Hello, I work for a software company that specializes in business simulations. We basically allow the user to create a bunch of scenarios and simulate the scenarios, allowing them to compare the results. The scenarios are made up of building blocks called profiles. There are about 10 profiles per scenario. About 5 years ago at Ford Motor Co., a new software company approached my group with a product that could hook up to another software product (simulations) outputs, and try to optimize outputs using a genetic evolution algorithm. I was amazed at how negative a reaction this product received. The software company had to bring in all their phds and were grilled about how does one define "optimal", what case studies has this algorithm solved, etc, etc, etc. Overall, it created in my head, a fear of optimization algorithms and the promise of "nearly optimal", etc. 5 years later, i now want to take a stab at implementing this into my current companies tools. However, I am fearful of possible buyers treating me like Ford did. They will question my educational background, previous experience, etc. So, my questions are: 1. Have times changed? Do you still see this reaction? 2. Should a person who can code, but does not have the AI background attempt to apply it? If yes, what would you recommend I avoid? Thanks and I cant wait to hear your response! Dan

Time will not change anything as their reaction is legitimate.

It''s up to you to convince them. You''ll need to do some experimentation to show the GA is benefitial, and that it does what a normal employee would do -- just better.

Alex

Just to add to alex''s comment.

The concept that drives a genetic algorithm to converge upon a solution is something called a ''fitness function''. How good a GA is at optimising something is completely dependent upon the fitness function. You, the programmer, are responsible for creating your GAs fitness function. Therefore how good your GA will be is dependent on how good you are at defining the problem you want it to solve.

Surely though, if you can demonstrate that a GA can do better than any other method they may have tried, they would show some interest?



My Website: ai-junkie.com | My Book: AI Techniques for Game Programming

Can you cook? The proof is in the pudding.

When I was doing GA research, I received a LOT of negative feedback from traditional Operations Research people in my department. I later found out that their opinions were fairly widespread in the OR community, partly through ignorance and partly through arrogance and partly because they were right...

Their ignorance stemmed from the fact that they didn''t understand GAs sufficiently to be able to criticise them properly, so they tended to criticise the researcher as ignorant (of their ways).

Their arrogance came out in their belief that deterministic optimisation algorithms were always going to out-perform stochastic search methods and that all one needed was the computational power to implement them in reasonable time.

In part, they are correct. GAs aren''t sufficiently understood (even today) to be able to say with confidence how exactly they will perform on a given problem. We can only give heuristic estimates of convergence rate and thus run time and we have no guarantees about not returning sub-optimal solutions.

However, if one understands the problem well and can design a good objective function (some people call it a fitness function), then performance can be evaluated over many trials.

Ultimately, I would say that whether or not you should implement a GA depends on the problem. It''s a blind search problem (in that you only have function evaluations to base the search on as opposed to an analytic solution) so a GA is a reasonable place to start. You might want to consider Simulated Annealing as well.

If you do decide to press ahead with this, let us know. I have some research information you should read about GAs.

Regards,

Timkin