The video game industry is rapidly growing, increasing the focus on user experience (UX) and player retention. A common approach is Dynamic Difficulty Adjustment (DDA), which adapts game difficulty based on player performance to maintain the flow state. Rule-based DDA implementations are limited in scalability and require extensive manual effort. To address this, we propose using Neural Networks to infer the game state and suggest real-time difficulty adjustments. A roguelike game was developed as a test environment, and gameplay sessions with 38 players were conducted to build the training dataset. The trained model was able to generate difficulty adjustments consistent with the game’s intended design logic. It reached an R \(^2\) of 0.887 on the test set, demonstrating high accuracy and robustness in performance prediction.