Fuzzy Inference System Tsukamoto–Decision Tree C 4.5 in Predicting the Amount of Roof Tile Production in Kebumen

ABSTRACT


A. INTRODUCTION
Tile is in great demand by various groups (Wati, Rohmah, & Rahmadani, 2021). This condition provides many opportunities for producers to implement strategies to optimize their production (Ezhilmaran & Joseph, 2017). In the industrial era 4.0, the application of technology in various ways, can be done to increase the optimization of tile production. One of the processes carried out is predicting the amount of production from producers.
One of the efforts made in a prediction is by using the Fuzzy Logic method (Tundo & Sela, 2018). Fuzzy logic is logic with a value of similarity (fuzziness) with a value of true or false (Sheena, Ramalingam, & Anuradha, 2017). This concept was implemented and introduced by Lotfi Asker Zadeh in 1965 in fuzzy set theory (Solesvik, Kondratenko, & Kharchenko, 2017). There are several types of fuzzy belonging to the type of Fuzzy Inference System (FIS), namely Sugeno, Mamdani, and Tsukamoto. This study uses the Tsukamoto FIS concept -Decision tree C 4.5 which is used to implement. Use of decision tree C 4.5 to create rules that are built based on collected datasets, then processed using WEKA. These two concepts make the flexible, simple structure, tolerant of the data used, and speed up the creation of rules without expert intervention (Tseng, Konada, & Kwon, 2016).
Several studies support this research, including the application of Mamdani FIS in predicting the amount of woven fabric production (Tundo & Saifullah, 2022). Rules are made automatically using Random Tree with the criteria used are production costs, stock, and demand. The resulting accuracy shows values with results that are close to actual production with an accuracy of 97%. But unfortunately not in detail explained how the concept of Random Tree is used. Furthermore, other predictions are also made with the application of FIS in predicting palm oil production (Tundo & 'Uyun, 2020) carried out using the Tsukamoto method. The rules used are the results of the decision tree J48 and REPTree with the criteria: the amount of oil palm, demand, and supply of palm oil. The decision tree shows values with results that are close to actual production. However, the resulting classification accuracy is lower than J48. Furthermore, there is a wind power prediction using the comparison of Fuzzy Mamdani and Sugeno (Topaloglu & Pehlivan, 2018). The criteria used are wind speed, power density, capacity factor, and suitability factor. The experimental results give a better result using the Sugeno method compared to Mamdani. However, no detailed accuracy results have been submitted yet, it should be explained so that it can be confirmed and seen.
Modeling in this study was carried out using decision tree rules C 4.5 which were then processed with WEKA to form a rule. These rules are used to predict tile production using the FIS Tsukamoto method. The existence of C 4.5 rules makes it easier to determine the rules that are built without having to consult with experts because C 4.5 will study existing datasets to serve as a reference in forming these rules according to conditions that often occur. In addition, evidenced by the accuracy presented (Mujahid & Sela, 2019). The modeling results produce relevant rules after being compared with the actual results. In addition, this research can also assist in estimating predictions of tile production which can estimate related losses or profits that will occur.

Create Rules
Before presenting the automatic C 4.5 generation rules using WEKA, here are the general working steps of the C 4.5 algorithm in building decision trees; a. Select a variable as root. b. Create a branch for each value. c. Divide cases into branches. d. Repeat the process for each branch until all cases in the branch have the same class.
The process of making rules using WEKA by changing the production output value in the dataset is a fuzzy set (Sheena et al., 2017). This study divides 3 fuzzy sets consisting of Little, Enough, and Many (Tundo, 2022). The initial process for changing the value of production output by determining the minimum, middle, and maximum values (Tundo & Nugroho, 2020). Then it is assumed that the minimum value until it approaches the midpoint is Little, approaches the midpoint until it approaches the maximum value is Enough and the rest is Many. The following are the maximum rules obtained after experiencing 5x test trials, where the maximum rules obtained produce an accuracy of 80%, as shown in Figure 1. Based on the resulting accuracy and detail accuracy for each class based on Figure 1 above, the rules formed by selecting the C 4.5 visual tree from the process produce rules that look like those in Figure 2. Based on Figure 2, the IF...THEN rule is obtained by writing the classification for each parameter first to make it easier to translate the rules that are formed (Selvachandran et al., 2019). The following classification for each parameter looks like the following.

FIS Tsukamoto
Tsukamoto's FIS is a method in which each consequence is in the form of an IF...THEN rule must be represented in the concept of a fuzzy set with a monotonous membership function (Rahmalia, 2021). As for the stages in this method, there are 3 processes to get the output, namely: a. Formation of Fuzzy Sets This process requires a rule model that must exist Tundo (2020), such as fuzzy sets, and representation of membership functions and domains (Haghpanah & Taheri, 2017). This study implements the concept based on Table 3, which is in the form of rules, criteria in fuzzy sets, and membership function representations as shown in Figure 3 and Figure 3.

Accuracy
Accuracy at this stage is used to measure the success rate of a method that has been used (Hamsa, Indiradevi, & Kizhakkethottam, 2016). In this method, the accuracy calculation uses the Average Forecasting Error Rate (AFER) method with the formula: Where is the actual data in the data and is the predicted value for the i data. Meanwhile, n is the amount of data (Tuan et al., 2020).

C. RESULT AND DISCUSSION 1. Calculation of FIS Tsukamoto
The following is an example of calculating the prediction of the amount of manual production using FIS Tsukamoto based on test data, which will be used as an example of the manual calculation, namely, in August 2022, with a land inventory of 49,000 Kg, a demand of 27,790, and an existing supply of 220.
Step 1: Look for the membership degree value of each criterion based on the modeling that has been made in Table 3 and Figure 3. The following is the value of each membership degree of each criterion. Step 2: The implication function application uses the MIN function, for each rule, to find the a and z values for each rule. Where the value of a and z is a parameter to produce a weighted average. Each Rule will have a and z as many rules as are formed like the following calculation; After all of the test data from May to September 2022 is calculated, predictions for tile production are produced as shown in Table 4.

Analysis of Comparative Results
The predicted results are compared directly with actual production, in detail shown in  The prediction comparison results (Table 4) were tested using the Average Forecasting Error Rate (AFER) method. The values obtained have error and truth values as shown in Table 5 and Figure 5.  Based on the calculation, the error value obtained with AFER is 29.34%, so the accuracy of the truth obtained is 70.66%.

D. CONCLUSION AND SUGGESTIONS
The results of the study show that FIS Tsukamoto -Decision tree C 4.5 can predict tile production at TH ABADI Kebumen. The prediction results with actual production from May to September 2022 using AFER have an error percentage of 29.34% with a truth value of 70.66%. The overall results of predictions on actual production did not exceed anything, so it can be concluded that the FIS Tsukamoto -Decision tree C 4.5 method is sufficiently optimum in providing predictive estimates. It is said to be quite optimum because all customer requests are met, either generated by the production prediction itself or the prediction results are added up with inventory data, and all predictions are close to actual production. This research also shortens the time in making rules because it is sufficiently processed with decision tree C 4.5 using WEKA, so there is no need for expert intervention. The rules that are formed from decision tree C 4.5 using WEKA can be accounted for due to the resulting accuracy, where the accuracy of the rule formation reaches 80%. Suggestions for future researchers can make comparisons with other FIS, namely Mamdani and Sugeno, as well as the decision tree that is formed can use other than C 4.5 to find out the differences in each of the methods used.