The Symmetric Pattern Fuzzy Discretizationin Predicting Plastic Type for a Sorting System Using Decision Tree Methods

Plastics made from petroleum as the main ingredient cannot decompose quickly like organic materials, but rather take 500 years to 1000 years. Recycling plastic waste can significantly reduce the negative impacts that damage the environment. In addition, it can also reduce the use of natural resources as raw materials and energy use in the extraction process from mining to ready-for-use. In the process of recycling plastic bottle waste in industry, a sorting system is needed to sort the types of plastic bottle waste automatically, effectively, and efficiently. The system needs a prediction model with satisfactory performance. This research aims to build the prediction models of a plastic bottle waste sorting system with a fuzzy approach using the Decision Tree method. The main focus is fuzzy discretization with asymmetric pattern where the membership functions for the first and the last categories are balanced. Seven Decision Tree models are proposed in this study, six models with symmetric fuzzy discretization and one model with crisp discretization as a comparison. Three types of plastics are the objects of the study, namely Polyethylene Terephthalate (PET/PETE), High-Density Polyethylene (HDPE), and Polypropylene (PP). All three are types of plastics that are widely found in household waste. The unique contribution of this paper is that the symmetric pattern fuzzy discretization in the decision tree method can improve the performance of the decision tree model, but the combination of fuzzy membership functions used also contributes. Not all combinations used can improve the performance of the prediction model. Four of six models of symmetric fuzzy discretization have better performance than the decision tree model with crisp discretization. The combination of fuzzy membership functions consisting of linear and triangular functions provides the highest performance. Two models that do not perform better than crisp discretization are the linear-trapezoidal combination and the linear-Gaussian combination.