A Hybrid of Ant Colony Optimization Algorithm and Simulated Annealing for Classification Rules

Ant colony optimization (ACO) is a metaheuristic approach inspired from the behaviour of natural ants and can be used to solve a variety of combinatorial optimization problems. Classification rule induction is one of the problems solved by the Ant-miner algorithm, a variant of ACO, which was initiated by Parpinelli in 2001. Previous studies have shown that ACO is a promising machine learning technique to generate classification rules. However, the Ant-miner is less class focused since the rule’s class is assigned after the rule was constructed. There is also the case where the Ant-miner cannot find any optimal solution for some data sets. Thus, this thesis proposed two variants of hybrid ACO with simulated annealing (SA) algorithm for solving problem of classification rule induction. In the first proposed algorithm, SA is used to optimize the rule's discovery activity by an ant. Benchmark data sets from various fields were used to test the proposed algorithms. Experimental results obtained from this proposed algorithm are comparable to the results of the Ant-miner and other well-known rule induction algorithms in terms of rule accuracy, but are better in terms of rule simplicity. The second proposed algorithm uses SA to optimize the terms selection while constructing a rule. The algorithm fixes the class before rule's construction. Since the algorithm fixed the class before each rule's construction, a much simpler heuristic and fitness function is proposed. Experimental results obtained from the proposed algorithm are much higher than other compared algorithms, in terms of predictive accuracy. The successful work on hybridization of ACO and SA algorithms has led to the improved learning ability of ACO for classification. Thus, a higher predictive power classification model for various fields could be generated.