Designing World Useful Portfolio and Benchmarks for Evolutionary Algorithms

Project: Research

View graph of relations


Evolutionary algorithms (EAs) are versatile black box optimization techniques. There is no need to know the mathematical form of the optimization objective function. Hence it is applicable to search landscapes which are non-differentiable, non-smooth, plagued with large number of deceptive local optima, or simulation based functions, which cannot be easily solved by conventional techniques. EAs also offer the possibility of cooperating with traditional techniques in a memetic computing framework. Finally, it is applicable to multiple objective (MO) problems which are less amendable to conventional techniques.However, there is very little study on how a portfolio of EA may synergistically interact to give even better performance. In this research, a novel automatic and parameter-less EA portfolio composition method is advanced. Our method will compose a portfolio of one or more algorithms that would give the best average performance. It gives an answer to which algorithms and the number of algorithms that should be in the portfolio, and allow the special case of one algorithm. Moreover, the constructed portfolio is guaranteed to have better average performance than any individual algorithm. Finally, this approach will output good portfolio type algorithm that is more stable than individual algorithms!Next, we investigate the problem of how to derive a suite of benchmarks that reflects the world, an extremely important topic within evolutionary computation. We propose a novel solution to this problem using the idea of reality measures. Researchers will now have a measure that they can use to justify the degree that their EAs are useful to the real world problems. This measure will be the first of its kind in evolutionary computation.Finally, we extend our method in novel ways based on the results of the above. Two important extensions are to MO problems and algorithms that permit restart. These will deliver more powerful portfolio type EA algorithms that solve real world problems more effectively.


Project number9041927
Grant typeGRF
Effective start/end date1/12/1323/05/18