Revising Fuzzy Genetic Algorithms: Rule Generation and a Cached Nearest-Neighbor Strategy
Name
Oleksandr Syzonov
Abstract
Genetic algorithms (GA) can be more helpful in finding global minima/maxima than other methods, like gradient descent or random search, especially for nondifferentiable functions with lots of local minima and maxima. One of the drawbacks of standard GA methods is that lots of hyperparameters need to be set, and selection pressure is applied based on complex rules as opposed to more intuitive fuzzy rules. Variants of genetic algorithms that adjust such parameters via fuzzy logic, to make parameter update principles more interpretable, constitute the class of fuzzy genetic algorithms (FGAs).
This thesis proposes modifications to two relatively recent Fuzzy Genetic Algorithms (FGAs) with N-terms and auto-generated rules, along with a computational optimization aimed at improving simulation runtime. The modifications are evaluated on benchmark functions (Ackley, Griewank, Rastrigin, and Schwefel) and the best setting for each modified method is selected (i.e. membership functions, number of terms, t-norm and t-conorm). The outcome is compared to standard GA and particle swarm optimization (PSO). The results showed that FGA methods could be optimized using caching and nearest-neighbor methods without losing accuracy and convergence. Both of the modified (and unmodified) methods were shown to perform statistically significantly worse than the baseline methods. As a result, we have proposed two optimizations of the existing two algorithms: extrapolation with rule generation and nearest-neighbor estimate with caching and tested their performance.
This thesis proposes modifications to two relatively recent Fuzzy Genetic Algorithms (FGAs) with N-terms and auto-generated rules, along with a computational optimization aimed at improving simulation runtime. The modifications are evaluated on benchmark functions (Ackley, Griewank, Rastrigin, and Schwefel) and the best setting for each modified method is selected (i.e. membership functions, number of terms, t-norm and t-conorm). The outcome is compared to standard GA and particle swarm optimization (PSO). The results showed that FGA methods could be optimized using caching and nearest-neighbor methods without losing accuracy and convergence. Both of the modified (and unmodified) methods were shown to perform statistically significantly worse than the baseline methods. As a result, we have proposed two optimizations of the existing two algorithms: extrapolation with rule generation and nearest-neighbor estimate with caching and tested their performance.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Stefania Tomasiello
Defence year
2023