Genetic Algorithms, Multi-Dimensional Problem Solving, and Trading

Discussion in 'Psychology' started by bwolinsky, Apr 20, 2012.

How Many?

Poll closed May 1, 2012.
  1. 25,000 to 50,000

    1 vote(s)
  2. 50,001 to 75,000

    0 vote(s)
  3. 75,001 to 100,000

    0 vote(s)
  4. 100,001 to 125,000

    0 vote(s)
  5. 125,001 to 150,000

    2 vote(s)
Thread Status:
Not open for further replies.
  1. I don't know if it's normal for someone to wrap their week up doing these calculations, but I find it calming for the next week to have my genetic algorithms problem solve 5 dimensional ranges with 2 variations of stops and targets before packing it in and just running an exhaustive study.

    The last line is the final genetic optimization range, and I wanted to see if anybody knew how many combinations there were out of 1000^5 possibilities.

    How much did I narrow it down, and would you say this is something you do for your trading?

    I write down these ranges from genetic algorithms, and since there are 5 variables it becomes a 5 dimensional problem that I can solve after 4-5 iterations, 4 genetic optimizations, and 1 exhaustive.

    If you know how many combinations there are in the last set of ranges, then you also realize 1000^5 is an impossibly large amount of data to ask a computer PC to solve, so what I've done here is put my scratch paper for solving a 5 dimensional problem using genetic algorithms and exhaustive analysis.

    I use these to have confidence in what I'm doing, but any time I discuss it with non-systematic traders they don't see the value in doing it, when I see many million in this piece of paper.

    So, guess, or calculate how many possibilities there were after the 4th optimization and compute the possiblities for the final 5th line.

    Yes, I wrote in pen and did make one notation error.
  2. Nobody voted.

  3. You may want to check out the "curse of dimensionality."

    Five dimensions is probably too many, depending on each input's error range for your NN/genetic algorithm/other to resolve.
  4. My process will solve it perfectly.

    Start in 25 steps, narrow down, 10 steps, narrow down, repeat in 2's, narrow down, 1's followed by exhaustive, and you will get the best answer, guaranteed.

    The two variations are stop's and targets, which are computed after identifying the best cores somewhere in the middle.

    The steps to solving this 5 dimensional problem came from a Topology PhD from Stanford, and I can safely say that is what works, even if it may take 10-12 hours it usually finds a solution before that.
  5. Are you trying to burn CPU cycles, or is there an actual problem underlying this?
  6. There is a problem. How do you optimize a 3rd derivative genetically to narrow it down into an exhaustive range that can be completed in a reasonable amount of time?

    The answer is this 5th dimensional iterative process.

    More or less, it is the right method to optimize a 5 parameter dataset in successively smaller increments before whittling it down to an exhaustive analysis that will guarantee you to find the best solution. It is an excercise in genetic algorithms I wish I knew two years ago.

    Before I would waste time with 1 to 1000 in 5 variables in intervals of 1 by setting the population size to 1000 and doing 250 minimum 240 generations. Had I been doing it 1 to 1000 stepping by 10 rather than 1 in 5 variables, the answers would have come more quickly and been a lot more accurate.

    Now that I know how to use genetic programming my quantitative automatons are performing much better these days, and I find it is useful after every losing trade to re-train the robot and reacquaint it with market oscillations.

    I don't agree that you can expect genetic algorithms to guaranteeably find the best solution. I believe you must examine what it's thinking, then have it compute the most logical choices once you've narrowed down ranges to less than 150,000 iterations when you can then do an exhaustive study on those rather than 1000000000000000, 1*10^15, 1000^5 possibilites naively expecting to find the solution after 250,000 proximally convergent elitist generations.

    This is the apex of technology. Using a program that would have taken 10,000 years to optimize so that it works in about 12-36 hours. After carefully training your robotic army, the only thing left is to apply it to the markets and know that what you are doing is sending your program to bark fire.

    I consider this artificial intelligence because the sheer probability of humans to be what they are today is just as improbable as the success a robot facing 1*10^15 possibilities has so that it can narrow itself down iteratively, do well in a walk forward test after training, and obey its QuantMaster.
  7. Not even ProfLogic?
  8. alexvnew


    CPU time is a terrible thing to waste...overfitting....:)
  9. Is this something like Sudoku?
    #10     Apr 24, 2012
Thread Status:
Not open for further replies.