GIB pass to double

[Yu18772]

http://tinyurl.com/7jc8n8j

[barmar]

The book bid is 4♣. It then simulated Pass, everything from 3♥ through 4♥, and 5♣. The simulations told it that it could probably set them 2, which is better than any of the games. And game didn't always make in its simulations, in which case setting them 1 was enough.

 

It looks like the problem is that it believed North's bidding -- it assumed North would have raised the preempt with at least 3 ♦. So in the simulations, West always had a couple of ♦, and therefore it had full HCP strength for its bid. That left less for NS, so they were probably going down. And slam is unlikely because of the 2 ♦ losers.

 

The simplest fix is to disable simulations in this rule, although then it won't find 4♥, either. So we'll need to tweak the rules to make it bid this (currently, the rule for 4♥ wants a 5-card suit, because it's not sure that West has 4 cards in both majors).

[barmar]

And before you ask, I did consider not being so trusting of the opponents bidding. But that's hard. After all, if the opponent opens 1♠, you want most of your simulations to include a 5+ ♠ suit, don't you? So how do we teach the robot to be trusting of some bids and not others?

[xxhong]

I think the bot should always trust opp's bots' bidding, but not trust human opp's bidding too much. At least, it should set a wider range for opps' bidding. For example, a human opp may open 1NT with 12-18 HCPs, with singletons or 6 or even 7 card suits.

 

In a more sophisticated practice, it should certainly build a database on human opp's bidding history and determines how likely opps's bidding deviates from the system. That's certainly an enormous task. I think those poker bots do have methods of historic play of opps implemented.

 

And before you ask, I did consider not being so trusting of the opponents bidding. But that's hard. After all, if the opponent opens 1♠, you want most of your simulations to include a 5+ ♠ suit, don't you? So how do we teach the robot to be trusting of some bids and not others?

[cloa513]

I think the bot should always trust opp's bots' bidding, but not trust human opp's bidding too much. At least, it should set a wider range for opps' bidding. For example, a human opp may open 1NT with 12-18 HCPs, with singletons or 6 or even 7 card suits.

 

In a more sophisticated practice, it should certainly build a database on human opp's bidding history and determines how likely opps's bidding deviates from the system. That's certainly an enormous task. I think those poker bots do have methods of historic play of opps implemented.

The bots weren't reliable in this case so trusting humans is not relevant.

No the solution is do what a human would do and make the reasonable bid- preempts work. Don't require 5 card major to jump to game- that's the risk of doubling that you might end in a Moyesian fit game. Sometimes choose 4 clubs or 3NT even with 4-3 majors.

[barmar]

The bots don't know which players are humans versus other bots. The underlying GIB software allows EW and NS to play different systems, but we don't make use of that facility, and it's not really the same thing (you're talking about differentiating between North and South, not between NS and EW).

 

The reason the North bot wasn't "reliable" in this case is because it did simulations, and they told it that passing was better than competing. However, when the East bot is selecting deals for its simulations, it assumes book bids for all the previous bids by the other players. Doing simulations for every previous bid would result in exponential explosion, because we'd then have to simulate the other players doing simulations of all the previous bids when they're deciding what to do, and so on. So it needs to make simplifying assumptions to make the computation tractable.