Evolutionary Design Optimization
My Master's thesis deals with the handling of uncertainties in Reliability Based Design Optimization (RBDO) using Evolutionary Algorithms (EAs), for the special case when the probability distributions of the uncertainties are not available.
Though I just mentioned 'special', it is only special since this situation is not generally dealt with in optimizaiton. If fact, practical optimization tasks often are closer to this case, since the uncertainties in the variables/parameters are
seldom available as well defined probability distributions. Most of the time, such distributions are constructed from a large amount of samples, or assumed based on experts' opinion. However, this solution methodology is:
1) Not economical since large number of samples are required, and
2) Not technically sound since the results do not reflect the actual knowledge about the uncertainties.
My work has been based on two approached to deal with this task:
Bayesian approach - Uses Bayesian inference to estimate probability of constraint satisfaction using small sample sets. See the SEAL 2010 paper here.
Evidence based approach: Utilizes expert opinions and available evidence to establish upper bounds on failure probability ('plausibility' in Dempster-Shafer terms). Paper accepted at GECCO 2011.
EAs prove to be very useful in such design scenarios. They handle constraints well, are derivative-free, and can provide well distributed trade-off fronts between objective function and reliability (or plausibility) which can be extremely useful for designers. Also, their parallelization can circumvent the resource-hurdles effectively.
Go back home