Developing a better hybrid learning algorithm using evolutionary algorithm and decision tree

Abstract

TIlls thesis presents a new hybrid algorithm for decision tree construction. The proposed algorithm exploits the strengths of Evolutionary Algorithms and Simulated Annealing to come up with a new multiobjective population-oriented simulated annealing technique for decision tree construction that performs well on standard datasets of va repository. We have investigated the issues related to decision tree construction, different evolutionary techniques and simulated annealing technique. A brief survey of the existing greedy algorithms and EA based hybrid systems for decision tree construction has also been incorporated in this thesis. The new hybrid algorithm, as presented in this thesis, explores the search space structurally and sequentially, other than using genetic operators like mutation or crossover to breed new decision trees. To avoid a complete search, the new algorithm utilizes simulated annealing technique and restrains from searching regions of the search space having a very low probability of containing an optimal solution. On the other hand, to make the search process faster and parallel, separate sets of search agents (populations of decision trees) are maintained for the current search points aild the suboptimal solutions that have so far been discovered. Finally the new system has been compared with the most recent genetic algorithm based hybrid systems like GALE and GATree as well as the traditional greedy heuristic based system, 01.5. The experimental results reveal the effectiveness and superiority of the new system over the above-mentioned systems.