Title: Achieving Human-Like Goalkeeping in Realistic Football Simulations via a Sample-Efficient Reinforcement Learning Framework

Abstract:

Although high-profile video games have frequently functioned as testbeds for Deep Reinforcement Learning (DRL), the gaming industry has seldom utilized these techniques to develop authentic artificial intelligence behaviors. Prior studies have predominantly concentrated on training super-human agents using extensive models, a strategy that remains impractical for game studios with constrained resources that prioritize human-like agent performance. To address this gap, this paper introduces a sample-efficient DRL methodology specifically designed for the training and fine-tuning of agents within industrial contexts, such as the video game sector. By utilizing pre-collected datasets and enhancing network plasticity, our approach significantly boosts the sample efficiency of value-based DRL.

We validated our method by developing a goalkeeper agent for EA SPORTS FC 25, currently one of the top-selling football simulation titles. The results demonstrate that our agent surpasses the game’s native AI by 10% in terms of ball-saving rate. Furthermore, ablation studies reveal that our technique accelerates agent training by 50% relative to conventional DRL methods. A qualitative assessment by domain experts confirms that our approach yields more human-like gameplay than hand-crafted agents. Highlighting the practical impact of this work, the methodology has been integrated into the latest installment of the series.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC