An ant colony optimization-based classifier system for bacterial growth

Abstract

Motivation: In predictive microbiology, identification of different combination of environmental factors (such as temperature, water activity, pH), which lead to growth/ no-growth of microorganism, is a problem of potential importance. Ant colony optimization (ACO) is one of the most recently developed nature-inspired metaheuristic techniques, based on the foraging behavior of real life ants and has already exhibited superior performance in solving combinatorial optimization problems. This work explores the search capabilities of this metaheuristic for learning classification rules in bacterial growth/no growth data pertaining to pathogenic Escherichia coli R31 as affected by temperature and water activity. The discovered rules thus can be used to verify whether any combination of temperature and water activity belong to either growth or no-growth of the microorganism. Method: The ant algorithm for classification works iteratively as follows: At any iteration level, software ants construct rules using available heuristic information and dynamically evolved pheromone trails. A rule that has highest prediction quality is said to be a discovered rule, which represents information extracted from the database. Examples correctly covered by the discovered rule are removed from the training set, and another iteration is started. Guided by the modified pheromone matrix, the agents build improved rules and the process is repeated for as many iterations as necessary to find rules covering almost all cases in the training set. Results: The developed ACO classifier system is utilized on several datasets and its performance is compared with the performance of other well known algorithms in terms of the average accuracy attained in 10-fold cross validation. The results obtained by this algorithm compare very favorably with other classifiers. Additionally, for discovery of classification rules in the dataset pertaining to bacterial growth/no-growth, the performance of the ACO classifier is compared with the C4.5 system with respect to the predictive accuracy and the simplicity of discovered rules. In both these performance indices the ACO classifier compares very well with the C4.5. Conclusions: The results obtained on several data sets indicate that the algorithm is competitive and can be considered a very useful tool for knowledge discovery in a given database.