-
Notifications
You must be signed in to change notification settings - Fork 0
Problems
lymacasm edited this page Aug 13, 2017
·
2 revisions
This page lists current problems by data (starting at the creation of this page). I'll keep old problems (if we ever get to that point).
At this point, we have a classifier with 99% accuracy. The only problem is the straights:
- Ace can represent either 1 or 14. Since Ace is stored in memory as 1, the current set of conditions cannot detect straights if an ace represents 14 (straights use arithmetic sequence detection primarily). Because of this, the algorithm has trouble distinguishing between straights and the nothing hand when the ace could be used to represent 14 (such as in the case where the hand is 10,J,Q,K,A). This hand can be hard to distinguish from a hand that is not a straight, such as (J,Q,K,A,2), as they have the same number of arithmetic sequence transitions, and max - min value is 12 in both cases, yet one hand is a straight, and the other is nothing. However, if the straight has an Ace with value 1, it poses no problem (such as the case where the hand is A,2,3,4,5).
These are problems I think we have currently with the code:
- Rules for straight are too difficult to learn. Might need better rules.
- New rules I made in the hopes of learning the pairs/3-of-a-kind/etc are too specific (i.e., how many times is this specific number in the hand?). Need to maybe generalize somehow.
- It's hard to learn the classes with lots of data entries with our current fitness function. Maybe it should be tailored to learning them?
- Maybe we should hack something by limiting what our algorithm can do, e.g., force default class to 0, and don't generate any rules for class 0 (class 0 is the "Nothing" hand)
- Understand distribution of our test dataset. One thing I might want to do is for someone to write code that will correctly label our test dataset so we can get more info from it, and get a better idea of what we are doing wrong.
- Maybe try and see which classes our classifier is getting confused with
- Try and find problems with our current rules, and ways to fix them (maybe investigate classifier.py for this)