Suggestions of how to improve this needed
ThorvaldAagaard opened this issue · 3 comments
In the following position:
the suggested card from the neural network is just a bad play.
i agree. it's better to test the spades first. if they don't split, we still have the option of a squeeze. the squeeze is clearly superior if East has both spades. If West has the two spades then both the squeeze and the finesse are an option.
Now Ben has committed itself to the finesse before knowing what is going on.
I think that there are several problems
- the NN is not advanced enough to understand cardplay technique described above
- the DD solver believes that playing D is exactly as good as playing SA, because it will only make the finesse when it actually works and will try for the spades to split otherwise. in the samples there is no case where a squeeze against West would be the correct play.
a relevant question: did West show out of spades earlier in the play?
In fact it is better to play CT first, as there is a finesse possible in spades too.
Opening lead was ST from East, that ran to the Queen.
Declarer ruffed a heart, and drew the trump.
With open cards it is easy to take the Ace of spades, but playing S6, seems like an ok chance, as ST could be from length or shortage
After ruffing the heart-return, the position in hearts are clear, but still not in spades and diamonds.
It is OK, that the DD-dummy reports all is equal, but interesting why the NN is suggesting D3.
The CT is not considered by NN, perhaps because as it normally is good to keep a trump.
I think this position should be solved by NN, and perhaps we should make a specific training for 5-6 card end positions
yes, as it is, the NN is just too attracted to play towards the AQ because it has seen that done millions of times probably :)
the NN doesn't try to play well, it tries to play similar to what is in the data.
the NN doesn't "see" the spade finesse because it sees the 7 as x.