Comments on: Towards billion bit optimization via efficient genetic algorithms http://www.kumarasastry.com/2007/02/15/towards-billion-bit-optimization-via-efficient-genetic-algorithms/ Tue, 02 Feb 2010 08:24:39 +0000 hourly 1 http://wordpress.org/?v=3.1 By: Kumara Sastry http://www.kumarasastry.com/2007/02/15/towards-billion-bit-optimization-via-efficient-genetic-algorithms/comment-page-1/#comment-210 Kumara Sastry Wed, 10 Oct 2007 02:28:34 +0000 http://kumarasastry.com/?p=233#comment-210 David, Thanks for the comments! Yes thats a typo. It should be minus. Regarding multimodal function, it depends on the type of multimodality. Note that cGA solved noisy problems. The addition of noise introduces many local optima (the global optima is not hill-climbable anymore) and cGA still is able to get to the global optima. However, if the multimodality comes from higher order building blocks, then yes cGA will converge to local optima. However, work is currently on-going to extend this work to develop efficient GAs that can solve million(s)-variable problems that require identification and exchange of higher order building blocks. David,

Thanks for the comments! Yes thats a typo. It should be minus.

Regarding multimodal function, it depends on the type of multimodality. Note that cGA solved noisy problems. The addition of noise introduces many local optima (the global optima is not hill-climbable anymore) and cGA still is able to get to the global optima. However, if the multimodality comes from higher order building blocks, then yes cGA will converge to local optima.

However, work is currently on-going to extend this work to develop efficient GAs that can solve million(s)-variable problems that require identification and exchange of higher order building blocks.

]]> By: David http://www.kumarasastry.com/2007/02/15/towards-billion-bit-optimization-via-efficient-genetic-algorithms/comment-page-1/#comment-208 David Fri, 05 Oct 2007 16:42:54 +0000 http://kumarasastry.com/?p=233#comment-208 Congrats for this barrier breaking results and the best paper award at GECCO in particular and for the good job done at IlliGAL in general. I would have a small remark and some questions. For me it seems that in the probabilistic model update (Eq. 1) if $x_{w,i} = 0$ then the update should be $p_i^{t+1} = p_i^t - 1/n$ not + 1/n. Q: How multimodal billion-variable problems can be solved quickly, reliably and accurately? My first impression is that the parallel cGA might easily converge to a suboptimal solution in a multimodal search landscape. To put it more generally, if there are competing schemata in the problem, is there a way to handle them somehow with small memory requirements? Thanks for the answer. Keep up the good work! Congrats for this barrier breaking results and the best paper award at GECCO in particular and for the good job done at IlliGAL in general. I would have a small remark and some questions.

For me it seems that in the probabilistic model update (Eq. 1) if $x_{w,i} = 0$ then the update should be $p_i^{t+1} = p_i^t – 1/n$ not + 1/n.

Q:
How multimodal billion-variable problems can be solved quickly, reliably and accurately? My first impression is that the parallel cGA might easily converge to a suboptimal solution in a multimodal search landscape.

To put it more generally, if there are competing schemata in the problem, is there a way to handle them somehow with small memory requirements?

Thanks for the answer.

Keep up the good work!

]]>