UCSM-DNN: User and Card Style Modeling with Deep Neural Networks for Personalized Game AI

Daegeun Choe, Youngbak Jo, Shindong Kang, Shounan An, Insoo OhAAAI 2022 · DOI

Introduction

In competitive online sports games, maintaining a seamless player experience often requires filling matchmaking gaps with artificial intelligence. However, traditional game AI frequently feels rigid and predictable, lacking the nuanced behaviors of human players. While transfer learning offers a path to personalization by fine-tuning models for individual users, it faces significant scalability challenges—maintaining thousands of distinct models for a large player base is computationally prohibitive.

To address this, researchers at Netmarble developed UCSM-DNN, a unified deep neural network architecture designed to mimic thousands of unique human playing styles. By encoding user behavior and character-specific traits into compact style vectors, UCSM-DNN provides a scalable solution for creating personalized AI agents that retain a "human" feel without the overhead of individual model management.

The UCSM-DNN Framework

The core innovation of UCSM-DNN lies in its ability to condition a general neural network on specific "style vectors." The architecture consists of two primary modeling components:

  1. User Style Modeling (USM): This component generates a User Style Vector (USV) to capture an individual's behavioral patterns. It utilizes binary USVs to represent high-pressure decision-making (e.g., pitch selection on a full count) and distributed USVs to capture general action probabilities.
  2. Card Style Modeling (CSM): Recognizing that players adapt their style based on the character they control, the Card Style Vector (CSV) encodes the specific playing style associated with in-game "cards" (representing real-world athletes).
Figure 1: The UCSM-DNN architecture integrates User Style Modeling and Card Style Modeling to condition a general DNN for personalized game agent behavior.

These style vectors are concatenated with standard game features and processed by a general DNN. This allows the AI to dynamically adjust its actions—such as pitching location or swing timing—based on the specific user-card identity it is currently simulating.

Experimental Evaluation

The model was evaluated using real-world data from the mobile baseball game MaguMagu 2021. The dataset included gameplay logs from 4,950 user pairs collected over four weeks. To quantify the effectiveness of personalization, the authors introduced the Distinguishability of Playing Styles (DPS) metric, which measures how distinctly the AI mimics different users while maintaining consistency for the same user.

Key Results

UCSM-DNN demonstrated superior performance across all evaluated game actions, including pitching, ball positioning, and batting.

Figure 2: Performance comparison of UCSM-DNN against various baselines using the Distinguishability of Playing Styles (DPS) metric.
  • Personalization Gains: The inclusion of style vectors significantly boosted distinguishability. In pitching scenarios, the DPS of UCSM-DNN was 246.3% higher than a standard model without style vectors.
  • Synergy of USM and CSM: The results indicate that modeling both the user and the card is critical. UCSM-DNN outperformed models that relied solely on USM or CSM, confirming that personalization is an interaction between the human player and the character they control.
  • Operational Efficiency: The system is highly optimized for mobile deployment. With only 112K parameters, the model requires just 488KB of memory to support 1,000 users and 2,000 cards, with inference times under 30ms on a standard mobile device.

Conclusion and Future Work

UCSM-DNN represents a significant step forward in scalable, personalized game AI. By treating playing style as a feature space rather than a model-per-user problem, it enables high-fidelity human mimicry in real-time environments.

Future research may focus on addressing "style drift" as player habits evolve over time and extending the architecture to more continuous-motion sports like soccer or basketball, which may require temporal modeling through LSTMs or Transformers. Additionally, investigating "cold start" scenarios—determining the minimum data required to generate an accurate style vector for new users—remains an open area for exploration.

Summary updated

BibTeX
@inproceedings{choe2022ucsm,
  author = {Daegeun Choe and Youngbak Jo and Shindong Kang and Shounan An and Insoo Oh},
  title = {{UCSM-DNN: User and Card Style Modeling with Deep Neural Networks for Personalized Game AI}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21713},
}
Figure