Penelitian ini bertujuan mengimplementasikan Artificial Neural Network (ANN) sebagai mekanisme pengambilan keputusan atau decision making pada Non-Player Character (NPC) customer dalam game simulasi toko kelontong UMKM. Permasalahan utama yang diangkat adalah keterbatasan pendekatan deterministik seperti Finite State Machine dan rule-based decision yang cenderung kaku ketika kondisi permainan menjadi lebih dinamis. Penelitian ini menggunakan pendekatan supervised learning dengan arsitektur feedforward ANN atau Multi-Layer Perceptron. Dataset dibentuk secara sintesis melalui simulasi Unity menggunakan Mode GURU sebagai teacher policy yang menghasilkan label keputusan berbasis rule-based heuristic. Perekaman data dilakukan pada dua decision point, yaitu keputusan NPC untuk masuk atau tidak masuk toko dan keputusan NPC untuk melanjutkan atau membatalkan antrean kasir. Dataset akhir terdiri dari 1.227 sampel pada decision point masuk toko dan 403 sampel pada decision point antre kasir. Arsitektur ANN yang digunakan adalah 14-8-8-2 untuk keputusan masuk toko dan 10-5-2 untuk keputusan antre kasir. Evaluasi analitik dilakukan menggunakan Google Colab melalui confusion matrix, accuracy, precision, recall, F1-score, Matthews Correlation Coefficient, Cohen’s Kappa, ROC-AUC, Stratified 5-Fold Cross Validation, serta baseline comparison terhadap Decision Tree, Random Forest, dan XGBoost. Hasil evaluasi hold-out validation menunjukkan bahwa ANN/MLP memperoleh accuracy sebesar 91,06% pada decision point masuk toko dan 91,36% pada decision point antre kasir, dengan ROC-AUC masing-masing sebesar 0,9775 dan 0,9832. Hasil baseline menunjukkan bahwa model tree-based memperoleh performa lebih tinggi pada dataset berbasis aturan, tetapi ANN tetap berhasil diimplementasikan sebagai sistem decision making NPC pada Unity. Penelitian ini berkontribusi pada penerapan ANN ringan untuk NPC customer pada game simulasi bisnis berbasis dataset sintesis expert demonstration.


Deskripsi Lain

This study aims to implement an Artificial Neural Network (ANN) a decision-making mechanism for customer Non-Player Characters (NPCs) in an MSME (Micro, Small, and Medium Enterprises) grocery store simulation game. The primary problem addressed is the limitation of deterministic approaches, such Finite State Machines and rule-based decisions, which tend to be rigid in dynamic game environments. This research employs a supervised learning approach using a feedforward ANN Multi-Layer Perceptron (MLP) architecture. The dataset was synthetically generated through Unity simulations using a "GURU Mode" the teacher policy, producing decision labels based on rule-based heuristics. Data recording was conducted at two decision points: the NPC's decision to enter avoid the store, and the decision to either proceed with abandon the cashier queue. The final dataset comprises 1,227 samples for the store entry decision and 403 samples for the queueing decision. The applied ANN architectures were 14-8-8-2 for the store entry decision and 10-5-2 for the queueing decision. Analytical evaluation was conducted via Google Colab using confusion matrices, accuracy, precision, recall, F1-score, Matthews Correlation Coefficient (MCC), Cohen’s Kappa, ROC-AUC, and Stratified 5-Fold Cross-Validation, alongside baseline comparisons against Decision Tree, Random Forest, and XGBoost models. Hold-out validation results demonstrated that the ANN/MLP model achieved an accuracy of 91.06% for the store entry decision and 91.36% for the queueing decision, with ROC-AUC scores of 0.9775 and 0.9832, respectively. Although baseline comparisons indicated that tree-based models yielded higher performance on this rule-based dataset, the ANN was nonetheless successfully implemented the NPC decision-making system within the Unity engine. This research contributes to the application of lightweight ANNs for customer NPCs in business simulation games, utilizing synthetic datasets derived from expert demonstrations.