Computer Animation and Virtual Worlds Software
Learning to Play Like Humans: A Deep Neural Framework for Believable Agent Development in Virtual Worlds
Abstract
ABSTRACT Many existing intelligent agents still behave in a mechanical way leading to user dissatisfaction and lack of engagement. To create more realistic and accurate imitations, this study proposes an imitation learning framework that explores deep learning (DL) and traditional machine learning, to build a computational model. The approach is tested using two case studies: Pac‐Man to predict human behavior, and Player Unknown's Battlegrounds (PUBG) to evaluate how well the model could imitate it. Recurrent neural networks (RNN), gated recurrent units, long short‐term memory networks (LSTM), random forest, extreme gradient boosting (XGBoost), etc. To assess model performance, mean squared error, root mean squared error, and accuracy are used. Among the employed models, RNN and LSTM achieved near‐identical and highest accuracy on Pac‐Man (99.98% and 99.96% on the combined dataset), clearly ahead of GRU, random forest and XGBoost, while XGBoost performed best on PUBG with 97.37% accuracy. A dedicated discussion explains why RNN performs well in the low‐dimensional, short‐horizon Pac‐Man setting and lists its advantages and limitations. The Mann–Whitney U Test and the Cosine similarity tests are also applied to measure how believable the imitated behaviors were. Neither analysis detected a statistically significant difference between human and model traces, and cosine similarity was approximately 0.98.