Home 
 1. CEC'2002 FDEC Workshop 
 Session 1: Presentations 
 Session 2: Discussions  
 Research Groups Posters 
 2. CEC'2003 FDEC Workshop 
 Session 1: Presentations 
 Session 1: Identified Themes 
 Session 2: Discussions 
 3. CEC'2004 FDEC Workshop 
 Discussions of Themes 
 4. CEC'2005 FDEC Workshop 
 Presentation 
 5. WCCI 2006 ECTC Panel Session 
 White Paper (rev2.5) 
 Presentations 
 Discussions 
 6. CEC'2007 FDEC Workshop 
Session 2: Discussions
Workshop on Future Direction in EC
CEC’2002
Session 2 – Discussions
Notes

Eiben
    • Unable to create generalize results.
    • There are not many improvements and experimental results that could claim that selection of test functions and the results are generalizable.
    • Need a good methodology for experimental research.
    • Need similarity definition for test suits.
    • Use machine learning approach – separate the training set and test set.
    • Use statistical approach for analyzing the data.

Thomas English
    • Facing the problem of different techniques that are doing different things.
    • Need a much more mature and scientific approach, e.g., characterization of algorithms instead of championing the algorithms.
    • Find meaningful problem classes and identify which algorithms are suitable for which classes.
    • Proposal for a special session in CEC 2003 – attempt to characterize the algorithms and problem classes. Session chair introduces a few selected problem instances and authors will present results for their techniques. Also, authors need to submit their codes and session chair will run the codes against new problem instances and show the results for these new problem instances where authors could not tune their algorithms to.

Peter Angeline
    • Need to know what steps to be taken to develop the unified framework in EC?
    • EC is currently a collection of techniques and trying very hard to be a field of study.
    • Darwinian’s idea is an algorithm and has nothing to do with the mechanism of DNA.
    • We have algorithms, and all the particular classes of mutation, selection and etc., is immaterial to this algorithm – should think in that level and making bridges at that level.

Question – will there be a real world problem for the session?

Eiben
Is that practical to have real-world problem in the special session since real- world problem is difficult to guarantee similarity of the problem and results produced?

Nik Kosabov
    • EC will stay but should not stay as it is.
    • Need to be improved further, e.g., capable of working in a dynamic environment.
    • One good example of real-world problem is life sciences applications.
    • EC has not been applied in the area of life sciences, although a lot other techniques like NN, clustering techniques have been used.

Audience
    • Customers are only interested in solutions - not the techniques or any particular technology that we have.
    • In the process of finding solutions for complex problem, iterations and fine-tuning are often needed.
    • We need to know what type of problem needs what kind of tools.

Thomas English
Theoretical bases are that it is very difficult to tell what kind of tools is for what kind of problems.

Rob Shipman
Should consider the question more broadly, e.g., which optimization techniques including heuristics approaches (instead of which EC) is suitable for which problems.

Nik Kosabov
    • Solving complex problems needs more than just EC – hybridization of EC.
    • Optimization process includes implementation and learning.
    • Need better integration of EC with NN, machine learning, symbolic approaches.

Bob Mackay
    • EC may not be just an area that includes a collection of tools that could be used for solving problems.
    • Only two techniques available for solving relational problem – inductive logic programming and genetic programming but none of them could do it well.

Xu JianXin
    • In control research area, there is a systematic approach for comparing different algorithms.
    • EC research is very diversify - everyone proposes their own algorithm and compare it with others.
    • Could we come up with some systematic approach for comparing the algorithms?

Peter Angeline
We are not the only group facing the problem, e.g., nonlinear system conference.

Rob Shipman
    • We need to be clear on what tools to be used for what applications.
    • One particular area is working in the dynamic environment where the feedback signal is not static.

Nik Kosabov
    • Ken’s framework is only on integration within existing EC – we need to integrate EC to other techniques like machine learning, symbolic manipulation.
    • Integration should be more general and more sophisticated than existing level.

Ali Zalzala
    • The field of GA research can only go for a few years? – Obviously it is wrong.
    • We have a big field on EC research now – what are the basis of what we are at the moment. Need to know those bases so that we could build up on that to move forward.

Eiben
What is the niche of EC?
Deb - Multi-objective problems?

Audience
One niche is complex problems – couldn’t solve the problem yet even with GP.

Hans-Paul Schwefel
    • ES was introduced because traditional approaches couldn’t solve his problem, e.g., wind tunnel optimization problem.
    • ES was first introduced for discrete parameters. Later ES was applied for continuous parameters because theory was easier to be developed for real parameters.

Ali Zalzala
Is NN a collection of tools? Do they think of that way?

Nik Kosabov
EC was used to learn parameters in NN online in adaptive process control for speed recognition.

Audience
The difference between EC and NN is that the fitness function in EC is unique.

Audience
Explore learning aspect for EC.

Thomas English
Techniques exist for improving learning in EC, e.g., via mutation operator.

Hans-Paul Schwefel
There is a lack of theory in EC.

Audience
Theory could provide the structure for EC research, and needs to be problem driven – rather than for a particular principle.

Audience
    • Outsider couldn’t figure out which techniques to be chosen in EC, e.g., GA, ES, GP or others.
    • One practical issue is the fitness function evaluation – it could take a very long time to evaluate the fitness function for real-world optimization problem.

Thomas English
A unified approach in EC could be useful for researchers in other fields.

Xu JianXin
EC is not only a nonlinear system but involve the stochastic process, which makes it very difficult to be analyzed.

Peter Angeline
EC is just a training method, like other methods such as BP or gradient- based.

Audience
Depending on the fitness function, BP can be used for simple shape fitness function. For real world problem - fitness noisy, multi-dimensional, discontinuity – EC is good for it.

Nik Kosabov
It is difficult for NN to evolve a structure in a dynamic environment or in an incremental way and to evolve the NN to suit it, e.g., dynamic speed recognition problem. Need EC or better techniques to solve this on-line optimization or on-going learning problem.

Audience
How to get a good evaluation of the fitness function in a short time during the on-line or real-time process?

Nik Kosabov
Use enough examples to start with. While the environment changes, more and more examples emerge and then optimize the NN structure on-line. We can have a set of NN running while the other one is improving, and select the best NN from the population, e.g., keep the NN improving while the control process is going.

Audience
For some of the systems, it is difficult to construct the fitness function until the optimization process starts, e.g., combat games.

Rob Shipman
In some problems, we are only interested in the mappings, and there has been less emphasis on the representation and what that does to subscribe in a higher algorithm to move around it.

Thomas English
The best solution is often a collection of solutions from the population – a two-state chaotic system.

Ali Zalzala
How could we communicate what we discussed here to students?

Audience
    • Should offer students more on the inside of biology so that they understand the wheel of evolution in more details.
    • Cross linking to biology and computational science.

Nik Kosabov
    • Researchers should learn more biology as well, e.g., brain study, evolution etc.
    • This is also a good application area and good for funding.

Audience
    • Many complex problems need tools and thinking beyond individual technique.
    • Generalized education - Symbolic AI, fuzzy, expert system, evolutionary computation, NN and etc – for Ph.D. students.
    • Use hands on applications – let students try for themselves and feel what is ready an EC technique.

Audiences
    • Need to collaborate with other communities.
    • Courses should be cross disciplinary, e.g., include computational sciences and biology.
    • Other view - biology may complicate the situations.
    • To show how EC can solve the problems faced by biologist or neurologist, which they don’t know how to solve.
    • Strong interest from students for combining techniques in biology with computer science.

Nik Kosabov
    • A future work is to combine genetic level with high level neural network process, e.g., bringing genetic research to NN conferences should be done in the future - there is a lot of interaction between these two areas.

Audience
    • Any other conference doing this kind of workshop? – GECCO did it occasionally.

Eiben
    • EvoNet is offering a wide range of activities and shared materials.
    • An excellent Website has also been set up for this purpose.

Audience
In US, independent form is competing for survival and needs to keep their materials proprietary – don’t want to collaborate and give away latest developments.

Ali Zalzala
European parliament will announce how the funding is to be distributed in September. They have requested researchers to make expression of interest. EvoNet has a white board where we could provide this information and they would link these together and perhaps make bids on behalf of the community. The funding is available globally and there are special interest coming from different parts. A bid could be put up for funding to do what EvoNet is currently doing in Europe.
Top