd suggest longer matches (say 21-pters or more) and a series of matches to determine the winner (at least 9, better 15). Also, a round robin system, i.e. league, may be better than a knock-out-tournament, especially with few contestants. Despite this the winning bot will still be determined by luck only. (For all you believers: not because the bots are all lucky, but because the differences between their playing strengths are rather small.)

[diane]

Hmm - i didnt really have in mind to have very long matches to rule out luck - since we all know they play very well, i wanted to see how the luck factored in all of this - so whilst long matches are ok - since they would be played fast anyways - I want also to see if any bot is particularly lucky - and if it affects their standard of play. When webrunner posted the screenshot of mgnu_avdanced in a somewhat compromised position, I am sure the analysis said it hadnt played very well.  I am sure this isnt the case - but it would be a nice side interest.

So how do we get Harmony playing then?

I will email and see if we can get parlorbot in.  

[tryout]

I want also to see if any bot is particularly lucky - and if it affects their standard of play. When webrunner posted the screenshot of mgnu_avdanced in a somewhat compromised position, I am sure the analysis said it hadnt played very well.  I am sure this isnt the case - but it would be a nice side interest.

If a certain bot is (un?)fortunate enough to have a lucky streak over the matches in this particular tourney it will get marked as always lucky or renamed to "LuckyBot"?
And of course luck (or its absence) doesn't effect the decisions of a bot, just the outcome of the match.

I haven't seen this screenshot, but suppose you mean an unfavourable board position. Neither then played the bot any worse than else. So there's nothing to see in the analysis.

FYI, mgnu_avdanced plays actually worse than the neural net is trained for. On this level it intentionally adds noise to the neural net output which often leads to not choosing the best decision. So of all gnubg based bots we have the best chance against mgnu_avdanced.

[diane]

I haven't seen this screenshot, but suppose you mean an unfavourable board position.

screenshot post
OK - I was mistaken - and it was mgnu_expert - but the rating it ahieved was intermediate - I cant remeber if that is normal for that bot - but they should be playing consistently regardless of the luck for sure.

I was explaining myself badly - I know all will have different reasons to want to watch this - but since I am learning about luck and how to manage it - I am curious as to how the outcomes are affected by luck, if the level of play is not affected.  So - if a bot is on a particularly unlucky streak, but continues to play very well - does the luck have less effect or the same?  That kind of thing.

Anyways - there are lots of points of interest I think - and I dont think a bot lucky in one tournament wil be deemed forever lucky, except by the very dim - surely???

So of all gnubg based bots we have the best chance against mgnu_avdanced.

Grrrrrrrrrr - this is the one I fare the worst against!!

[spielberg]

The result of battles between skilled opponents is generally affected much more by luck.

This is easy to prove :

One footballl team (lets call them the Giants) play two other foootball teams.

Their first match is with a kids mixed sex side - luck will have very little influence on the result.

Their second is with a professional well financed side ( lets call them the Raiders) - suddenly the way the ball bounces becomes a major factor in the eventual outcome of the game.

[purol]

hi,


for those who cant wait to see a bot play another bot ...

... you can let gnubg play gnubg on your own computer  (and choose your own settings).  

however, no kibitzing ...



purol

[dorbel]

Far more matches are won by the side that is luckiest on the day than is generally realised. The shorter the match, the greater the mean variation in luck. If one side has a Snowie luck advantage of +5, it is almost impossible to overcome this with good play and the majority of short matches are decided by the dice. Although bots appear to be "luckier", they do not on average have better luck than humans. The appearance of luck is the outcome of their better play; strong play creates more good rolls next turn than weak play.
Watching bots play each other is no more or less  interesting than watching two strong human players, in fact for learning purposes, I find all matches to be equally interesting, even if the protagonists play quite badly. I am not sure what the bot tourney is supposed to show.
No lews, bots roll the same number of doubles as humans, because like humans they get what the server gives them.

[diane]

for learning purposes, I find all matches to be equally interesting

For sure this is true, and I watch as many as I can.

The only difference here is that the bots shouldnt have off days - momentary lapses etc.  Watching players is relatively easy to set up - this would just be a one off for extra interest.  I know what the theory is on luck v skill, but being the sort of person I am, I tend to prefer empirical data - it kinda makes it more real to me.

This is totally aside to the fact it would be fun to see which bot triumphs  

Although so far - there doesn't seem to be a great deal of interest - and I think (I am guessing a bit - since no one went into what is needed) there is quite a bit of setting up to do - so unless anyone else wants it - it may not happen.

And I think Lews knew that    

[lewscannon]

No lews, bots roll the same number of doubles as humans, because like humans they get what the server gives them.

Then how do you explain the last time I played mpgnu_expert, it kibitzed 'eat my dust, sucker' & 'your momma's ugly', and rolled 4 double 6s in a row while bearing off? It doesn't even know my mother!

[webrunner]

No lews, bots roll the same number of doubles as humans, because like humans they get what the server gives them.

Then how do you explain the last time I played mpgnu_expert, it kibitzed 'eat my dust, sucker' & 'your momma's ugly', and rolled 4 double 6s in a row while bearing off? It doesn't even know my mother!

That is because you dropped him before Lews  

[diane]

These bots just get cheekier and cheekier - has anyone seen I Robot??  

[diane]

Since not many people have said whether or not they are interested in this - i thought it might be easier to guage the interest via a poll.  Please vote - and then we can see if the work required is worth it.