Selected work

Representative Papers

FootsiesGym: A Fighting Game Benchmark for Two-Player Zero-Sum Imperfect-Information Games

Jul 07, 2026

This work addresses the lack of efficient, open-source reinforcement learning environments tailored to nontrivial two-player zero-sum imperfect-information games. It introduces Footsies, an open-source RL environment based on a minimalist 2D fighting game that, for the first time, distills core strategic mechanisms from real fighting-game neutral phases—such as cyclicity and non-transitivity—into an analytically tractable and computationally efficient benchmark. Implemented in Python with a vectorized simulator, Footsies enables high-throughput training on standard hardware and includes integrated support for systematic evaluation across multiple reinforcement learning algorithms. By balancing strategic complexity with scalability, Footsies establishes a reproducible and accessible benchmark for advancing research in imperfect-information games.

0 citationsRead paper

EMAgnet: Parameter-Space EMA Regularization for Policy Gradient Self-Play in Large Games

Jun 22, 2026

This work addresses the inefficiency of exploration in existing policy gradient self-play methods for large-scale two-player zero-sum imperfect-information games, which rely on uniform distribution regularization and struggle to distinguish between superior and inferior actions. To overcome this limitation, the authors propose EMAgnet, a novel approach that employs an exponential moving average (EMA) of the policy network parameters as an adaptive regularization target. This dynamic prior effectively tracks the evolving policy during training, replacing the static uniform prior. Integrated within a PPO-based self-play framework and enhanced with linear and power-law annealing schedules, EMAgnet significantly reduces exploitability across multiple benchmark and high-exploration-difficulty game environments. Notably, it demonstrates superior stability and performance in settings containing a large number of strictly dominated strategies.

0 citationsRead paper

Human-like Bots for Tactical Shooters Using Compute-Efficient Sensors

Dec 30, 2024

Addressing the industry challenges of high computational overhead and insufficient behavioral realism in AI opponents for tactical shooter games (e.g., *VALORANT*), this paper proposes a lightweight, pixel-free anthropomorphic AI framework. Methodologically, it replaces image-based perception with sparse ray-casting sensors for environment representation and integrates human player trajectory imitation via a compact neural network, enabling real-time 2v2 decision-making on CPU-only hardware. Key contributions include: (1) the first end-to-end anthropomorphic imitation learning framework based solely on ray-based perception deployed in a commercial-grade tactical shooter; (2) over 90% reduction in inference cost compared to conventional vision-driven AI; and (3) human evaluation confirming behavioral naturalness statistically indistinguishable from professional players (*p* < 0.01), significantly enhancing adversarial realism and deployment practicality.

0 citationsRead paper

Recent publications

Latest Papers

FootsiesGym: A Fighting Game Benchmark for Two-Player Zero-Sum Imperfect-Information Games

Jul 07, 2026

This work addresses the lack of efficient, open-source reinforcement learning environments tailored to nontrivial two-player zero-sum imperfect-information games. It introduces Footsies, an open-source RL environment based on a minimalist 2D fighting game that, for the first time, distills core strategic mechanisms from real fighting-game neutral phases—such as cyclicity and non-transitivity—into an analytically tractable and computationally efficient benchmark. Implemented in Python with a vectorized simulator, Footsies enables high-throughput training on standard hardware and includes integrated support for systematic evaluation across multiple reinforcement learning algorithms. By balancing strategic complexity with scalability, Footsies establishes a reproducible and accessible benchmark for advancing research in imperfect-information games.

0 citationsRead paper

EMAgnet: Parameter-Space EMA Regularization for Policy Gradient Self-Play in Large Games

Jun 22, 2026

This work addresses the inefficiency of exploration in existing policy gradient self-play methods for large-scale two-player zero-sum imperfect-information games, which rely on uniform distribution regularization and struggle to distinguish between superior and inferior actions. To overcome this limitation, the authors propose EMAgnet, a novel approach that employs an exponential moving average (EMA) of the policy network parameters as an adaptive regularization target. This dynamic prior effectively tracks the evolving policy during training, replacing the static uniform prior. Integrated within a PPO-based self-play framework and enhanced with linear and power-law annealing schedules, EMAgnet significantly reduces exploitability across multiple benchmark and high-exploration-difficulty game environments. Notably, it demonstrates superior stability and performance in settings containing a large number of strictly dominated strategies.

0 citationsRead paper

Human-like Bots for Tactical Shooters Using Compute-Efficient Sensors

Dec 30, 2024

Addressing the industry challenges of high computational overhead and insufficient behavioral realism in AI opponents for tactical shooter games (e.g., *VALORANT*), this paper proposes a lightweight, pixel-free anthropomorphic AI framework. Methodologically, it replaces image-based perception with sparse ray-casting sensors for environment representation and integrates human player trajectory imitation via a compact neural network, enabling real-time 2v2 decision-making on CPU-only hardware. Key contributions include: (1) the first end-to-end anthropomorphic imitation learning framework based solely on ray-based perception deployed in a commercial-grade tactical shooter; (2) over 90% reduction in inference cost compared to conventional vision-driven AI; and (3) human evaluation confirming behavioral naturalness statistically indistinguishable from professional players (*p* < 0.01), significantly enhancing adversarial realism and deployment practicality.

0 citationsRead paper