Understanding how Knowledge is exploited in Ant Algorithms
McCallum, Thomas Edward Reid
Ant algorithms were first written about in 1991 and since then they have been applied to many problems with great success. During these years the algorithms themselves have been modified for improved performance and also been influenced by research in other fields. Since the earliest Ant algorithms, heuristics and local search have been the primary knowledge sources. This thesis asks the question "how is knowledge used in Ant algorithms?" To answer this question three Ant algorithms are implemented. The first is the Graph based Ant System (GBAS), a theoretical model not yet implemented, and the others are two influential algorithms, the Ant System and Max-Min Ant System. A comparison is undertaken to show that the theoretical model empirically models what happens in the other two algorithms. Therefore, this chapter explores whether different pheromone matrices (representing the internal knowledge) have a significant effect on the behaviour of the algorithm. It is shown that only under extreme parameter settings does the behaviour of Ant System and Max-Min Ant System differ from that of GBAS. The thesis continues by investigating how inaccurate knowledge is used when it is the heuristic that is at fault. This study reveals that Ant algorithms are not good at dealing with this information, and if they do use a heuristic they must rely on it relating valid guidance. An additional benefit of this study is that it shows heuristics may offer more control over the exploration-exploitation trade-off than is afforded by other parameters. The second point where knowledge enters the algorithm is through the local search. The thesis looks at what happens to the performance of the Ant algorithms when a local search is used and how this affects the parameters of the algorithm. It is shown that the addition of a local search method does change the behaviour of the algorithm and that the strength of the method has a strong influence on how the parameters are chosen. The final study focuses on whether Ant algorithms are effective for driving a local search method. The thesis demonstrates that these algorithms are not as effective as some simpler fixed and variable neighbourhood search methods.