CMA stop conditions #401
CMA stop conditions #401
Conversation
|
Thanks for opening your first pull request in this repository! Someone will review it when they have a chance. In the mean time, please be sure that you've handled the following things, to make the review process quicker and easier:
Thank you again for your contributions! 👍 |
| @@ -274,6 +291,17 @@ typename MatType::elem_type CMAES<SelectionPolicyType, | |||
| return overallObjective; | |||
| } | |||
|
|
|||
| // Terminate if sigma/sigma0 is above 10^4 * max(D) | |||
There was a problem hiding this comment.
| // Terminate if sigma/sigma0 is above 10^4 * max(D) | |
| // Terminate if sigma/sigma0 is above 10^4 * max(D). |
To be consistent with the style.
I think perhaps some review would be fine, but I believe its okay. It should be ready to be merged |
|
Since this is so close, we should keep it open to get it merged. I marked it keep-open |
|
Hello, is this issue still open? |
|
@NilotpalMaitra still open; needs someone to take it across the finish line. |
Sure! I would try my best to resolve this issue! |
|
Can anyone just tell whats the final issue, its hard to decipher from the above chats |
|
Hi @NilotpalMaitra, the goal of this was to expand some of the stopping conditions the CMA-ES currently had. If you check the original post I left some references to the papers as well as a Python library maintained by the author of those papers. If you go through them, you will see there are some of these stopping criteria missing in the current implementation. The modifications aren't as straightforward as I originally thought, with some modifications needing to change quite a bit of code to add arrays/history of previous iterations and substantially modifying some of the API. This would then need to be replicated for all CMA-ES variants in the library to maintain coherence. Overall, there is a lot of code that would need to be overhauled and tested, and I believe it would be beneficial to have some discussion with the maintainers before committing to these changes. Feel free to take it up and finish it. |
|
Thank you so much for summarizing the issue! as you have mentioned that the issue is not as simple as it was, So I would definitely go through the paper first to give myself a gist of what all is happening . |
Hi,
I am currently enhancing the implementation of the CMA-ES algorithm by introducing additional stopping conditions. These modifications are intended to improve numerical stability and are advocated in A Restart CMA Evolution Strategy With Increasing Population Size and The CMA Evolution Strategy: A Tutorial. The changes are also in line with implementations found in Python libraries maintained by the same author.
These are basically:
maxiter): Stop if the number of iterations reaches a predefined limit.ftarget): Stop if the best objective value is less than or equal to a target value.maxfevals): Stop if the number of function evaluations reaches a predefined limit.tolfun): Stop if the range of function values of the best solutions falls below a threshold over a certain number of iterations.tolflatfitness): Stop if the best objective value does not change significantly over a number of generations.tolconditioncov): Stop if the condition number of the covariance matrix exceeds a certain threshold.noeffectaxis): Stop if adding 0.1 times the standard deviation in the search space in a principal axis direction does not change the solution.noeffectcoord): Stop if adding 0.2 times the standard deviation in the search space along one coordinate does not change the solution.tolx): Stop if the change in the variables between generations falls below a threshold.tolfacupx): Stop if any scaled step sizes exceed the product of the initial step sizes and a specified tolerance factor.tolfunhist): Stop if the range of fitness values in the historical best fitness list remains below a specified threshold.tolupsigma): Stop if the ratio of the current step size to the maximum eigenvalue of the covariance matrix exceeds a specified multiple of the initial step size.tolstagnation): Stop if there is no significant improvement over a large number of generations.tolfunrel): Stop based on a relative condition involving the historical and current best function values.These conditions are the ones currently present in the Python repository I mentioned and align with the papers while adding a few other stopping criteria. These are what I plan to implement, but I'm open suggestions.
Additionally, I have made modifications in lines 94 and 112 to align with the initialization presented in The CMA Evolution Strategy: A Comparing Review (page 26, equations 35 and 37). These initialization parameters are also supported by other papers I have reviewed.