Quote:
Originally Posted by PairTheBoard
What stopping strategy is this based on? How would you know if it's the best stopping strategy? Seems like some stopping strategies could get pretty complex.
PairTheBoard
You need to somehow weight the future potential of every branch. The last flip ie position (n-1,k) where k is the running sum at the completion of n-1 flip if one is alive still there flipping (+ and - for head/tails) is easy to decide based on the updated profile of current run if its plus EV or not to flip. If your chance to have good coin is p then do p*(0.51-0.49)+(1-p)*(0.48-0.52) and see if its positive to flip. If its negative stop,avoid the last flip. Here you need to have p>=2/3 for the one flip level left to flip.
The one before gets more complicated. For example it may be leading to a stop in which case you lose exactly 1 (if you were to flip and arrive there ending the branch) or it may lead to a win (+1) that forces another flip next which is also positive (hence overall 1+ ie more than 1). So the EV of flipping with 2 left is the EV of the total branch future not the EV of the next single flip outcome. The probability of each branch is calculated based on the updated profile so far up to that run but then the outcome gives the 2 branches different profiles again so the ++ future for example is coming with a different probability than say q^2 where q was the chance at the previous flip to have been +1 (based on how often good coin that leads to plus + lucky bad coin leading to plus). It is instead q*q' type thing (q'in general different than q because it derives updated information conditional on the first flip being positive with chance q forcing now a new q' for the next flip). I have to rethink that a bit to see how reasonable this is. (maybe one should stick with a profile for all the future branch and not update it when doing this)
Recursively that way the true equity of a position that a flip happens can be seen vs where one doesnt happen.
For example for some very strong ones flipping 2 in a row is a given (a negative flip wont stop from running another say if the coin is 75% to be good). For something that is borderline negative to flip, one may still flip if its possible that by flipping you arrive on one branch to a position that is more positive than the negative that is the other development. I mean one path ends in one flip but the other takes 2 up flips or 1 upa nd 1 down so its overall better to flip for that path even if the 1 flip level was borderline negative say. For example if it is 66.3% chance to be good and 33.7% to be bad then 1 flip alone is negative EV (need 2/3 to be ok and 66.3%<2/3). But if you have 2 flips left maximum the bad flip stops things and leads to eg -1 with some probability (probability to be bad say 0.337 times 0.52 + probability to be good say 0.663* 0.49)=.50011 at which point the decision to stop happens. The chance to flip heads though ie 0.49989 leads to +1 and another flip after that is now positive say (over 2/3 profile updates probability ) by a little bit say if the new updated probability becomes eg 67% it will be 0.67*(0.02)-0.33*0.04=0.0002.
So the EV of flipping with 2 left from a 66.3% position becomes really 0.49989*(1+0.002)-0.50011*1=+0.00077978. So although the 1 flip left level is negative EV (0.50011*(-1)+0.49989*1) the 2 flips level is positive, so you chose to flip then based on the entire future EV not the EV of the single next flip.
Now of course we need to properly formalize this with equations that start by solving the problem backwards from the end assigning every outcome position some net EV and stopping only when the net future EV is negative.
What is your proposal?
Last edited by masque de Z; 12-14-2015 at 06:42 PM.