tag:blogger.com,1999:blog-6927558.post7710086357559969484..comments2023-08-29T10:30:31.267+01:00Comments on Biofractal: An Introduction to Applied Evolutionary Metaheuristicsbiofractalhttp://www.blogger.com/profile/11229842750489025189noreply@blogger.comBlogger4125tag:blogger.com,1999:blog-6927558.post-53823142594736840632009-09-18T23:06:52.218+01:002009-09-18T23:06:52.218+01:00Thanks for your interest.
I believe the maximum c...Thanks for your interest.<br /><br />I believe the maximum complexity comes from the truncation operator which is O(M 2 log M) where M=N+N', N is the population size, and N' is the archive population size. This operator is a component of the original SPEA2 algorithm. <br /><br />For more info see Zitzler, Eckart and Laumanns, Marco and Thiele, Lothar (2001) SPEA2: Improving the Strength Pareto Evolutionary Algorithm. Evolutionary Methods for Design, Optimisation, and Control.biofractalhttps://www.blogger.com/profile/11229842750489025189noreply@blogger.comtag:blogger.com,1999:blog-6927558.post-29398529140937676882009-09-09T22:51:57.740+01:002009-09-09T22:51:57.740+01:00Thanks for the answers, you made most of it clear....Thanks for the answers, you made most of it clear. Somehow I skipped through how the multi-objective EA works on the problem until it being Pareto optimal. The exprectation of the real world test were not stessed out this much but now I get that one as well.<br /><br />As for my third question I thought that the problems were set as single-objective but reading back I now see that the multi-objective EA returned multiple Pareto optimal solutions. The pay-off matrix got only clear to me now so I think I understand all of it :)<br /><br />One more thing I'm curious about - even though it is clearly out of the domain of the paper itself - is wether you could estimate the complexity of EA algorithms (or specifically your algorithm)? It could really be an important decision factor when working with multiple dimensions (objectives) if it turned out that the complexity was extremely high. As you've referenced Alba, Enrique, Marti, Rafael (Eds). 2006 neural nets may have limitations when training manually but the great thing about is that it's easy to implement an O(n) complexity algorithm n being the training data quantity.Gergely Oroszhttps://www.blogger.com/profile/05299003916141837135noreply@blogger.comtag:blogger.com,1999:blog-6927558.post-64598320753202360532009-09-09T21:15:40.374+01:002009-09-09T21:15:40.374+01:00Thanks for the questions. Here are my answers.
1....Thanks for the questions. Here are my answers.<br /><br />1. A multi-objective EA does not use weightings. Weightings are used by single-objective algorithms to simulate multi-objective capability. Instead a multi-objective EA attempts to maximises or minimise all the objectives at the same time, in parallel. No objective is superior to another. A solution is considered fully optimised when adjusting any one of its objectives compromises one of the other objectives, such a solution is considered 'Pareto optimal'.<br /><br />2. Yes, the difference is small but, and I did not make this clear, the expectation was very different. The farmers, form their many years of experience, predicted a completely different outcome, namely that the *only* way to maximise net revenue was to maximise weight gain through feeding with the most expensive feed stuff. The alternative strategy showed that this is not true. That approx the same money can be made by feeding with cheaper food stuff to create lighter animals. This preserves the farmer's margins whilst having positive sustainability implications.<br /><br />3. The problem was multi-objective because its objectives were in conflict with one another (when objectives do not conflict they can be treated as a single objective). The only way to properly solve such a multi-objective problem is with a genuinely multi-objective EA, such as IC-SPEA2. Using a single objective EA and simulating multi-objective capabilities through the use of weightings etc has been shown not to work very well (the solution tends to reflect the weightings not the inner problem landscape).biofractalhttps://www.blogger.com/profile/11229842750489025189noreply@blogger.comtag:blogger.com,1999:blog-6927558.post-37903979713976961852009-09-09T15:24:14.459+01:002009-09-09T15:24:14.459+01:00Thanks for sharing this paper, it was a truly inte...Thanks for sharing this paper, it was a truly interesting read.<br /><br />I have three questions:<br /><br />1. As you've mentioned in the section "What is a Multi-Objective Problem" weighing multi-objective problems has the danger of resulting in a solution favoring one of the objectives more. However it is not clear to me how evolutionary algorithms deal with this. Are all objectives set as equal weighs? Or is the problem solved with different weighs and then evaluated accordingly?<br /><br />2. In the Mexican farm real world example you wrote as conclusion "Yet IC-SPEA2 found a strategy that gave the highest net revenue while delivering lighter animal". Are you referring to solution 2? To me it is not seem really covnincing that there is about 0.1% difference between the two solutions. This seems so marginal that it could be easily be because of the nature or implementation of the specific problem.<br /><br />3. Was the implemented algorithm a multi-objective problem? If so, were any weighs assigned to the objectives (max profit/max weight/min diet costs)?Gergely Oroszhttps://www.blogger.com/profile/05299003916141837135noreply@blogger.com