Imitation of Human Behavior in 3D-Shooter game

Ilya Makarov, Mikhail Tokmakov, Lada TokmakovaAIST (Supplement) 2015 · OpenReview

Abstract

Traditional artificial intelligence in first-person shooters (FPS) often relies on perfect information from the game engine, leading to robotic and unfair behavior. This research proposes a framework that mimics human limitations to improve behavioral realism. By integrating image-based visual recognition, statistical "danger zone" prediction, and a human-like motor reflex model, the authors developed a BOT that perceives and interacts with its environment more like a human player.

Introduction

In competitive 3D shooters, game AI (BOTs) typically "cheat" by accessing exact player coordinates through the game engine. While effective, this results in perfect aim and zero reaction time, which developers often counteract by artificially reducing the BOT's health or accuracy. This approach creates an "uncanny valley" where BOT behavior feels predictable and unnatural. This paper addresses the gap in behavioral realism by imposing human-like constraints on the AI, focusing on visual perception delays and physical motor errors.

Methodology

The system was implemented in Unreal Engine 4 using a multi-layered architecture designed to simulate human cognitive and physical processes.

Visual Recognition and Perception

Instead of using internal engine data, the BOT processes the visual scene. The researchers evaluated several machine learning models, including Bayesian Classifiers, Neural Networks, and Support Vector Machines (SVM). A critical challenge identified was the "Shadow Filter" problem, where the AI initially confused player shadows with the players themselves.

The visual recognition system identifies characters and filters out shadows based on floor-plane geometry.

To solve this, the visual unit was linked to the geometry unit; if a detection occurs strictly on the floor plane, it is classified as a shadow and ignored, mirroring human visual filtering.

Geometry-Based Attention

Humans do not scan environments uniformly; they focus on high-probability areas like doorways and corners. The framework uses a priority queue based on the 3D environment's geometry to calculate "danger zones."

Statistical priority map highlighting 'danger zones' where enemies are most likely to appear.

By focusing visual processing power on these zones, the BOT simulates human situational awareness, making its movements appear intentional rather than random.

Rule-Based Decision Making (RBS)

High-level strategic decisions are governed by a Rule-Based System. This logic tree allows the BOT to weigh factors like health and enemy presence to decide between aggressive or defensive actions. The authors also explored using Neural Networks to modulate these choices based on simulated "psychological temper" or team-oriented goals.

Motor Reflex Modeling

To avoid instantaneous aiming, the authors modeled the physical act of aiming in three human-like phases:

  1. Inertia: The initial delay during target recognition.
  2. Instant Aiming: Rapid movement toward the target.
  3. Braking/Correction: Final micro-adjustments to align the crosshair.

Aiming error is treated as a feature, with accuracy decreasing as rotation speed or target angle increases, mimicking the difficulty of "flick shots" compared to tracking slow-moving targets.

Results

The study utilized custom datasets from multiplayer game sessions to train the visual models. The combination of Sparse Network of Winnows (SNoW) and Classifier Boosting achieved a 94% correct detection rate. Furthermore, SVM and SNoW provided the necessary speed for real-time performance. The Rule-Based System successfully enabled the BOT to adapt its playstyle dynamically, switching between survival-focused and objective-focused behaviors.

Conclusion

This framework demonstrates that incorporating human limitations—rather than perfect logic—is key to creating realistic game AI. Future work will focus on handling environmental occlusion (e.g., partially visible targets), managing multiple simultaneous threats, and extending the model to multi-story environments with vertical gameplay.

Summary updated

BibTeX
@inproceedings{makarov2015imitation,
  author = {Ilya Makarov and Mikhail Tokmakov and Lada Tokmakova},
  title = {{Imitation of Human Behavior in 3D-Shooter game}},
  booktitle = {AIST (Supplement)},
  year = {2015},
  url = {https://openreview.net/forum?id=-Mm5XAvqpu},
}
Figure