Optimization methods (genetic/brute force)

[pureporsche]

I'm curious, I have read that sometime soon you will be releasing an updated version of ninja trader (7 I think). Can you please comment on the changes to the optimization algorithms used.

I'm assuming you are using a standard brute force "solve" right now. Iterating through all possible sets of variables within a given range. Have you looked into a genetic optimizing algorithm - which is drastically faster and yields better results due to the fact that larger ranges may be employed.

Also will the new version be multi threaded, specifically when running an optimization - taking full advantage of multi processor machines?

Thanks in advance for your comments.

Mike

[pureporsche]

PS. By genetic, I mean stochastic optimizer - one that takes advantage of random chance in it's operation and focusing on an important region of a solution space while mostly ignoring "dead ends".

[NinjaTrader_Jason]

Hello Mike,

NinjaTrader 7 will support multi core processors. I have contacted Development regarding your first inquiry.

[pureporsche]

Perfect - Thanks Jason!

[NinjaTrader_Dierk]

NT7 will come with a GA (genetic algorithm). We will provide details as we publish first beta.

[Anagoge]

That is great news! I'll be very pleased to see a genetic optimizer built into NT 7. It would be nice if we can define out own fitness function to rank the trials the genetic optimizer runs (and maybe even set limits like only accept trials with a max drawdown < 20%). The original genetic optimizer hack posted here used SQN for the fitness function, but that isn't ideal for everyone.

I don't like to be greedy, but it never hurts to suggest things for NT8, so...
If you ever get the urge to implement something (possibly) even better, check out CMA-ES and the benchmark comparisons here with other optimization algorithms:

http://www.bionik.tu-berlin.de/user/niko/cmaesintro.html

Some advantages are that CMA-ES auto-tunes the population size and search space, and it also notices when there have been enough iterations and stops the process early when no more progress is being made, so it auto-tunes the number of generations as well. This removes the requirement for the user to even think about things like population sizes, generations, etc.

[NinjaTrader_Jason]

Hello Anagoge,

I will forward your suggestion to our development team and ask them if they can add this to the list of future considerations for the software.

Thank you for your feedback.

[jdfagan]

I don't like to be greedy, but it never hurts to suggest things for NT8, so...
If you ever get the urge to implement something (possibly) even better, check out CMA-ES and the benchmark comparisons here with other optimization algorithms:

http://www.bionik.tu-berlin.de/user/niko/cmaesintro.html

Some advantages are that CMA-ES auto-tunes the population size and search space, and it also notices when there have been enough iterations and stops the process early when no more progress is being made, so it auto-tunes the number of generations as well. This removes the requirement for the user to even think about things like population sizes, generations, etc.

CMA-ES (Covariance Matrix Adaptation Evolutionary Strategy) is impressive and I've used it with AmiBroker (http://www.amibroker.com/devlog/wp-c...eadme5140.html - search for CMA-ES) to nice effect. Perhaps this can be done with NT 7 since other 3rd party GO optimizers have already been written (e.g., Pete's, wmGO, PH Genetics, MoGo, etc.).

JD

[NinjaTrader_Jason]

Hello jdfagan,

I will forward your suggestion as well. Thank you for your feedback.