Year: 2020
Type: dissertation
Abstract: We focus on the clonal selection inspired computational models of the immune system developed for general-purpose optimisation. \nOur aim is to highlight when these artificial immune systems (AIS) are more efficient than evolutionary algorithms (EAs). \nCompared to traditional EAs, AIS use considerably higher mutation rates (hypermutations) for variation, give higher selection probabilities to more recent solutions \nand lower selection probabilities to older ones (ageing). We consider the standard Opt-IA that includes both of the AIS distinguishing features and argue why it is of greater applicability than other popular AIS. Our first result is the proof that the stop at first constructive mutation version of its hypermutation operator is essential. \nWithout it, the hypermutations cannot optimise any function with an arbitrary polynomial number of optima. Afterwards we show that the hypermutations are exponentially faster \nthan the standard bit mutation operator used in traditional EAs at escaping from local optima of standard benchmark function classes and of the NP-hard Partition problem. \nIf the basin of attraction of the local optima is not too large, then ageing allows even greater speed-ups. For the \n Cliff benchmark function this can make the difference between exponential and quasi-linear expected time. \nIf the basin of attraction is too large, then ageing can implicitly detect the local optimum and escape it by automatically restarting the search process. \nThe described power of hypermutations and ageing allows us to prove that they guarantee (1+epsilon) approximations for Partition in expected polynomial time for any constant epsilon. \nThese features come at the expense of the hypermutations being a linear factor slower than EAs for standard unimodal benchmark functions and of eliminating the power of ageing at escaping local optima in the complete Opt-IA. \nWe show that hypermutating with inversely proportional rates mitigates such drawbacks at the expense of losing the explorative advantages of the standard operator. \nWe conclude the thesis by designing fast hypermutation operators that are provably a linear factor faster than the traditional ones for the unimodal benchmark functions and Partition, while maintaining explorative power and working in harmony together with ageing.
Author: D. Yazdani
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3083329625