Knowledge Acquisition from Computer Log Files by ADG with Variable Agent Size.

We had previously proposed an outstanding evolutionary method, Automatically Defined Groups (ADG), for generating heterogeneous cooperative agents, and then we had developed a rule extraction algorithm from computer log files using ADG. In this algorithm, agents search multiple errordetection rules cooperatively based on the difference between normal state log files and abnormal state log files. The more frequent applicable and the more accurate the error-detection rule is, the more agents are allocated for searching the rule. Therefore, the number of agents allocated for each rule can represent the important degree of the rule. However, when the rule extraction method was applied to the large scale log files, which may have a number of latent rules, a problematic situation on the number of agents could be observed. In the previous proposed method, the number of agents is not adaptive, therefore the number of agents may be lack for evaluating the each rule's importance minutely. In this paper, we propose an improved method, where the number of agents is adaptively increased depending on the discovered rules. As a result, the importance of respective rules could be evaluated minutely by increasing the number of agents. In addition, the proposed method could acquire more rules than those by the method with the fixed number of agents. [ABSTRACT FROM AUTHOR]