Knowledge Incorporation in Evolutionary Computation


Interactive Evolutionary Computation

The basic idea of interactive evolution is to involve a human expert on-line into the evolution process. During evolution, the human expert selects one or more individuals that survive and reproduce to form a new generation.

However, it would be arduous for human beings to sit in front of the computer and make endless choices. This entails a method to abstract the knowledge from the human expert and build a surrogate so that the evolution process can be carried out freely and automatically. This is also related to evolutionary computation with approximate fitness functions.

Multiple Objective Optimization Using Fuzzy Preferences

Multiple Objective Optimization (MOO) has received wide attention in evolutionary design optimization. In many cases, not all objectives are equally important. A common situation is that "objective A is more important than objective B, and objective C is equally important as objective D" and so on. In this case, the preferences expressed in human language need to be translated to numbers in optimization. To this end, fuzzy preference modeling and fuzzy multiple criteria decision making can be applied. More importantly, the preferences can be changed interactively during evolution.

Knowledge Incorporation in representation, population initialization

One most direct method for incorporating knowledge into evolutionary computation is to encode the a prior knowledge in the initial population. In evolutionary fuzzy rule generation, knowledge can also be incorporated by checking the consistency between the rules generated by data and those generated by human experts.


Knowledge Incorporation in Recombination and Mutation


Knowledge Incorporation in selection and reproduction


Knowledge Incorporation in Fitness Evaluations


Knowledge Incorporation through Life-time Learning


Reference

[1] Y. Jin (Editor). Knowledge incorporation in evolutionary computation. Springer, Berlin Heidelberg, 2005

For discussions, please contact Yaochu Jin.