Several Problems For Evolutionaryalgorithms And Intelligent Data Mining

Nature has always been a rich source of human’s creativity. The human’s abilityof recognizing things often comes from the interaction of nature. Many self-adaptiveoptimization phenomenons from nature constantly give enlightenment on humans.Intelligent algorithms are originated from natural (biological) rules. They is designedto solve some practical problems by means of simulating some rules. They will assigna complex task to the population and employ their cooperation to complete. Theirrelated concepts are easy to understand and it is usually convenient to execute. Due toflexibility and robustness, intelligent optimization algorithms have become a hotresearch area and have been widely used in many fields, such as computer science,knowledge discovery, telecommunication network, robots and so on. In recent decades,some bionic intelligent optimization algorithms, which are entirely different from theclassical mathematical programming principle and attempt to simulate the naturalecological system mechanism to solve the complicated optimization problems, havebeen successively proposed and studied, such as simulated annealing algorithm,genetic algorithm and artificial neural network technology and artificial immunealgorithm and swarm intelligence algorithm, etc. These algorithms greatly enrich themodern optimization technique, and also provide feasible solution for the optimizationproblems which the traditional optimization techniques are difficult to handle.Data mining is the process of extracting hidden, unknown, and potentially usefulinformations and knowledge from a large number of incomplete, noisy, fuzzy andrandom data, where the data can be stored in databases, data warehouses, or otherinformation repositories. The most commonly used data mining technologies containassociation rule, sequential pattern analysis, classification analysis and clusteringanalysis, and so on. Many of them involve optimization models. Therefore, they can besolved and handled by means of intelligent optimization algorithms.It is undoubtedly a challenging task how to combine the data mining techniqueswith the intelligent algorithms. Namely, the data mining technology can serve for theintelligent algorithm, or the intelligent algorithm can be applied into the data mining. Ifthese two kinds of techniques or algorithms can be successfully combined, get rid oftheir own shortcomings, and make full use of mutual advantages, this will be a verygood issue. Meanwhile, it will provide an effective method to mine the massive data,and provide a very good, novel application for the intelligent algorithms. The maincontributions of this thesis can be summarized as follows: 1. In the parallel particle swarm optimization (PSO), Partitioning Around Medoid(PAM) is introduced to to divide the swarm into several non-overlapping sub-swarms.Through clustering, the particles within the same sub-swarm are relative concentrative,so that they can learn each other easier.This makes the limited time be spent on themost effectively searching, so as to improve the searching efficiency of an algorithm.In order to explore the whole solution spaces evenly, uniform design is introduced togenerate an initial population, in which the population members or individuals arescattered uniformly over the feasible solution space. In evolution, the better individualsgenerated using uniform design replace some worse individuals, so as to achieve thesurvival of the fittest.2．In multi-objective particle swarm optimization (MOPSO), to maintain orincrease the diversity of the swarm, and help an algorithm to jump out of the localoptimal solution, PAM clustering algorithm and uniform design are respectivelyintroduced to generate and select the Pareto optimal solutions. A novel algorithm, themulti-objective particle swarm optimization based on PAM and uniform design, isproposed.3．In association rule mining, evaluating an association rule needs to repeatedlyscan database to compare the whole database with the antecedent, consequent of a ruleand the whole rule. In order to decrease the number of comparisons and timeconsuming, we present an attribute index strategy. It only needs to scan database onceto create the attribute index of each attribute. Thereafter, it is unnecessry to scandatabase any more in getting any metric value for evaluating an association rule,instead, this value can be obtained only via the attribute indices. The association rulemining is modeled as a multiobjective rather than a single objective problem. In orderto acquire the solutions which are uniformly scattered toward the Pareto frontier in theobjective space, the elitism policy and uniform design are introduced. This paperproposes attribute index and uniform design based multiobjective association rulemining with evolutionary algorithm. It does not require the user-specified minimumsupport and minimum confidence any more, but uses a simple attribute index. Itdesigns a novel real encoding so as to extend its application scope.4. To measure the performance or validity of clustering algorithms, severalevaluation indices, such as successful rate, successful number and full successful rateare defined. To acquire them correctly, two novel class assignment algorithms aredesigned. One can maximize several proposed evaluation indices; the other can ensureeach cluster to contain at least one vector data. To testify their effectiveness, they are employed to evaluate the performance or validity of several clustering algorithms.5. In order to overcome the premature convergence in PSO, dynamicallycrossover is introduced to PSO, briefly denoted as CPSO. It is introduced to k-meansalgorithm so as to get rid of the drawbacks of simply finding the spherical cluster andbeing sensitive to the initial partitions in k-means algorithm. Finally, the proposedalgorithm is applied to the image segmentation, and acquires very good segmentationresults.