TLDR: DECOY is a novel, data-driven multi-agent simulation environment that uses Counter-Strike: Global Offensive (CS:GO) as a testbed for strategic planning research. It abstracts complex 3D gameplay into a high-level, discretized simulation using a waypoint system and neural predictive/generative models trained on real tournament data. DECOY accurately simulates movement and combat outcomes without modeling low-level mechanics like aiming, achieving significant speed-ups and high fidelity in replicating human gameplay, making it a valuable tool for multi-agent AI research.
In the rapidly evolving world of artificial intelligence, creating intelligent agents that can navigate and strategize in complex, multi-agent environments remains a significant challenge. While AI has achieved superhuman performance in games like StarCraft II and Dota 2, these successes often come at an enormous computational cost, relying on vast datasets and extensive simulation time. This challenge is even more pronounced in real-world scenarios like team sports or emergency response, where high-fidelity simulators are either unavailable or prohibitively expensive.
Addressing this gap, researchers Yunzhe Wang, Volkan Ustun, and Chris McGroarty have introduced DECOY, a novel multi-agent simulation environment designed to facilitate strategic multi-agent planning research. DECOY uses Counter-Strike: Global Offensive (CS:GO), a popular 5v5 first-person shooter, as its testbed, abstracting complex 3D terrains and strategic planning into a high-level, discretized simulation. The core innovation lies in its ability to accurately simulate gameplay using only movement decisions, without explicitly modeling low-level mechanics such as aiming and shooting.
The DECOY Approach: Balancing Fidelity and Efficiency
DECOY’s framework is built on a waypoint system that simplifies and discretizes continuous states and actions. This system allows agents to navigate the 3D environment using a predefined graph of interconnected points, making movement decisions more manageable for AI. To ensure realism, the simulator is paired with neural predictive and generative models, which are trained on real CS:GO tournament data to reconstruct event outcomes like shooting damage and engagement results. This data-driven approach allows DECOY to maintain environmental fidelity while significantly reducing computational complexity.
The simulation environment, developed using Panda3D and the Bullet Physics library, accurately replicates the 3D map geometry of CS:GO’s iconic de_dust2 map. Agent physics, including dimensions, speed, and jumping, adhere to CS:GO specifications. The system processes discrete events with incremental time steps, and the state space includes detailed information for up to 10 agents (position, view angle, weapon, armor, helmet) and the bomb’s status.
Intelligent Damage Modeling
A key innovation in DECOY is its approach to modeling combat outcomes without requiring agents to control precise aiming. This is achieved through two probabilistic models:
- Damage Indicator Predictor (DIP): A binary classifier that determines whether any damage is likely to occur between an attacker and a victim based on contextual game state features.
- Damage Outcome Generator (DOG): If DIP predicts damage, DOG, a Conditional Variational Auto-Encoder (CVAE), generates detailed combat outcomes such as damage amount, death events, and hit locations. The use of a generative model is crucial here, as it captures the complex, multi-factor nature of damage in CS:GO (e.g., weapon type, hit group) without needing to simulate aiming directly.
Performance and Realism
Extensive evaluations demonstrate DECOY’s effectiveness. The simulator achieves approximately a 16x speed-up over real-time performance in a standard 5v5 (10-agent) scenario, making it highly efficient for large-scale training and evaluation. The damage models also show strong performance: DIP achieved an accuracy of 90.8% in predicting damage events, while DOG effectively reconstructed damage outcomes with a Mean Absolute Error (MAE) of 13.97 HP and a hit group classification accuracy of 96.6%. In generative mode, DOG produced statistically accurate damage distributions, closely matching real game data.
Crucially, replays generated from human data in DECOY closely match those observed in the original game. The waypoint-based trajectories align with original human movements with high spatial and temporal fidelity, showing an average Euclidean distance of just 0.443 meters. The simulation also accurately reproduces health progression (0.961 correlation) and death timings (0.809 correlation), with the final match outcome aligning with the original 91.0% of the time. This strong alignment underscores DECOY’s potential as a valuable tool for advancing research in strategic multi-agent planning and behavior generation.
Also Read:
- Reinforcement Learning: The Core Driver for Advanced AI Research Systems
- Optimizing Industrial Scheduling: A Novel DRL Environment for Flexible Job-Shops
Future Directions
While DECOY represents a significant step forward, the authors acknowledge areas for future improvement. These include addressing compounding errors in movement and damage calculations, incorporating longer-term strategic factors like team economy, and modeling utility mechanics such as grenades. Future work will also focus on integrating more expressive temporal modeling methods, such as Graph Neural Networks and Transformers, to capture the complex joint dynamics and temporal relationships inherent in multi-agent engagements.
DECOY is publicly available at github.com/HATS-ICT/decoy, offering a flexible and extensible platform for researchers. This work, detailed in the paper A Data-Driven Discretized CS:GO Simulation Environment to Facilitate Strategic Multi-Agent Planning Research, paves the way for developing more sophisticated AI agents capable of complex tactical decision-making in both virtual and real-world scenarios.
